<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:content="http://purl.org/rss/1.0/modules/content/"><channel><title>Dingbo AI field notes</title><link>https://ai.dingboai.com/en/</link><description>What we learned running AI agents on real customer service work.</description><language>en</language>
<item><title>What Is an AI Agent? A Plain Answer With a Real Example</title><link>https://ai.dingboai.com/en/what-is-an-ai-agent.html</link><guid>https://ai.dingboai.com/en/what-is-an-ai-agent.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>An AI agent is software that uses a language model to choose its own steps and tools to finish a task. The parts, the types, and one working example.</description><content:encoded><![CDATA[<p>An AI agent is software that uses a language model to decide which steps to take, and which tools to use, in order to finish a task. You give it a goal, such as "answer this customer email". It works out what it needs to know, looks those things up, and produces a result. A text generator answers from what it was handed. An agent goes and gets what it needs.</p>
  <p>We run AI agents on live customer service for an online brand. The example below is what one of them does with a single email. The volumes are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>A working example</h2>
  <p>A customer writes that their parcel is marked delivered and nothing has arrived. The agent:</p>
  <ul>
    <li>reads the email, plus a summary of the earlier messages in the same thread;</li>
    <li>finds the order from the sender's address and the order number in the text;</li>
    <li>asks the carrier's tracking service for the current status;</li>
    <li>searches the knowledge base for how the company handles this situation;</li>
    <li>looks up past mistakes recorded for this kind of case;</li>
    <li>writes the reply in the customer's language;</li>
    <li>decides whether this is a reply it may send alone or one a person must approve.</li>
  </ul>
  <p>Nobody listed those steps for it in that order. A different email, such as a question about battery range, leads to different steps: no order lookup, one knowledge base search, an answer.</p>

  <h2>The parts of an AI agent</h2>
  <ul>
    <li><strong>A language model.</strong> It reads the input and decides the next step.</li>
    <li><strong>Tools.</strong> Functions the model can call: look up an order, read tracking, search the knowledge base, create a ticket. Without tools, a model can only talk.</li>
    <li><strong>Instructions.</strong> What the job is and what a good result looks like.</li>
    <li><strong>Context.</strong> The facts of this case, prepared for the model: what the customer wants, what has been promised, what has been done.</li>
    <li><strong>Limits.</strong> What the agent may do alone and what needs a person.</li>
  </ul>

  <h2>AI agent, chatbot, workflow</h2>
  <p>A chatbot is a conversation window. What sits behind it can be a script, a model, or an agent. A workflow is a set of steps drawn in advance: every case follows the same path. An agent chooses its path case by case. The three are compared in <a href="https://ai.dingboai.com/en/ai-agent-vs-chatbot.html">AI agent vs chatbot</a> and <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>.</p>

  <h2>Types of AI agents</h2>
  <p>Textbooks sort agents by how they decide: simple reflex agents, model-based agents, goal-based agents, utility-based agents and learning agents. That list describes the inside.</p>
  <p>For a business, the useful split is by what the agent is allowed to do:</p>
  <ul>
    <li><strong>Answers.</strong> It reads and replies. Nothing changes in any system.</li>
    <li><strong>Prepares.</strong> It gathers the facts, proposes a solution and drafts the reply. A person approves.</li>
    <li><strong>Acts within limits.</strong> It sends simple replies and creates tickets on its own, inside rules a person set.</li>
  </ul>
  <p>All three run side by side in the deployment we operate. In the 30 days to 19 November 2026, the chat agent handled 15,725 conversations with no message typed by a human agent. In email, 26,451 replies went out using an AI draft: the AI sent 3,331 on its own, agents sent 3,781 without changing a word, and agents edited 19,339 before sending.</p>

  <h2>What an AI agent is not</h2>
  <ul>
    <li><strong>It is not a colleague with a memory.</strong> It knows about yesterday only if the system hands it the facts.</li>
    <li><strong>It is not always right.</strong> A model can state things that are false. The cases we caught are in <a href="https://ai.dingboai.com/en/ai-customer-service-hallucination.html">hallucination examples and fixes</a>.</li>
    <li><strong>It is not the answer to every job.</strong> If the steps can be drawn in advance, a workflow is cheaper and behaves the same way every time.</li>
  </ul>

  <h2>What makes an agent work on real cases</h2>
  <p><strong>Facts, not raw material.</strong> A model handed a whole email thread or a whole order misses the point. On one email about spare parts, we passed only the order lines the email concerned. The input fell from 37,604 tokens to 4,722, and all four parts in the reply were then correct.</p>
  <p><strong>One source for the rules.</strong> How to speak to a customer is written once, and every part of the system refers to that one text.</p>
  <p><strong>A person where money moves.</strong> Refunds, compensation and replacements are approved by a person. The agent prepares them.</p>
  <p><strong>Measurement by what people change.</strong> Every edit an agent's reviewer makes is a signal about what the agent got wrong.</p>

  <h2>Do you need one?</h2>
  <p>An AI agent fits a job that repeats every day, in real volume, with a system behind it to look things up in, and where the steps differ from case to case. Customer service email is the plainest example. Nine more are in <a href="https://ai.dingboai.com/en/ai-agent-use-cases.html">AI agent use cases</a>.</p>
  <p>Tell us the job in one line and the free plan describes, within two working days, what an agent would handle and what stays with your people.</p>]]></content:encoded></item>
<item><title>What Is Agentic AI? Definition, Examples, and Limits</title><link>https://ai.dingboai.com/en/what-is-agentic-ai.html</link><guid>https://ai.dingboai.com/en/what-is-agentic-ai.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Agentic AI is AI that works toward a goal over several steps, choosing actions and using tools while a person supervises. Definition, examples and limits.</description><content:encoded><![CDATA[<p>Agentic AI is AI that works toward a goal over several steps. It plans what to do, uses tools to do it, checks what came back, and continues until the task is finished or it reaches a limit. A person supervises the work instead of directing each step.</p>
  <p>"Agentic" is the adjective of "agent": having the ability to act. An AI system is agentic to the degree that it chooses its own actions.</p>

  <h2>An example</h2>
  <p>A customer opens the chat on a store's website and writes: "I ordered yesterday, can I still change the delivery address?" In the support system we run, the chat agent:</p>
  <ul>
    <li>identifies the order;</li>
    <li>reads its fulfilment status, because the answer depends on whether it has shipped;</li>
    <li>tells the customer what the status is;</li>
    <li>creates a ticket carrying the order number, the requested address and the status, because changing an order is a job for a person;</li>
    <li>gives the customer the ticket number and the follow-up email address.</li>
  </ul>
  <p>Five steps, two systems, no script that fixed their order. In the 30 days to 19 November 2026, the chat agent created 1,694 tickets for requests that needed a person.</p>

  <h2>Agentic AI vs generative AI</h2>
  <p>Generative AI produces content when asked: a paragraph, an image, a piece of code. Agentic AI uses a generative model as its engine and adds tools, a goal and the freedom to choose steps. One writes. The other gets a job done, and writing is one of its steps. The comparison in detail: <a href="https://ai.dingboai.com/en/agentic-ai-vs-generative-ai.html">agentic AI vs generative AI</a>.</p>

  <h2>Agentic AI vs RPA</h2>
  <p>Robotic process automation follows a script recorded in advance: open this screen, copy that field, paste it there. It handles structured input, and it stops working when the screen or the input changes.</p>
  <p>An agentic system reads free text, such as an email in Italian, and decides which steps the case needs. RPA is the right tool when the input is structured and the steps never vary. Agentic AI is the right tool when every case arrives in different words and needs a different path.</p>

  <h2>Agency is a dial</h2>
  <p>A system is not simply agentic or not. Each kind of action has its own setting:</p>
  <ul>
    <li><strong>Look up:</strong> the AI reads orders, tracking and the knowledge base freely.</li>
    <li><strong>Reply:</strong> the AI sends simple answers on its own and drafts the others for a person.</li>
    <li><strong>Commit the company:</strong> a refund, a replacement, a compensation amount. The AI proposes, a person decides.</li>
  </ul>
  <p>Where to set each one is covered in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">who approves what</a>. The settings move as evidence comes in. On 18 November 2026, the AI sent 290 of the 1,397 email replies that went out that day on its own.</p>

  <h2>Where agentic AI goes wrong</h2>
  <p>Giving a model the freedom to act creates failures that a text generator does not have. These are ones we met in production:</p>
  <ul>
    <li><strong>Acting on old information.</strong> A tracking lookup failed, and the reply presented the last status the system had seen as the current one. The rule now: a failed lookup means no current status, and such a reply goes to a person.</li>
    <li><strong>A promise with no task behind it.</strong> Replies said "we will check with the carrier" on parcels that were moving normally, announcing a follow-up nobody was going to write. The sentence now appears only when tracking shows a real exception, and the carrier check is created as a task.</li>
    <li><strong>Recording the wrong version.</strong> The AI's proposed amount was written to the compensation ledger, and the agent then changed the reply. Of 276 entries over 14 days, 38 did not match what was sent. The ledger now takes the version a person approved.</li>
    <li><strong>A monitor that cannot see.</strong> A check meant to judge the AI's replies returned the same verdict for every case. It was not being shown the reply it was supposed to judge.</li>
  </ul>
  <p>More in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>

  <h2>Examples of agentic AI in production</h2>
  <ul>
    <li>Answering customer chat with order and tracking lookups: 15,725 conversations in 30 days.</li>
    <li>Drafting a reply to every customer email, with the facts gathered first: 26,451 replies sent using an AI draft in 30 days.</li>
    <li>Sending simple email replies alone: 3,331 in 30 days, a median of 6.3 minutes after the customer's email.</li>
    <li>Turning a chat request into a ticket for a person: 1,694 in 30 days.</li>
  </ul>
  <p>All four are counted on 19 November 2026 from one deployment.</p>

  <h2>Questions to ask before buying agentic AI</h2>
  <ul>
    <li>Which actions does it take alone, and who set that list?</li>
    <li>What does it do when a lookup fails?</li>
    <li>Where is each action recorded, and which version is recorded when a person changes the AI's proposal?</li>
    <li>How do you find out it made a mistake?</li>
    <li>How many steps does it take before it stops and asks?</li>
  </ul>
  <p>Describe the job you have in mind and the free plan answers these five for your case, within two working days.</p>]]></content:encoded></item>
<item><title>Agentic AI vs Generative AI: The Difference at Work</title><link>https://ai.dingboai.com/en/agentic-ai-vs-generative-ai.html</link><guid>https://ai.dingboai.com/en/agentic-ai-vs-generative-ai.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Generative AI writes when asked. Agentic AI uses the same models to look things up and finish a task. One customer email handled both ways, and what each needs.</description><content:encoded><![CDATA[<p>Generative AI produces content when you ask for it. Agentic AI uses the same kind of model to carry out a task: it decides what it needs, looks it up, and acts. The first answers a prompt. The second finishes a job.</p>
  <p>The clearest way to see the difference is to give both the same piece of work.</p>

  <h2>One email, two ways</h2>
  <p>A customer writes: "Where is my order? It has been ten days."</p>
  <p><strong>Generative AI.</strong> You paste the email into a model and ask for a reply. It writes a polite, well-formed message: an apology for the wait and a promise to look into it. It contains no tracking status, because the model has none. The customer learns nothing, and someone still has to look the order up.</p>
  <p><strong>Agentic AI.</strong> The system reads the email, finds the order, asks the carrier for the status, checks the delivery time the store's policy states, and writes a reply that gives the latest tracking event with its date and place. If tracking shows a real problem, it also creates the task for a person to contact the carrier.</p>
  <p>Same model family, same email. The second version did the work the first one only talked about.</p>

  <h2>What each one needs</h2>
  <ul>
    <li><strong>Generative AI</strong> needs a prompt and, for accuracy, the source material to write from.</li>
    <li><strong>Agentic AI</strong> needs tools connected to real systems, permission to use them, a definition of the goal, limits on what it may do alone, and a record of what it did.</li>
  </ul>
  <p>This is why a generative tool is running in an afternoon and an agentic system takes a project. The connection work is described in <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">connecting orders and shipping</a>.</p>

  <h2>The risks are different</h2>
  <p>Generative AI gets words wrong. It states a fact that is false, or writes in the wrong tone. A person reading the output catches it.</p>
  <p>Agentic AI gets actions wrong. It tells a customer an address was changed when nobody changed it. It records an amount that a person later altered. It sends a reply built on a lookup that failed. Nobody reads these in advance if the system acts alone.</p>
  <p>So the controls differ. Generative output needs review of the text. Agentic work needs:</p>
  <ul>
    <li>a list of actions the AI may take alone, set by a person;</li>
    <li>approval by a person where money or goods move;</li>
    <li>a record of every action, taken from the version a person approved;</li>
    <li>a rule for failure: when a lookup does not succeed, the AI does not guess, and the case goes to a person.</li>
  </ul>

  <h2>The cost is different</h2>
  <p>A generative request is one call to a model. An agentic task is several: decide, call a tool, read the result, decide again, write. The cost per task is higher and so is the time. In the deployment we run, the replies the AI sends alone go out a median of 6.3 minutes after the customer's email, counted on 19 November 2026.</p>

  <h2>They work together</h2>
  <p>A production system is not one or the other. In the email flow we operate, the path is fixed: gather the facts, decide the solution, write the reply, check it. Inside the "gather" and "decide" steps, the model is free to choose which tools to call. The writing step is plain generation from the facts the earlier steps produced. The design is described in <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>.</p>
  <p>Keeping generation and action apart has a practical benefit: when a reply is wrong, you can tell whether the facts were wrong or the wording was.</p>

  <h2>Which one do you need?</h2>
  <ul>
    <li><strong>Generative AI is enough</strong> when a person supplies the facts and reviews the output: marketing copy, a first draft of a document, a summary of a meeting.</li>
    <li><strong>You need agentic AI</strong> when the answer depends on data in your systems, when the volume is too high for a person to feed each case in by hand, or when the task ends in an action, such as a ticket, a record or a message sent.</li>
  </ul>
  <p>Customer service is the second kind. In 30 days to 19 November 2026, the system we run received 36,789 customer emails and 15,725 chat conversations. Answering them means looking up an order, a tracking status or a policy first.</p>

  <h2>A test for any product called "agentic"</h2>
  <p>Ask the vendor to show one case from start to finish, with the tool calls visible. If the system produced a fluent answer and made no lookups, it is generative AI with a new label.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/what-is-agentic-ai.html">what is agentic AI</a> and <a href="https://ai.dingboai.com/en/what-is-an-ai-agent.html">what is an AI agent</a>. Tell us the job and the free plan shows which parts need an agent and which do not, within two working days.</p>]]></content:encoded></item>
<item><title>AI Agent vs Chatbot: What Is the Difference?</title><link>https://ai.dingboai.com/en/ai-agent-vs-chatbot.html</link><guid>https://ai.dingboai.com/en/ai-agent-vs-chatbot.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>A chatbot is a conversation window. An AI agent looks things up and finishes tasks. How to tell them apart, with a real chat example and questions for vendors.</description><content:encoded><![CDATA[<p>A chatbot is a program you talk to in a chat window. An AI agent is a program that works out the steps a task needs and carries them out with tools. The two overlap: a chatbot can have an agent behind it. The question that matters is what happens after the customer presses Enter.</p>

  <h2>Three things called "chatbot"</h2>
  <ul>
    <li><strong>The scripted bot.</strong> It follows a tree of buttons and keywords written in advance. It answers what its author predicted and nothing else.</li>
    <li><strong>The model that talks.</strong> A language model answers in fluent sentences from the documents it was given. It understands free text in many languages. It cannot look anything up in your systems.</li>
    <li><strong>The agent in a chat window.</strong> A model that can call tools: find the order, read the tracking, search the knowledge base, create a ticket. It decides which of these the question needs.</li>
  </ul>
  <p>All three look the same on the website. They differ in what they can do for the customer.</p>

  <h2>The same question to all three</h2>
  <p>"My order has not arrived. Where is it?"</p>
  <ul>
    <li>The scripted bot shows a button: "Track your order", linking to the tracking page.</li>
    <li>The model that talks explains how tracking works and suggests contacting support.</li>
    <li>The agent finds the order, reads the carrier's status and answers: where the parcel is, since when, and the delivery time the store's policy states.</li>
  </ul>
  <p>Only the third answered the question. How the lookup is made safe is in <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">connecting orders and shipping</a>.</p>

  <h2>How to tell which one you are looking at</h2>
  <ul>
    <li>Ask something that needs your data: "Has order 1234 shipped?" A talker cannot answer.</li>
    <li>Ask in a language the vendor did not demo.</li>
    <li>Ask two things in one message. A scripted bot handles one.</li>
    <li>Ask for something it must not grant, such as a refund. Watch whether it agrees, refuses, or hands over properly.</li>
    <li>Ask the vendor for the log of one conversation, showing the lookups made.</li>
  </ul>

  <h2>What an agent in chat should not do</h2>
  <p>More ability needs tighter limits. Chat replies are sent the moment they are written. Nobody reviews them first. In the system we run, the chat agent answers facts and published policy: products, stock, tracking status, troubleshooting, the text of the terms. It does not agree to anything a person has to carry out: a refund, a replacement, a cancellation, an invoice, an amount of money.</p>
  <p>For those, it creates a ticket with the facts it has gathered and gives the customer the ticket number and the follow-up email address. A person takes it from there by email. The handover is described in <a href="https://ai.dingboai.com/en/ai-customer-service-escalation.html">when AI hands over</a>.</p>

  <h2>Numbers from one deployment</h2>
  <p>Counted on 19 November 2026, for the previous 30 days:</p>
  <ul>
    <li>15,725 chat conversations: 11,010 on the website, 2,865 in the Android app, 1,850 in the iOS app.</li>
    <li>Messages typed by a human agent in those chats: 0.</li>
    <li>Tickets the chat agent created for requests that needed a person: 1,694.</li>
    <li>Conversations that started outside 09:00 to 18:00 Central European Time: 6,902, which is 44 percent.</li>
  </ul>
  <p>The second and third lines belong together. No human typed in the chat, and the customers who needed a human decision were passed to one in a ticket.</p>

  <h2>When a simple chatbot is enough</h2>
  <p>If customers ask the same ten questions and none needs their order, a scripted bot or a page of answers does the job. An agent earns its cost when answers depend on the customer's own order, when questions arrive in several languages, or when a share of conversations must end in a ticket for your team.</p>

  <h2>Questions for a vendor</h2>
  <ul>
    <li>Which of our systems does it read, and can it change anything in them?</li>
    <li>What is it not allowed to promise, and how is that enforced?</li>
    <li>When it hands over, what does the person receive: the whole chat, or the facts?</li>
    <li>What does it say when its lookup fails?</li>
    <li>How is it priced: per seat, per conversation, per resolved case?</li>
  </ul>
  <p>The pricing question is covered in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>. Related: <a href="https://ai.dingboai.com/en/what-is-an-ai-agent.html">what is an AI agent</a> and <a href="https://ai.dingboai.com/en/best-ai-chatbot-for-shopify.html">nine checks for a store chatbot</a>. Send us your ten most common chat questions and the free plan shows which need an agent, within two working days.</p>]]></content:encoded></item>
<item><title>AI in Customer Service: What It Does Today, in Numbers</title><link>https://ai.dingboai.com/en/ai-in-customer-service.html</link><guid>https://ai.dingboai.com/en/ai-in-customer-service.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>What AI does in customer service today, from one live operation: 15,725 chats and 36,789 emails in 30 days, what AI handles alone and what people decide.</description><content:encoded><![CDATA[<p>AI in customer service today does four things: it answers chat questions, drafts replies to email, sends the simple replies on its own, and prepares decisions for people. It does not decide refunds. That is the short version. The rest of this page is one live operation, counted on 19 November 2026, so the claims come with numbers.</p>
  <p>The operation is an online brand selling in eight languages through nine storefronts. The background is in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>1. Chat: answered by AI</h2>
  <p>15,725 chat conversations in 30 days. 11,010 came from the website and 4,715 from the mobile apps. A human agent typed in none of them.</p>
  <p>The chat AI answers questions of fact and published policy: product specifications, stock, where an order is, how to fix a fault, what the warranty says. It looks the order and the tracking up itself. Of these conversations, 44 percent started outside 09:00 to 18:00 Central European Time and 22 percent at the weekend.</p>

  <h2>2. Chat: requests passed to people</h2>
  <p>When a customer in chat needs something a person must carry out, such as a refund, a replacement, a cancellation or an invoice, the AI does not agree to it. It creates a ticket with the facts and gives the customer a ticket number. It created 1,694 tickets in the 30 days. A chat reply is sent the moment it is written, with no review, so the AI commits the company to nothing there.</p>

  <h2>3. Email: a draft for every message</h2>
  <p>36,789 customer emails arrived in 30 days, across five mailboxes. For each, the AI gathers the facts first: the order, the tracking status, what was said earlier in the thread, what the knowledge base says. Then it proposes a solution and drafts the reply in the customer's language.</p>
  <p>26,451 replies were sent using an AI draft:</p>
  <ul>
    <li>19,339 were edited by an agent before sending.</li>
    <li>3,781 were sent by an agent without changing a word.</li>
    <li>3,331 were sent by the AI on its own.</li>
  </ul>
  <p>Seven in ten drafts are edited. That is the honest state of AI in email support: the draft saves the lookup and most of the typing, and the agent's judgement finishes it.</p>

  <h2>4. Email: simple replies sent alone</h2>
  <p>The 3,331 replies the AI sent on its own are replies with no decision in them: an order status, a warranty term, an invoice request, a public discount code, an address change before shipping. They went out a median of 6.3 minutes after the customer's email. On 18 November 2026, the AI sent 290 of the day's 1,397 replies this way.</p>
  <p>These replies are listed in a review queue, sorted by type, where a person checks them. When a reply involves an action, the action is created after the reviewer passes it. The list of types the AI may send alone grows one type at a time, by risk.</p>

  <h2>5. Decisions: prepared, not made</h2>
  <p>Refunds, compensation, replacements and cancellations are decided by people. The AI's part is to have everything ready: the facts, a proposed solution from a fixed list, the draft. When the agent approves, the task goes to the person who carries it out and one line is written to a shared ledger, taken from the version the agent approved. Where the lines sit is in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">who approves what</a>.</p>

  <h2>What customers think</h2>
  <p>Over two days from 17 November 2026, 217 customers rated an email reply. Among the 64 ratings for replies the AI sent alone, 7 were low, 11 percent. Among the 107 for tickets an agent closed, 21 were low, 20 percent. The AI handles the simple cases, so this is not a like-for-like contest. It does show that the replies the AI sends alone are not where the low scores come from. The full reading is in <a href="https://ai.dingboai.com/en/csat-ai-customer-service.html">CSAT for AI customer service</a>.</p>

  <h2>What AI gets wrong</h2>
  <ul>
    <li><strong>It invents.</strong> A shop name, an order number, a policy detail. See <a href="https://ai.dingboai.com/en/ai-customer-service-hallucination.html">hallucination examples and fixes</a>.</li>
    <li><strong>It promises.</strong> In a review of 400 edited drafts, 53 percent of the edits removed a promise the agent did not want to make.</li>
    <li><strong>It misses what is in front of it.</strong> In long threads it overlooks what was already agreed, unless the thread is turned into a short list of facts first.</li>
    <li><strong>It misreads tone.</strong> A sarcastic "thanks" was once read as thanks.</li>
  </ul>
  <p>Each of these has a fix in the system around the model, not in the model.</p>

  <h2>What it changes for the team</h2>
  <p>Agents stop looking things up and start deciding. The work that remains is the work that needs a person: money, exceptions, angry customers. The team also gains cover it did not have: 35 percent of emails and 44 percent of chats arrive outside office hours. See <a href="https://ai.dingboai.com/en/after-hours-customer-service-ai.html">after-hours customer service with AI</a>.</p>

  <h2>Where to start</h2>
  <p>Start with one channel and one type of question, keep a person reviewing everything, and measure what they change. The method is in <a href="https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html">running a pilot on real work</a>. Tell us your channels and volume and the free plan sets out the first step within two working days.</p>]]></content:encoded></item>
<item><title>What Is BPO? And What Changes When AI Does the Work</title><link>https://ai.dingboai.com/en/what-is-bpo.html</link><guid>https://ai.dingboai.com/en/what-is-bpo.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>BPO means business process outsourcing: paying another company to run a process for you. The types, how it is priced, and what changes when AI does the work.</description><content:encoded><![CDATA[<p>BPO stands for business process outsourcing. A company pays another company to run one of its processes, such as answering customers, entering invoices or processing payroll. The buyer keeps the result. The provider supplies the people, the supervision and the routine.</p>

  <h2>Types of BPO</h2>
  <ul>
    <li><strong>Front office:</strong> work that faces the customer. Support by phone, email and chat, order taking, sales calls.</li>
    <li><strong>Back office:</strong> work the customer never sees. Data entry, accounting, payroll, document processing, claims handling.</li>
  </ul>
  <p>By location, BPO is onshore when the provider is in the same country, nearshore when it is in a neighbouring one, and offshore when it is far away. A BPO call center is front-office BPO delivered by phone.</p>

  <h2>How BPO is priced</h2>
  <ul>
    <li><strong>Per seat or per hour.</strong> You pay for people being available, whether or not the work arrives.</li>
    <li><strong>Per transaction.</strong> You pay for each call, ticket or document handled.</li>
    <li><strong>Per outcome.</strong> You pay for a result, such as a resolved case.</li>
  </ul>
  <p>Seat pricing is the traditional form, because the provider's cost is people and people are paid by the hour.</p>

  <h2>Why companies use it</h2>
  <p>To cover hours and languages they cannot staff, to handle volume that rises and falls, and to stop managing a routine that is not their core business. The price they pay is distance: the provider's staff know the process as written, and every exception comes back to the client.</p>

  <h2>What changes when AI does the work</h2>
  <p>We run customer service work for an online brand with AI agents doing the routine and the brand's own team approving decisions. The volumes are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>. Compared with staffed BPO, five things are different.</p>
  <p><strong>Capacity is not headcount.</strong> In the 30 days to 19 November 2026, the system handled 15,725 chat conversations, received 36,789 customer emails and sent 26,451 replies from AI drafts. A rise in volume does not need recruitment or training.</p>
  <p><strong>Hours and languages come as standard.</strong> 44 percent of those chats started outside 09:00 to 18:00 Central European Time, and 22 percent at the weekend. The work covers eight languages from one knowledge base. With staff, each of those is a separate shift or a separate hire.</p>
  <p><strong>Pricing can follow finished work.</strong> When the cost is not hours of people, there is no reason to bill hours. We charge per completed task: a conversation answered, an email replied to, a document produced. The models are compared in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>.</p>
  <p><strong>Data can stay with you.</strong> Staffed BPO means the provider's employees log in to your systems. An AI system can be installed on your own server at go-live, so customer data stays inside your network.</p>
  <p><strong>Decisions stay with your people.</strong> In our setup the AI prepares refunds, replacements and compensation, and your team approves them. In staffed BPO those decisions are either delegated to the provider's staff under a rulebook or sent back to you in a queue.</p>

  <h2>What does not change</h2>
  <ul>
    <li><strong>Someone has to own the process.</strong> Policies, exceptions and the knowledge base need an owner on your side.</li>
    <li><strong>Quality has to be measured.</strong> With people you sample calls. With AI you count what reviewers change: in the same 30 days, agents edited 19,339 of 26,451 AI-drafted replies before sending.</li>
    <li><strong>Hard cases need people.</strong> Angry customers, legal threats, unusual requests. AI routes these. It does not settle them.</li>
  </ul>

  <h2>Where staffed BPO is still the better choice</h2>
  <ul>
    <li>Phone support, which we do not offer.</li>
    <li>Work with low volume and high variety, where there is nothing to repeat.</li>
    <li>Work that depends on judgement no document describes.</li>
    <li>Processes with no system behind them to look things up in.</li>
  </ul>

  <h2>Questions for any provider, staffed or AI</h2>
  <ul>
    <li>What exactly do I pay for: time, transactions or results?</li>
    <li>Who decides when money is involved?</li>
    <li>Where does my customer data sit, and who can read it?</li>
    <li>How do I see the quality of the work each week?</li>
    <li>How do I stop, and what do I keep when I do?</li>
  </ul>

  <h2>How an AI engagement starts</h2>
  <p>Ours has three steps: a free written plan within two working days, a pilot of two to four weeks on your real cases with your team reviewing every output, and then payment per completed task. There is no long contract. The detail is in <a href="https://ai.dingboai.com/en/ai-customer-service-outsourcing.html">AI customer service outsourcing</a> and <a href="https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html">running a pilot</a>.</p>
  <p>Tell us which process you are considering outsourcing and the free plan says which parts AI would take and which stay with people.</p>]]></content:encoded></item>
<item><title>AI Automation Agency: What It Does and How to Pick One</title><link>https://ai.dingboai.com/en/ai-automation-agency.html</link><guid>https://ai.dingboai.com/en/ai-automation-agency.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>An AI automation agency builds and runs AI systems that take over repetitive work for other businesses. The three kinds, what to ask, and warning signs.</description><content:encoded><![CDATA[<p>An AI automation agency is a company that builds, and sometimes runs, AI systems that take over repetitive work for other businesses. AI automation itself means using AI models to do tasks that used to need a person reading, deciding and typing: answering customers, sorting email, filling documents, producing reports.</p>
  <p>The label covers three different kinds of company. Knowing which one you are talking to saves months.</p>

  <h2>Three kinds of AI automation agency</h2>
  <ul>
    <li><strong>The connector.</strong> It links your existing apps with ready-made automation tools and adds a model call in the middle. Fast and cheap to start. It suits simple flows: a form comes in, a summary goes out.</li>
    <li><strong>The builder.</strong> It writes a custom system for your process and hands it over. You get what you specified. Keeping it working afterwards is your job.</li>
    <li><strong>The operator.</strong> It builds the system, runs it on your real work, and is paid for the work the system finishes. It carries the risk that the system does not perform.</li>
  </ul>
  <p>We are the third kind. This page is written from that seat, and the questions below apply to all three.</p>

  <h2>What an agency delivers</h2>
  <ul>
    <li>A study of the job: what comes in, what a person looks up, what they decide, what goes out.</li>
    <li>Connections to your systems: the store, the mailbox, the order system, the spreadsheet.</li>
    <li>The AI part: what the model reads, the rules it follows, the tools it calls.</li>
    <li>The review screen where your people approve what needs approving.</li>
    <li>Measurement: what was handled, what people changed, what went wrong.</li>
  </ul>
  <p>The model is the smallest item on that list. Projects fail on the other four. See <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>

  <h2>Which jobs are worth automating</h2>
  <p>A job is a candidate when it repeats every day, arrives in real volume, has a system behind it to look things up in, and follows rules someone can state. Customer service email and chat fit best. In the deployment we run, the system handled 15,725 chat conversations and received 36,789 customer emails in the 30 days to 19 November 2026. Other jobs of the same shape are in <a href="https://ai.dingboai.com/en/ai-agent-use-cases.html">AI agent use cases</a>.</p>
  <p>A job that happens twice a month, or that depends on judgement nobody can write down, is not worth it.</p>

  <h2>Ten questions to ask an agency</h2>
  <ul>
    <li>Show me one system of yours running on real work today. What volume does it handle?</li>
    <li>Which numbers are measured, and which are estimates?</li>
    <li>What does the AI do alone, and what waits for my staff?</li>
    <li>What happens when a lookup fails or the model is unsure?</li>
    <li>Where does my data sit, and which outside services see it?</li>
    <li>How will I see what my staff corrected each week?</li>
    <li>Do you test on my real past cases before launch?</li>
    <li>Who fixes it when my policy changes?</li>
    <li>What do I pay for: hours, a project, usage or results?</li>
    <li>If I stop, what do I keep?</li>
  </ul>

  <h2>Warning signs</h2>
  <ul>
    <li><strong>A demo on invented cases.</strong> Real cases are messy: long threads, missing order numbers, three questions in one email.</li>
    <li><strong>An accuracy figure with no method.</strong> Ask how many cases, which ones, and who judged them.</li>
    <li><strong>"Fully autonomous" for work involving money.</strong> Someone has to approve refunds.</li>
    <li><strong>No plan for mistakes.</strong> Every system makes them. Ask how they are found.</li>
    <li><strong>Payment in full before anything has run on your work.</strong></li>
  </ul>

  <h2>How pricing works</h2>
  <p>Agencies charge in four ways: by the hour, by fixed project, by usage, or by result. Each puts the risk in a different place. With hourly and project pricing, you pay whether or not the system performs. With pricing by result, the agency is paid only for finished work, so it has a reason to keep the system working. The comparison is in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>, and what drives a custom quote is in <a href="https://ai.dingboai.com/en/custom-ai-agent-development-cost.html">custom AI agent development cost</a>.</p>

  <h2>How we work</h2>
  <ul>
    <li><strong>Free plan.</strong> Within two working days you get a written plan: which systems we connect, what the AI handles, what stays with your people.</li>
    <li><strong>Pilot on real work.</strong> Two to four weeks on your real cases. Your team reviews every output, and you see every record.</li>
    <li><strong>Pay per completed task.</strong> The pilot shows how much the system handled and how accurately. We agree a unit price on those numbers. There is no long contract.</li>
  </ul>
  <p>At go-live the system is installed on your own server. We do not take on regulated finance or medical work.</p>

  <h2>Agency or in-house</h2>
  <p>Build in-house when you have engineers who can connect your systems and someone who will watch the output every week for as long as it runs. Use an agency when you want the result without building that team. The options are laid out in <a href="https://ai.dingboai.com/en/ai-customer-service-system-build-vs-buy.html">build, open source or buy</a>.</p>
  <p>Tell us in one line the job that takes the most hours on your team. The free plan comes back within two working days.</p>]]></content:encoded></item>
<item><title>Best AI Chatbot for Shopify: 9 Checks Before You Choose</title><link>https://ai.dingboai.com/en/best-ai-chatbot-for-shopify.html</link><guid>https://ai.dingboai.com/en/best-ai-chatbot-for-shopify.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Nine checks to find the best AI chatbot for a Shopify store: real order lookup, failed lookups, refunds in chat, email, languages, and how you pay for it.</description><content:encoded><![CDATA[<p>Every AI chatbot for Shopify looks good in a demo, because a demo asks questions the product page already answers. The questions that fill a real support inbox are different: where is my order, why was I charged twice, this part does not fit, I want to return it. The nine checks below come from running an AI agent for an e-bike brand on nine Shopify storefronts in eight languages, where it answered 15,445 chat conversations in the last 30 days. Use them on any chatbot you are considering, including ours.</p>

  <h2>1. Can it look up a real order?</h2>
  <p>Ask the chatbot about a specific order from your store. It should find the order by order number or by the email used at checkout, and confirm the match before it shows any details. A chatbot that answers "you can check your order status on the order page" has not connected to anything.</p>

  <h2>2. What does it say when the lookup fails?</h2>
  <p>This check separates products faster than any other. Give it an order number with a typo. A lookup has four possible results: found, confirmed that no such order exists, found part of the record, or could not check. The chatbot has to tell these apart. One that treats "could not check" as "does not exist" tells a paying customer that their order is not in the system.</p>

  <h2>3. Does it know which store the customer bought from?</h2>
  <p>If you run several storefronts, prices, shipping options and return policies differ by market. The same question from a customer of the German store and a customer of the UK store can need different answers. Ask the same policy question as two customers from two stores and compare.</p>

  <h2>4. Does it read carrier tracking?</h2>
  <p>Shopify holds a tracking number. The carrier knows where the parcel is. The chatbot should read the carrier's status and explain it in plain words, and it should only read: no writing into carrier systems, no starting a claim on its own. What this takes is covered in <a href="https://ai.dingboai.com/en/wismo-where-is-my-order-automation.html">automating where-is-my-order requests</a>.</p>

  <h2>5. What does it do with a refund request?</h2>
  <p>Type "I want a refund" into the chat. A chat reply goes out instantly and cannot be taken back, so the correct behaviour is narrow: the chatbot does not agree, does not offer an amount, and does not suggest compensation. It creates a ticket, gives the customer the ticket number and the email address where the team will follow up, and keeps answering factual questions. The same applies to returns, exchanges, cancellations, address changes and invoices. We made this rule absolute after chat replies agreed to reissue invoices on their own.</p>

  <h2>6. What happens after the handoff?</h2>
  <p>Ask to see the ticket the chatbot created. It should carry what the chatbot already established: what the customer wants, the order number, the verified order and shipping status, what the customer has sent, and what is still missing. If your agent has to reread the whole chat, the chatbot saved the customer's time and spent yours. See <a href="https://ai.dingboai.com/en/ai-customer-service-escalation.html">when AI hands over</a>.</p>

  <h2>7. Does it work on email?</h2>
  <p>In our deployment, refunds and returns get settled by email, and email carries the heavier cases. The brand received 36,294 customer emails in 30 days, and 25,576 replies were sent using an AI draft. A chatbot that only covers the chat widget leaves that volume untouched. Ask whether the same agent, with the same knowledge and the same order access, drafts email replies for your team to approve.</p>

  <h2>8. Where do its product facts come from?</h2>
  <p>Specifications differ between versions of a product and between markets, and stock changes daily. Ask where the chatbot gets a product fact and how it knows which version the customer owns. Then ask about an item that is out of stock. In a review of 400 AI drafts that our client's agents edited, 22 percent of the edits added stock status the draft had left out, so check this one on purpose.</p>

  <h2>9. What exactly do you pay for?</h2>
  <p>Per seat, per conversation started, per resolved conversation, or per completed task are four different bills for the same month. Ask what counts as a billable unit and where the count comes from. If a conversation that ends with "please contact support" counts as resolved, you are paying for the chatbot to send customers away. The models are compared in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>.</p>

  <h2>How to run the checks</h2>
  <p>Do not use questions you invent. Export a few hundred real conversations and emails from the last month, including the ugly ones, and replay them against the chatbot. Compare each answer with what your agent actually sent. Invented test questions hide the problems that real history exposes.</p>
  <p>Then split the results by what the request needed: information, which the chatbot should answer, and action, which it should hand to a person with the facts attached. A chatbot that gets both right on your own history is the best one for your store.</p>
  <p>The systems to connect are listed in <a href="https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html">Shopify AI customer service agent: what to connect</a>. Our free plan runs these nine checks against your store's setup within two working days.</p>]]></content:encoded></item>
<item><title>AI Customer Service for WooCommerce, Magento, Custom Stores</title><link>https://ai.dingboai.com/en/ai-customer-service-woocommerce-magento-custom-store.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-woocommerce-magento-custom-store.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>What an AI customer service agent needs from a WooCommerce, Magento, BigCommerce or custom store: six read-only data points and how each platform exposes them.</description><content:encoded><![CDATA[<p>An AI customer service agent does not depend on Shopify. It depends on six pieces of data that every online store holds, whatever it is built on. Our production deployment reads them from nine Shopify storefronts; this page lists the six, and how a WooCommerce, Magento, BigCommerce or custom-built store exposes each one. If your store can answer these six questions through an API, the agent can do the work described in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>The six things the agent reads</h2>
  <ul>
    <li><strong>An order, by order number.</strong> Items, quantities, amounts, payment status, the shipping address.</li>
    <li><strong>Orders, by customer email.</strong> Customers write without their order number. The agent searches by the email used at checkout and confirms the match before showing details.</li>
    <li><strong>Fulfilment status and tracking number.</strong> Whether the order has shipped, with which carrier, under which number.</li>
    <li><strong>The store or market the order belongs to.</strong> Prices, shipping options and return policies differ by country.</li>
    <li><strong>Product facts and stock.</strong> Which version the customer bought, and whether an item is in stock today.</li>
    <li><strong>Refund and return records.</strong> What has already been paid back or sent back, so the reply does not contradict it.</li>
  </ul>
  <p>All six are read-only. The agent does not edit orders, cancel them or issue refunds. Those are actions, and actions go to a person.</p>

  <h2>WooCommerce</h2>
  <p>WooCommerce has a REST API with keys that can be created as read-only. Orders can be fetched by number and searched by billing email, and products carry stock status. One point needs checking in every WooCommerce store: tracking numbers. WooCommerce itself has no tracking field. The number lives in whatever shipment tracking plugin the store uses, or in the order notes, and the agent has to be pointed at the right place. A store that keeps tracking numbers only in emails sent by the warehouse has nothing for the agent to read.</p>

  <h2>Magento and Adobe Commerce</h2>
  <p>Magento exposes orders, shipments with their tracking entries, products and stock through its REST API. Access is granted by an integration token whose permissions are set per resource, so the token can be limited to reading sales and catalogue data. Magento stores often run several store views for different countries and languages inside one installation. Each order records its store view, and the agent uses it to choose the right policy and language.</p>

  <h2>BigCommerce</h2>
  <p>BigCommerce API accounts are created with scopes, and orders, products and customers can each be set to read-only. Shipments, with tracking numbers, are a sub-resource of the order.</p>

  <h2>A custom-built store</h2>
  <p>A store built in-house has no standard API, and it does not need one. It needs two read-only endpoints: one that returns an order by number, and one that returns the orders for an email address, each including fulfilment status, tracking number, market and refund records. Each one is a single query. If no developer is available, a read-only database view serves the same purpose.</p>

  <h2>What happens when a lookup fails</h2>
  <p>On every platform, the integration must return one of four results: found, confirmed that none exists, found part of the record, or could not check. A plugin conflict or a timeout on a WooCommerce site is a "could not check", and the agent must say it could not read the order right now. An integration that returns an empty result in that situation makes the agent tell a paying customer that their order does not exist.</p>

  <h2>Where the conversations come from</h2>
  <p>The store platform is only the order side. Customer messages arrive through channels, and the agent connects to those separately:</p>
  <ul>
    <li><strong>Email.</strong> The support mailboxes, read over IMAP and answered in the same thread. See <a href="https://ai.dingboai.com/en/ai-shared-inbox-support-email.html">AI for a shared support inbox</a>.</li>
    <li><strong>Chat.</strong> A chat window on the website and inside the mobile app. See <a href="https://ai.dingboai.com/en/ai-chat-website-mobile-app-support.html">AI chat on website and app</a>.</li>
    <li><strong>A helpdesk.</strong> If your team works in Zendesk, Freshdesk or Gorgias, the agent attaches there. See <a href="https://ai.dingboai.com/en/ai-agent-zendesk-freshdesk-gorgias.html">adding an AI agent to a helpdesk</a>.</li>
  </ul>

  <h2>What the work looks like once connected</h2>
  <p>The volumes below come from the Shopify deployment, counted on 19 November 2026 for the previous 30 days. The same jobs run on the six data points above:</p>
  <ul>
    <li>15,725 chat conversations answered by AI, across nine storefronts.</li>
    <li>36,789 customer emails received, each with an AI draft that has the order and shipment already checked.</li>
    <li>26,451 replies sent using an AI draft: 3,331 by the AI on its own, 3,781 approved without an edit, 19,339 edited and then sent.</li>
    <li>1,694 tickets created by the chat AI for requests that needed a person.</li>
  </ul>

  <h2>A checklist before you start</h2>
  <ul>
    <li>Create a read-only API key or token. Do not hand over an admin login.</li>
    <li>Find where tracking numbers are stored, and confirm they are stored for every shipment.</li>
    <li>List your stores or store views with their countries and languages.</li>
    <li>Confirm that refunds are recorded on the order, not only in the payment provider.</li>
    <li>Export a few hundred real conversations for the replay test described in <a href="https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html">how to run an AI agent pilot</a>.</li>
  </ul>
  <p>For the Shopify version of this page, see <a href="https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html">what to connect on Shopify</a>. Tell us which platform your store runs on and the free plan names the endpoints we would read, within two working days.</p>]]></content:encoded></item>
<item><title>Adding an AI Agent to Zendesk, Freshdesk or Gorgias</title><link>https://ai.dingboai.com/en/ai-agent-zendesk-freshdesk-gorgias.html</link><guid>https://ai.dingboai.com/en/ai-agent-zendesk-freshdesk-gorgias.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Three ways to add an AI agent to a helpdesk such as Zendesk, Freshdesk or Gorgias: at the mailbox, through the ticket API, or with an approval workbench.</description><content:encoded><![CDATA[<p>A team that already works in Zendesk, Freshdesk or Gorgias does not have to leave it to get an AI agent that checks orders and drafts replies. There are three places to attach the agent, and they differ in how much the team's daily routine changes. Our production system uses the first and the third. This page describes all three so you can choose.</p>

  <h2>Option 1: attach at the mailbox</h2>
  <p>Every helpdesk receives email from a support address. The agent can read the same mailbox over IMAP, before or alongside the helpdesk. It distils each email, checks the order and the knowledge base, and prepares a draft and a proposed solution.</p>
  <p>This is how our deployment receives mail: five mailboxes, 36,789 customer emails in the 30 days to 19 November 2026. It needs nothing from the helpdesk vendor, and it works with a helpdesk that has no API at all. The details are in <a href="https://ai.dingboai.com/en/ai-shared-inbox-support-email.html">AI for a shared support inbox</a>.</p>

  <h2>Option 2: attach through the ticket API</h2>
  <p>Zendesk, Freshdesk and Gorgias each have an API that reads tickets and writes to them, and each supports internal notes that customers do not see. With this option the helpdesk stays the team's only screen:</p>
  <ul>
    <li>A new ticket or a new customer message triggers the agent.</li>
    <li>The agent reads the ticket and its history through the API, and the order through the store's API.</li>
    <li>It writes its output back to the ticket as an internal note: the facts it established with their sources, the proposed solution, and the reply text.</li>
    <li>The human agent reads the note, sends the reply as it is or edits it, and decides on any action.</li>
  </ul>
  <p>What the helpdesk does not provide is the part after the decision. A refund or a reshipment that an agent approves still has to reach the person who carries it out, with an owner and a record. In this option that step is built on tags, ticket fields and the helpdesk's own automation rules.</p>

  <h2>Option 3: a workbench built for approvals</h2>
  <p>Our production system has its own ticket workbench, built around the fact that most of an agent's work is now reviewing. What it shows for each ticket:</p>
  <ul>
    <li>The customer's request for this round, separated from quoted text and old messages.</li>
    <li>A summary of the earlier exchange as dated facts: what was promised, what was done, what is still open.</li>
    <li>The verified order and shipping status.</li>
    <li>The AI's proposed solution and the draft reply. Replies in other languages are shown with a translation, so one team reviews every market.</li>
    <li>Two buttons: send and keep open, or send and close. Closing asks whether someone has to act; if so, the task goes to the execution list and the team's group chat at once.</li>
  </ul>
  <p>Around the ticket sit the queues a helpdesk does not have: an "AI review" queue for solutions the AI handled on its own, an execution list split by the action a person must take, and a complaint record that archives every reply. One line per decision is written to a shared ledger, using the version a person finally approved.</p>

  <h2>How to choose</h2>
  <ul>
    <li><strong>You want drafts in front of agents with no change to their tools:</strong> option 2.</li>
    <li><strong>Your helpdesk plan has no API access, or you use a shared mailbox:</strong> option 1.</li>
    <li><strong>Most of your tickets end in an action, such as a part to ship or a refund to pay:</strong> option 3, because the cost sits in what happens after the reply.</li>
  </ul>
  <p>The options combine. A team can keep its helpdesk for phone and social channels and run email through the workbench.</p>

  <h2>What stays the same in all three</h2>
  <ul>
    <li><strong>Reading before the model.</strong> The thread is distilled into structured facts first. A raw thread of 25 messages produced a reply that contradicted an earlier promise; structured facts fixed it and cut the input from 37,604 tokens to 4,722.</li>
    <li><strong>Four lookup results.</strong> Found, confirmed none, partial, could not check.</li>
    <li><strong>The approval line.</strong> Anything that moves money or goods is decided by a person.</li>
    <li><strong>Counting.</strong> Sent by the AI, sent unchanged, edited before sending. In the 30 days to 19 November 2026 those were 3,331, 3,781 and 19,339.</li>
  </ul>

  <h2>Checking a helpdesk's built-in AI</h2>
  <p>Helpdesks sell AI features of their own. Check four things before relying on one for after-sales work: whether it looks up a specific order and its carrier status, what it says when that lookup fails, whether it keeps earlier promises consistent across a long thread, and what you are billed for. The nine checks in <a href="https://ai.dingboai.com/en/best-ai-chatbot-for-shopify.html">our chatbot checklist</a> apply to helpdesk AI as written.</p>

  <h2>What we need from you to start</h2>
  <ul>
    <li>For option 1: IMAP and SMTP access to the support mailboxes.</li>
    <li>For option 2: an API token for the helpdesk with permission to read tickets and add internal notes.</li>
    <li>For all: read-only access to orders, as listed for <a href="https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html">Shopify</a> and for <a href="https://ai.dingboai.com/en/ai-customer-service-woocommerce-magento-custom-store.html">WooCommerce, Magento and custom stores</a>.</li>
  </ul>
  <p>Tell us which helpdesk you use and the free plan states which option fits, within two working days.</p>]]></content:encoded></item>
<item><title>AI for a Shared Support Inbox: Gmail, Outlook, IMAP</title><link>https://ai.dingboai.com/en/ai-shared-inbox-support-email.html</link><guid>https://ai.dingboai.com/en/ai-shared-inbox-support-email.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How an AI agent works a shared support inbox on Gmail, Outlook or any IMAP mailbox: threading, quoted text, automated mail, several mailboxes, thread replies.</description><content:encoded><![CDATA[<p>Many support teams have no helpdesk. They have support@ and three people who share it in Gmail or Outlook. An AI agent can work that inbox directly: it connects over IMAP to read and SMTP to send, the same way a mail client does. The mailbox-level work is less glamorous than writing replies and decides whether the replies are any good. This is what it involves, from a system that received 36,789 customer emails through five mailboxes in the 30 days to 19 November 2026.</p>

  <h2>Connecting the mailbox</h2>
  <p>Gmail, Microsoft 365 and most hosted mail services offer IMAP. Each requires an app-specific password or an authorised connection for it. A mailbox that logs in on the web and refuses IMAP is missing that step. One more setting matters: the retrieval range. A mailbox set to expose only the last 30 days over IMAP hides older messages of a long-running thread from the agent.</p>

  <h2>Turning emails into tickets</h2>
  <p>An inbox is a list of messages. Support work is a list of cases. The agent merges messages into tickets by the reply headers each email carries, by the customer's address and by order number. Three details cause trouble if ignored:</p>
  <ul>
    <li>A customer who writes from two addresses about one order is one case.</li>
    <li>An order number has to be matched as a whole. A number that is the beginning of a longer number is a different order.</li>
    <li>Reference numbers that look like order numbers are not order numbers. A police report number was once read as an order number and pulled in a stranger's order.</li>
  </ul>

  <h2>Separating new text from quoted text</h2>
  <p>A reply email contains the customer's new message plus everything quoted from before. The agent has to answer the new part. Mail clients mark quotes differently, and customers write between quoted lines. When plain text is ambiguous, the HTML version of the email settles it. Text that the customer typed under a quoted block is new text, and treating it as a quote loses the request.</p>

  <h2>Closing what is not a customer</h2>
  <p>A shared inbox receives delivery failure notices, out-of-office replies, platform notifications and spam. The agent closes automated messages without drafting a reply. It also checks the spam folder, because real customers land there.</p>

  <h2>Several mailboxes, one queue</h2>
  <p>Teams grow into several addresses: support@, service@, one per country. In our deployment two mailboxes carry 96 percent of the volume and three smaller ones the rest. All of them feed one queue, and each ticket records which mailbox it came through, so the reply leaves from the address the customer wrote to.</p>

  <h2>Replying in the thread</h2>
  <p>The reply goes out as a reply to the customer's most recent email, with the original subject and the correct reply headers. In the customer's mail client it sits under their own message. A new email with a new subject splits the conversation and gets overlooked. A message that starts a conversation, such as a notice about a delayed order, is sent as a new email and does not pretend to be a reply.</p>

  <h2>Language detection on short emails</h2>
  <p>The agent replies in the customer's language, so it has to identify that language. Detection fails on short emails and on text written in capital letters. An all-capitals Italian message was classified as Vietnamese and answered in Vietnamese before we changed the detection for short and upper-case text. Check this on your own mail before any reply is sent automatically.</p>

  <h2>What the agent does with each email</h2>
  <ul>
    <li>Distils the new message and the thread into structured facts with sources and dates.</li>
    <li>Checks the order, the shipment and the knowledge base.</li>
    <li>Proposes a solution and writes the reply.</li>
    <li>Sends it, if the reply is simple and involves no money or promise. Otherwise it waits for an agent.</li>
  </ul>
  <p>In the same 30 days, 26,451 replies were sent using an AI draft. The AI sent 3,331 on its own, and half of those left within 6.3 minutes of the customer's email. Agents approved 3,781 without an edit and edited 19,339 before sending.</p>

  <h2>Email arrives when nobody is at work</h2>
  <p>Of the 36,789 emails, 12,693 arrived outside 09:00 to 18:00 Central European Time, and 5,163 on a Saturday or Sunday. In a shared inbox those wait for Monday. With the agent, the simple ones are answered and the rest are waiting as checked drafts. More in <a href="https://ai.dingboai.com/en/after-hours-customer-service-ai.html">after-hours customer service with AI</a>.</p>

  <h2>Who this fits</h2>
  <p>Teams of two to twenty people answering customers from Gmail, Google Workspace, Outlook, Microsoft 365 or a hosted mailbox, with orders in a store or an ERP system. You keep your addresses and your mail provider.</p>
  <p>The writing rules for the replies are in <a href="https://ai.dingboai.com/en/ai-reply-customer-emails-best-practices.html">AI replies to customer emails</a>, and the sorting in <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a>. Send us the number of mailboxes and the monthly volume and the free plan arrives within two working days.</p>]]></content:encoded></item>
<item><title>AI Chat Support on Website and Mobile App: Real Split</title><link>https://ai.dingboai.com/en/ai-chat-website-mobile-app-support.html</link><guid>https://ai.dingboai.com/en/ai-chat-website-mobile-app-support.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>AI chat support across a website and iOS and Android apps: 15,725 conversations split by channel and country, what chat answers, and what becomes a ticket.</description><content:encoded><![CDATA[<p>A brand with a website and a mobile app has at least three chat entry points: the web page, the iOS app and the Android app. Customers use them for different things, and one AI agent has to serve all three with the same knowledge. These are the numbers from a deployment for an e-bike brand, counted on 19 November 2026 for the previous 30 days, and the rules the chat runs on.</p>

  <h2>Where 15,725 conversations came from</h2>
  <ul>
    <li><strong>Website:</strong> 11,010 conversations.</li>
    <li><strong>Android app:</strong> 2,865.</li>
    <li><strong>iOS app:</strong> 1,850.</li>
  </ul>
  <p>Three in ten conversations started inside the app. People in the app already own the product and ask how to use it. Website visitors include people who have not bought yet. One agent covers both, and it knows which entry point a conversation came from.</p>

  <h2>Nine storefronts, one agent</h2>
  <p>The same 15,725 conversations by storefront: United Kingdom 2,684, France 2,601, Italy 2,573, Germany 2,390, Spain 1,705, the EU store 1,695, the Netherlands 1,171, Poland 696, United States 202. The agent answers in the customer's language and applies the prices, shipping options and policies of the store the customer is on. A question about delivery time from the Polish store and from the UK store gets two different answers, both correct.</p>

  <h2>What the chat answers</h2>
  <p>A chat reply goes out the moment it is written and cannot be taken back. So the chat answers facts and published policy:</p>
  <ul>
    <li>product questions, specifications and compatibility,</li>
    <li>stock,</li>
    <li>order and shipping status, after confirming the order belongs to the person asking,</li>
    <li>how to use the product and how to troubleshoot it, with the steps and the video guide for that model,</li>
    <li>policy text: warranty terms, return conditions, delivery times.</li>
  </ul>

  <h2>What the chat turns into a ticket</h2>
  <p>Anything that needs a person to act: an invoice, a refund, compensation, a reshipment, an exchange, a cancellation, an order change, an address change. The chat AI does not agree to it, does not suggest a remedy and does not quote an amount. It creates a ticket with the facts it has gathered, tells the customer the ticket number and the email address where the team will follow up, and continues with factual questions.</p>
  <p>In the 30 days counted, the chat AI created 1,694 tickets this way. Agents did not step into the chat window. The requests that needed people reached them as tickets with the facts attached.</p>

  <h2>Do not make the customer choose a department first</h2>
  <p>Our chat opens with two buttons, pre-sales and after-sales. In one week, 1,300 conversations began with a tap on one of them, and the reply to that tap was a question back to the customer. A button that produces a question is a wasted turn. The first reply to a tap should already help: for after-sales, ask for the order number and say what the agent can check with it.</p>

  <h2>Keep a standard answer whole</h2>
  <p>When the knowledge base holds a standard answer to a question, the chat gives all of it. A support lead asked the chat a question that has a standard answer in the knowledge base and received only part of it. Chat answers should be short, and the place to save words is the wrapping around the answer. The facts stay complete.</p>

  <h2>Send the right video</h2>
  <p>Video guides are matched to the product model by the video's own title, not by the name of the page it was listed on. A page shared by several models once supplied a tail-light video for the wrong one. When a customer says the video does not match their product, the agent must not send the same link again.</p>

  <h2>Chat does not sleep</h2>
  <p>Of the 15,725 conversations, 6,902 started outside 09:00 to 18:00 Central European Time, 3,537 on a Saturday or Sunday, and 2,079 between 22:00 and 06:00. A staffed live chat covers a fraction of those hours. The pattern is in <a href="https://ai.dingboai.com/en/after-hours-customer-service-ai.html">after-hours customer service with AI</a>.</p>

  <h2>What to connect</h2>
  <ul>
    <li>The chat window on the website, and the chat screen in the app through the same service, with the storefront and the entry point passed along.</li>
    <li>Read-only order lookup. See <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">AI chatbot order tracking</a>.</li>
    <li>The knowledge base. See <a href="https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html">what goes into it</a>.</li>
    <li>The ticket system that receives the handoff. See <a href="https://ai.dingboai.com/en/ai-customer-service-escalation.html">when AI hands over</a>.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Brands that sell in several countries, have an app or plan one, and receive the same product and order questions in every language. The free plan maps your chat entry points and what each should answer within two working days.</p>]]></content:encoded></item>
<item><title>After-Hours Customer Service With AI: Real Numbers</title><link>https://ai.dingboai.com/en/after-hours-customer-service-ai.html</link><guid>https://ai.dingboai.com/en/after-hours-customer-service-ai.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How much customer service arrives after hours and at weekends, with real counts, and how an AI agent covers it: what it answers and what waits for a person.</description><content:encoded><![CDATA[<p>Support teams work office hours. Customers do not write in office hours. Before deciding how to cover nights and weekends, count how much arrives then. We counted for an e-bike brand selling across Europe and the US, on 19 November 2026, for the previous 30 days, using Central European Time because most of its customers live there.</p>

  <h2>How much arrives after hours</h2>
  <p><strong>Chat:</strong> 15,725 conversations.</p>
  <ul>
    <li>6,902 started outside 09:00 to 18:00. That is 44 percent.</li>
    <li>3,537 started on a Saturday or Sunday. That is 22 percent.</li>
    <li>2,079 started between 22:00 and 06:00. That is 13 percent.</li>
  </ul>
  <p><strong>Email:</strong> 36,789 customer emails.</p>
  <ul>
    <li>12,693 arrived outside 09:00 to 18:00. That is 35 percent.</li>
    <li>5,163 arrived on a Saturday or Sunday. That is 14 percent.</li>
  </ul>
  <p>Chat leans further into evenings and weekends than email does. People chat when they are using the product, and they use an e-bike after work and at the weekend.</p>

  <h2>Three ways to cover those hours</h2>
  <ul>
    <li><strong>Shifts.</strong> Evening and weekend staff in every language you sell in. For eight languages that is a rota few brands can fill.</li>
    <li><strong>An auto-reply.</strong> "We have received your message and will reply within two working days." It confirms receipt and answers nothing.</li>
    <li><strong>An AI agent.</strong> It answers what can be answered now, and prepares the rest for the morning.</li>
  </ul>

  <h2>What the AI does at night</h2>
  <p><strong>In chat</strong> it answers as it does at any hour: product facts, stock, order and shipping status, usage, troubleshooting, published policy. In the 30 days counted, agents did not step into the chat window at any hour.</p>
  <p><strong>In email</strong> it reads every message, checks the order and the shipment, and writes a draft with a proposed solution. Replies that involve no money, no promise and no policy exception go out on their own. In the 30 days counted the AI sent 3,331 emails this way, and half of them left within 6.3 minutes of the customer's email. On 18 November 2026 it sent 290 of the 1,397 replies that went out that day.</p>

  <h2>What waits for a person</h2>
  <p>A refund, a reshipment, a return, a cancellation or a compensation amount is a decision, and nobody is there to make it at 23:00. The AI does not make it instead. What it does depends on the channel:</p>
  <ul>
    <li><strong>Chat:</strong> it creates a ticket with the facts gathered, and gives the customer the ticket number and the email address where the team will follow up. The customer knows the request is logged and where the answer will come from.</li>
    <li><strong>Email:</strong> the draft and the proposed solution sit in the agent's queue. In the morning the agent opens a case that is already checked, approves or edits, and sends.</li>
  </ul>
  <p>Each role in the system has a named person responsible for it. While that person is offline the AI keeps checking facts and preparing plans, and it carries out nothing that needs their decision.</p>

  <h2>The morning queue</h2>
  <p>Without an agent, Monday starts with the weekend's unread mail: read each one, find the order, check tracking, then write. With an agent, Monday starts with decisions. Every weekend email has its facts established, its order and tracking checked, and a draft waiting. The simple ones are already answered and gone from the queue.</p>

  <h2>Rules that matter more at night</h2>
  <ul>
    <li><strong>No promise the day shift has to break.</strong> The reply says what is known and what happens next. It does not agree to a remedy to end the conversation. In a review of 400 drafts that agents edited, 53 percent of the edits removed a promise the agent did not want to make.</li>
    <li><strong>A failed lookup is not an answer.</strong> Carrier pages fail at night too. When tracking cannot be read, the reply says so, an older status is not presented as current, and the reply is not sent automatically.</li>
    <li><strong>A thank-you is not a request.</strong> A closing "thanks" needs a short reply. Check that the agent tells sincere thanks from sarcasm: a customer's sarcastic thanks was once answered as if it were sincere.</li>
    <li><strong>Stop when a person has taken the case.</strong> If an agent has already replied on a ticket, the AI does not send on it overnight.</li>
  </ul>

  <h2>Surveys and after-hours replies</h2>
  <p>If you send a satisfaction survey when a ticket closes, an overnight AI reply that answered a simple question earns good scores. A survey sent while the customer is still waiting for something a person must do earns bad ones. The numbers are in <a href="https://ai.dingboai.com/en/csat-ai-customer-service.html">CSAT for AI customer service</a>.</p>

  <h2>How to count your own</h2>
  <p>Export one month of chat start times and email arrival times. Convert them to the time zone your customers live in. Count three groups: outside your staffed hours, weekends, and 22:00 to 06:00. Then read a hundred of the after-hours messages and mark each as information or action. The information share is what an agent answers the same night.</p>
  <p>Channel details are in <a href="https://ai.dingboai.com/en/ai-chat-website-mobile-app-support.html">AI chat on website and app</a> and <a href="https://ai.dingboai.com/en/ai-shared-inbox-support-email.html">AI for a shared support inbox</a>. Send us your staffed hours and the free plan shows what would be covered, within two working days.</p>]]></content:encoded></item>
<item><title>Pre-Sales AI Chat: Stock, Specs, Offers, Compatibility</title><link>https://ai.dingboai.com/en/ai-presales-chat-ecommerce.html</link><guid>https://ai.dingboai.com/en/ai-presales-chat-ecommerce.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>What a pre-sales AI chat answers before the order: specifications by version and market, stock, published offers, compatibility, questions with no set answer.</description><content:encoded><![CDATA[<p>A pre-sales question is a customer deciding whether to buy. The answer has to be right, because a wrong specification becomes a return, and it has to be fast, because the customer is on the product page now. Pre-sales is one of five roles in the support system from our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>. This page covers what that role answers, what it refuses to do, and where its facts come from.</p>

  <h2>What pre-sales customers ask</h2>
  <ul>
    <li><strong>Specifications:</strong> range, weight, dimensions, load, what is in the box.</li>
    <li><strong>Differences:</strong> between two models, or between versions of one model.</li>
    <li><strong>Stock and delivery:</strong> is it available, when does it ship, how long to my country.</li>
    <li><strong>Price and offers:</strong> is there a discount, does the code work on this item.</li>
    <li><strong>Compatibility:</strong> does this accessory fit that model.</li>
    <li><strong>Policy:</strong> warranty length, returns, payment methods, invoices for a company.</li>
    <li><strong>Local rules:</strong> may I use this product on the road in my country.</li>
  </ul>

  <h2>Specifications depend on version and market</h2>
  <p>The same model name can cover different versions, and the version sold in one country can differ from the version sold in another. The knowledge base names the version for each fact. When versions differ, the product page of the store the customer is on is the reference. A chat that answers from a single specification sheet gives a customer in Poland the motor power of the US version.</p>

  <h2>Stock is part of the answer</h2>
  <p>A customer asking about a product wants to know whether they can have it. In a review of 400 AI drafts that agents edited before sending, 22 percent of the edits added stock status the draft had left out. The pre-sales role reads stock for the customer's store and states it. When an item is out of stock, the reply says so in the first sentence and gives the restock information the store has published.</p>

  <h2>Offers: published ones only</h2>
  <p>The AI tells customers about discount codes and promotions the store has made public, and explains the conditions. It does not create a discount, extend an expired one or match a competitor's price. A special price is a decision for a person. In chat the AI creates a ticket for such a request. By email it drafts the reply and an agent decides.</p>

  <h2>Compatibility needs the exact model</h2>
  <p>"Does this rack fit my bike?" has no answer until the model and version are known. The agent asks for them if they are missing, then answers from the parts compatibility data in the knowledge base.</p>

  <h2>When there is no official answer</h2>
  <p>Some pre-sales questions have no definitive answer in any document: whether a product may be used on public roads in a particular country, or whether it suits a particular use. An early draft answered a customer in Denmark with "we have no official conclusion for Denmark, so we will not judge this for you". That reply is useless to someone deciding whether to buy.</p>
  <p>The rule now: the reply gives our assessment from the facts we have, such as the product's motor power, speed limit and equipment, states what the assessment rests on, and says what the customer can do next. It does not invent a legal ruling, and it does not hide behind the absence of one.</p>

  <h2>What pre-sales passes on</h2>
  <p>A pre-sales conversation turns into an order question the moment the customer has bought. "I ordered yesterday, can I add a helmet?" belongs to the orders role, which can look the order up. The pre-sales role passes it on with the conversation attached. Complaints that arrive through the pre-sales entry are passed on the same way.</p>

  <h2>Email pre-sales</h2>
  <p>Pre-sales questions also arrive by email, and they are among the replies the AI sends on its own: a public discount code, a general policy explanation, an answer to a question about invoices before ordering. These replies make no promise the store has not already published. They pass through the "AI review" queue, where pre-sales is one of six types a reviewer checks.</p>

  <h2>The first message</h2>
  <p>On the website, 11,010 chat conversations started in the 30 days to 19 November 2026. Many customers open with a tap on a "pre-sales" button. The reply to that tap should not be a bare "how can I help?". It should say what the agent can answer, and if the customer is on a product page, start from that product.</p>

  <h2>Languages</h2>
  <p>Pre-sales questions come in the language of the storefront. The agent answers in that language from facts kept in one working language, so a specification is the same in German and in Spanish. How this works is in <a href="https://ai.dingboai.com/en/multilingual-ai-customer-service.html">multilingual AI customer service</a>.</p>

  <h2>What to measure</h2>
  <ul>
    <li>The share of pre-sales conversations answered without "please contact support".</li>
    <li>Returns whose stated reason is "not as described", traced back to a chat answer.</li>
    <li>Questions the knowledge base could not answer, collected weekly. Each one is a gap in the product page as well.</li>
    <li>Conversations where the customer asked about stock and the reply did not state it.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Stores with technical products, several versions or markets, and accessories that fit some models and not others. If a buyer needs to compare before purchasing, the answers they get decide whether the order comes back.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/ai-chat-website-mobile-app-support.html">AI chat on website and app</a> and <a href="https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html">what goes into the knowledge base</a>. The free plan maps your pre-sales questions within two working days.</p>]]></content:encoded></item>
<item><title>WISMO: Automating Where-Is-My-Order Requests With AI</title><link>https://ai.dingboai.com/en/wismo-where-is-my-order-automation.html</link><guid>https://ai.dingboai.com/en/wismo-where-is-my-order-automation.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>WISMO means where is my order. How an AI agent answers these requests in chat and email: order lookup, carrier status, failed lookups, when a person steps in.</description><content:encoded><![CDATA[<p>WISMO stands for "where is my order". It is the name support teams give to every message that asks when a parcel will arrive, why tracking has not moved, or whether an order has shipped at all. These requests are simple one at a time and heavy in volume, which makes them the first job most online stores want an AI agent to take. This is what the job requires, from a system that answers them every day in chat and email for an e-bike brand with nine storefronts.</p>

  <h2>Why an FAQ bot cannot answer WISMO</h2>
  <p>A WISMO question is about one specific parcel. "Orders ship within three working days" does not answer it. The answer has three parts, and each comes from a different place: the order, from the store; the parcel's position, from the carrier; and what the customer should expect next, from your shipping policy. An agent that cannot reach all three sends the customer to a tracking page they have already looked at.</p>

  <h2>Step 1: find the right order</h2>
  <p>The agent looks the order up by order number, or by the email address used at checkout, and confirms the match before it shows any details. With several storefronts it searches all of them and notes which store the order came from, because shipping options and delivery times differ by market.</p>

  <h2>Step 2: read the carrier, not only the store</h2>
  <p>The store knows a tracking number and a status such as "fulfilled". The carrier knows where the parcel is. The agent reads the tracking status from the tracking service and the carrier's public tracking page, and it only reads. It does not write into carrier systems.</p>

  <h2>Step 3: say it in plain words</h2>
  <p>A carrier status line is written for logistics staff. The agent turns it into an answer: where the parcel is, what the next step is, and the delivery time your policy states for that destination. It writes in the customer's language. The first sentence gives the answer, because customers read the first line and decide whether the reply helped.</p>
  <p>When the parcel is moving normally, the reply says so and stops. It does not tell the customer to contact the carrier, and it does not announce an investigation. A sentence about checking with the carrier belongs only in a reply where tracking shows a real exception.</p>

  <h2>Step 4: handle the lookup that fails</h2>
  <p>Every lookup returns one of four results, and the agent treats them differently:</p>
  <ul>
    <li><strong>Found:</strong> answer from the record.</li>
    <li><strong>Confirmed none:</strong> no order matches. Say so and ask for the order number or the checkout email.</li>
    <li><strong>Partial:</strong> the order exists and has no tracking number yet. Say that no tracking number has been issued and give the dispatch time your policy states.</li>
    <li><strong>Could not check:</strong> the carrier's page did not respond. Say that tracking could not be read right now, and give what the order record shows.</li>
  </ul>
  <p>The fourth result is where automation goes wrong. A status from an earlier lookup is history, and presenting it as the parcel's current position misleads the customer. A reply built on a failed lookup must not be sent automatically. It goes to a person.</p>

  <h2>WISMO in chat and in email</h2>
  <p>In chat the agent answers WISMO directly. A chat reply is immediate, and a tracking answer promises nothing. In the last 30 days the chat handled 15,445 conversations, and agents stepped into the chat window zero times.</p>
  <p>In email the agent drafts a reply for every message, with the order and tracking already checked. Tracking answers are among the simple replies it sends on its own, part of the 2,756 emails it sent without waiting for a person in the same 30 days. Once a person is following a ticket, the agent stops sending on that ticket.</p>

  <h2>When WISMO stops being WISMO</h2>
  <p>Some order status messages are requests for action in disguise:</p>
  <ul>
    <li>The parcel is marked delivered and the customer does not have it.</li>
    <li>Tracking has not moved for many days, or the parcel is held at customs.</li>
    <li>The parcel is being returned to the sender.</li>
    <li>The customer wants to change the address or cancel because of the delay.</li>
  </ul>
  <p>Each of these needs someone to do something: file a claim, reship, refund, change an order. The agent does not agree to any of it on its own. In chat it creates a ticket with the facts it has gathered, gives the customer the ticket number and the follow-up email address, and a person decides. In email it prepares the proposed solution with the draft, and an agent approves it. The rule is in <a href="https://ai.dingboai.com/en/ai-customer-service-escalation.html">when AI hands over</a>.</p>

  <h2>Addresses: ask only with a reason</h2>
  <p>Carriers reject some addresses, for example one with no house number. When that blocks a shipment, the reply states the reason and asks for the missing part. If the customer answers that the original address is fine, that is not a confirmation the parcel can ship. The reply explains the carrier's requirement again and asks again. Shipping to an address the carrier has rejected produces the next WISMO ticket.</p>

  <h2>What to measure</h2>
  <ul>
    <li>The share of WISMO requests answered with no person involved.</li>
    <li>How many "could not check" results occur, by carrier. A carrier that fails often needs a second source.</li>
    <li>How many customers write again within a few days about the same parcel. A second message means the first answer did not settle the question.</li>
    <li>How many WISMO conversations turned into action tickets, and how quickly those were closed.</li>
  </ul>
  <p>The integration side is covered in <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">AI chatbot order tracking</a>, and the Shopify specifics in <a href="https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html">what to connect on Shopify</a>. The free plan maps your carriers and your order system within two working days.</p>]]></content:encoded></item>
<item><title>AI for Order Changes: Address Edits and Cancellations</title><link>https://ai.dingboai.com/en/ai-order-changes-address-cancellation.html</link><guid>https://ai.dingboai.com/en/ai-order-changes-address-cancellation.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How an AI agent handles address changes and cancellation requests: check whether the order has shipped, answer at once, and route the change to a person.</description><content:encoded><![CDATA[<p>"Can I change my address?" and "Please cancel my order" are short messages with a deadline. Every hour they wait, the order moves closer to the truck, and a request that was easy at 10:00 is impossible at 16:00. Speed matters here more than in any other support request, and one fact decides everything: has the order shipped? This is how the system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> handles order changes.</p>

  <h2>The one fact: shipped or not</h2>
  <p>Before any reply, the agent reads the fulfilment status of the order. Everything branches from it:</p>
  <ul>
    <li><strong>Not shipped:</strong> the change is possible. The agent confirms what the customer wants and passes the change on.</li>
    <li><strong>Shipped:</strong> the address on the parcel cannot be edited. The reply says so in the first sentence and gives the tracking information.</li>
  </ul>
  <p>We learned how fragile this fact is. The shipping status was once passed along as a sentence of description instead of a fixed value, and the code read any description as "shipped". Over three days, 25 address-change replies were held back, and 12 of those orders had not shipped. The fix was to pass the status as one of a few fixed values and nothing else.</p>

  <h2>Address changes</h2>
  <p>For an order that has not shipped, the AI replies on its own: it confirms the new address back to the customer, and the change goes to the execution list under "needs internal confirmation", where a named person makes the edit in the order system. The AI does not edit the order itself. The reply says the change has been passed on. It does not say the change is done before a person has done it.</p>

  <h2>When the carrier rejects an address</h2>
  <p>Sometimes the request comes from our side: the carrier will not accept the address on the order because the house number is missing or the address is too long for the label. The reply states that reason and asks for the missing part. If the customer answers that the original address is fine, that is not a confirmation the order can ship. The carrier's requirement has not changed. The reply explains it again and asks again.</p>

  <h2>Cancellations before shipping</h2>
  <p>A cancellation is a decision about money, so a person makes it. The AI reads the request, checks that the order has not shipped, and drafts the first reply. That reply addresses the reason the customer gave before it confirms the cancellation. When the customer still wants to cancel, the agent approves, and the cancellation goes to the execution list.</p>
  <p>The AI does not send cancellation replies on its own. It prepares them, with the order status checked, so the agent only has to decide.</p>

  <h2>Cancellations after shipping</h2>
  <p>Once the parcel has left, a cancellation is a return. The reply says the order has shipped, gives tracking, and explains the return route under the store's policy. The steps from there are in <a href="https://ai.dingboai.com/en/ai-refund-request-handling.html">refund request handling</a>.</p>

  <h2>Changing the item</h2>
  <p>A different colour, another version, an added accessory: these are order changes with a price or stock consequence. The request goes to a person, and the reply must state stock for the requested item. In a review of 400 drafts that agents edited, 22 percent of the edits added stock status the draft had left out.</p>

  <h2>In chat</h2>
  <p>Chat is where these requests arrive first, because the customer wants an answer now. The chat AI answers the factual half immediately: whether the order has shipped. For the change itself, it creates a ticket carrying the order number, the requested change and the fulfilment status, and gives the customer the ticket number and the follow-up email address. In the 30 days to 19 November 2026, the chat AI created 1,694 tickets for requests that needed a person. It does not say "your order is cancelled". Nobody has cancelled it yet.</p>

  <h2>Where the task goes</h2>
  <p>Approved changes land in one part of the execution list: cancellations before shipping, order changes and address changes, with one owner. Every change is its own record. Closing the ticket is never a condition: the change is passed on when it is approved, whether or not the conversation is over.</p>
  <p>Changes the AI handled on its own, such as an address change before shipping, pass through the "AI review" queue, where a reviewer checks them by type. Cancellations and order changes are two of the six types in that queue.</p>

  <h2>What to measure</h2>
  <ul>
    <li>Time from the customer's message to the first reply on a change request.</li>
    <li>How many changes arrived before shipping and were still missed.</li>
    <li>How many replies said a change was done before it was.</li>
    <li>The share of cancellation requests where the customer kept the order after the first reply.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Stores that ship within a day or two of the order, sell configurable or high-value items, or deliver across borders, where a wrong address means a returned parcel and a second shipping cost.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/wismo-where-is-my-order-automation.html">where-is-my-order requests</a> and <a href="https://ai.dingboai.com/en/ai-delivery-exception-handling.html">delivery exceptions</a>. The free plan maps your order-change flow within two working days.</p>]]></content:encoded></item>
<item><title>Delivered but Not Received: AI for Delivery Exceptions</title><link>https://ai.dingboai.com/en/ai-delivery-exception-handling.html</link><guid>https://ai.dingboai.com/en/ai-delivery-exception-handling.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How an AI agent handles delivery exceptions such as delivered but not received and returned to sender: what the reply says and who checks with the carrier.</description><content:encoded><![CDATA[<p>A delivery exception is a parcel that tracking says is somewhere the customer says it is not. Marked delivered, and the customer has nothing. On its way back to the sender, and the customer never refused it. These are different from an ordinary "where is my order" message, which is answered by reading the tracking page. An exception needs someone to contact the carrier. This is how the system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> separates the two and handles the second.</p>

  <h2>Exception or status question?</h2>
  <p>The agent reads the carrier's tracking and classifies the request:</p>
  <ul>
    <li><strong>Status question:</strong> the parcel is moving, or simply slow. The reply gives the latest status and the delivery time the policy states. See <a href="https://ai.dingboai.com/en/wismo-where-is-my-order-automation.html">where-is-my-order requests</a>.</li>
    <li><strong>Exception:</strong> tracking shows delivered and the customer reports no parcel; or tracking shows the parcel being returned to the sender.</li>
  </ul>
  <p>The line matters because the two get different replies. A customer whose parcel is moving normally should not be told that the carrier is being contacted. That sentence, sent as a habit, creates an expectation of a follow-up that nobody will write. In our system the carrier-check sentence appears only when tracking shows a real exception, and the definition of an exception is written in one place.</p>

  <h2>What the reply to an exception says</h2>
  <p>Three things, in this order:</p>
  <ul>
    <li>The latest tracking event, with its date and place, as the carrier shows it.</li>
    <li>That we will check with the carrier.</li>
    <li>That we will reply within 24 to 48 hours with what the carrier says.</li>
  </ul>
  <p>The time window is stated because a customer with a missing parcel needs to know when to expect news. It is written in one place in the reply rules, and every part of the system refers to that place.</p>
  <p>The reply does not ask the customer to chase the carrier, and it does not offer a refund or a reshipment at this point. Those are decisions for a person after the carrier has answered.</p>

  <h2>The promise creates a task</h2>
  <p>"We will check with the carrier" is only true if someone does. The reply and the task are tied together: a carrier check enters the execution list under delivery exception checks, which has a named owner. When an agent approved the reply, a short message also goes to the shipping team's group chat: one paragraph, once per case. The same order is not posted twice, and a correction does not produce a second post.</p>
  <p>The AI sends this kind of reply on its own for the two clear cases, delivered-not-received and returned-to-sender. Those replies go through the "AI review" queue, and the carrier-check task is written to the execution list after the reviewer approves.</p>

  <h2>When tracking cannot be read</h2>
  <p>Carrier websites fail, and some refuse automated requests. When the lookup does not succeed, the agent has no current status. It does not write that there has been no movement since the last event it saw, because it does not know that. A reply built on a failed lookup is not sent automatically. It goes to a person.</p>
  <p>Lookups also need a limit. During one test we queried a carrier's public tracking site dozens of times within 20 minutes, and the carrier blocked the server for a quarter of an hour, which also stopped lookups for live customers. Tracking results are now cached, and tests keep their request count low.</p>

  <h2>Telling the customer before they ask</h2>
  <p>The system also sends order-status notices on its own initiative, before the customer has written, and opens a ticket for each. In the 30 days to 19 November 2026, 738 tickets were opened by such notices. One rule we corrected: the clock for "unresolved for seven days" starts when the customer first replies, not when the notice was sent. A notice the customer never answered is not an overdue case.</p>

  <h2>After the carrier answers</h2>
  <p>The carrier's answer leads to a decision: wait, reship, refund, or file a claim. The agent drafts the reply with the proposed solution, and a person approves anything that costs money or sends goods. An approved reshipment goes to the execution list for reshipments and to the after-sales team's group chat. Each decision is recorded on its own line in the ledger.</p>

  <h2>In chat</h2>
  <p>The chat reads tracking and tells the customer what it shows. When the customer says the parcel marked delivered has not arrived, the chat does not promise an investigation result or a replacement. It creates a ticket with the order, the tracking status and the customer's statement, and gives the ticket number and the follow-up email address.</p>

  <h2>What to measure</h2>
  <ul>
    <li>How many exception replies were followed by a second reply within the promised 48 hours.</li>
    <li>How many "we will check with the carrier" sentences were sent on parcels with no exception. The target is zero.</li>
    <li>Failed tracking lookups, by carrier.</li>
    <li>Exceptions by carrier and by destination country, which shows where parcels go missing.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Stores that ship large or high-value parcels, ship across borders, or use several carriers, where each missing parcel is expensive and the carrier has to be contacted by a person.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">connecting orders and shipping</a> and <a href="https://ai.dingboai.com/en/ai-order-changes-address-cancellation.html">address changes and cancellations</a>. Send us your list of carriers and the free plan maps the exception flow within two working days.</p>]]></content:encoded></item>
<item><title>AI Technical Support: Troubleshooting Hardware Products</title><link>https://ai.dingboai.com/en/ai-technical-support-troubleshooting.html</link><guid>https://ai.dingboai.com/en/ai-technical-support-troubleshooting.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How an AI agent does technical support for hardware: the fix in the first lines, steps and videos for the exact model, and what it asks the customer to do.</description><content:encoded><![CDATA[<p>Technical support for a physical product is different from answering order questions. The customer has a device that does not work, the answer depends on the exact model, and a bad answer sends them to a repair shop for something they could have fixed in two minutes. In the support system from our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>, troubleshooting is done by AI in chat and email for an e-bike brand. These are the rules that came out of that work.</p>

  <h2>Rule 1: the fix comes first</h2>
  <p>When a customer asks how to solve a problem, the reply starts with the solution: the steps to try, the way to remove or reset the part, the likely cause, what to do if the steps fail. A customer wrote that the battery lock was jammed and the battery would not come out. An early draft told them to contact their dealer. The removal method was in the knowledge base the whole time.</p>
  <p>Dealer visits, warranty verification and requests for photos are steps in a process. They come after the fix, or beside it, never in place of it.</p>

  <h2>Rule 2: the steps must survive the pipeline</h2>
  <p>Between the knowledge base and the customer, text passes through filters: one removes internal fields, one shortens, one strips links. Each can delete the answer. We had a filter that removed video links by cutting whole sentences. In the knowledge base, the component, the operation, the cause and the fix were written as one long sentence ending in a link. The filter removed all of it, and for a day one draft in four lost its troubleshooting steps.</p>
  <p>The check we now run on any change to that pipeline: take one real knowledge base entry, pass it through, and compare before and after. The business content has to arrive complete.</p>

  <h2>Rule 3: the right model, the right video</h2>
  <p>Troubleshooting steps and video guides differ by model and version. Identify the customer's model from the order, not from their description, and pick the guide for it. A video is matched by its own title. A spreadsheet tab once listed videos for several models together, and the agent sent a tail-light video for the wrong bike. Another time a charging-port fault was answered with a video about a battery light cover. When the customer says the video does not match, the agent must not send the same link again.</p>

  <h2>Rule 4: ask for what a customer can do</h2>
  <p>A customer with a new bike that would not switch on was told, in the second email, to test the battery with a multimeter. Most customers do not own one. Steps escalate in order: what can be checked by looking, then by pressing and holding, then by reseating a connector, then by a photo or a short video for the team. A tool the customer has to buy is not a troubleshooting step.</p>

  <h2>Rule 5: do not ask twice</h2>
  <p>The thread is distilled before the model sees it, into facts that include what the customer has already sent. If a video of the fault is in the thread, the reply works from it. Asking again for material the customer supplied weeks ago is the fastest way to turn a fault report into a complaint.</p>

  <h2>Rule 6: give a judgement</h2>
  <p>Often the facts point to a likely cause without proving it. The reply says so: based on these symptoms, our judgement is that the controller is at fault, and this is what we suggest. It does not say "we cannot determine the cause" and stop. It also does not present a guess as a diagnosis: it names the facts the judgement rests on.</p>

  <h2>Rule 7: answer from the knowledge base, every time</h2>
  <p>Even when the case cannot be decided yet, the reply gives what the knowledge base has: the standard answer, the process, the video, the part information. In a review of 400 drafts that agents edited, 10 percent of the edits replaced "we have no information on this" with an answer the knowledge base already contained. The fix was in the search, described in <a href="https://ai.dingboai.com/en/rag-not-finding-answers.html">why RAG does not find answers</a>.</p>

  <h2>When troubleshooting is not enough</h2>
  <p>If the steps fail, the case becomes a warranty or repair matter. The agent proposes a solution, such as a replacement part or a local repair with a reimbursement limit stated in the same email, and a person approves it. When the fault makes the product unsafe, the reply tells the customer to stop using it. The path from there is in <a href="https://ai.dingboai.com/en/ai-warranty-claim-handling.html">AI warranty claim handling</a>.</p>

  <h2>Chat and email</h2>
  <p>Chat is good at troubleshooting: the customer is in front of the product, and steps can go one at a time. The chat AI does this directly, including inside the mobile app, where 4,715 of the 15,725 conversations in the 30 days to 19 November 2026 took place. When a part or a repair is needed, the chat creates a ticket listing what has already been tried, so the email team does not start from zero.</p>
  <p>By email the AI drafts the full troubleshooting reply. Product questions after delivery are one of the six types in the "AI review" queue for replies the AI sent on its own.</p>

  <h2>Keeping the knowledge base correct</h2>
  <p>Support agents find errors in the knowledge base while reviewing drafts. A correction made by the team is kept in a corrections layer that is applied when the entry is read, so the next sync from the source document does not overwrite it. When the source is fixed, the correction is removed.</p>

  <h2>What to measure</h2>
  <ul>
    <li>Faults solved by the first reply, with no part shipped.</li>
    <li>Replies where the customer answered "that is not my model" or "I already sent that".</li>
    <li>Drafts in which an agent added a troubleshooting step. Each one is a gap in the knowledge base or the pipeline.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Brands whose products have firmware, batteries, moving parts or assembly steps, and whose support team answers the same faults in several languages.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html">what goes into the knowledge base</a> and <a href="https://ai.dingboai.com/en/ai-customer-service-hallucination.html">hallucination examples and fixes</a>. Send us your top ten faults and the free plan maps them within two working days.</p>]]></content:encoded></item>
<item><title>AI Warranty Claim Handling: From Photo to Spare Part</title><link>https://ai.dingboai.com/en/ai-warranty-claim-handling.html</link><guid>https://ai.dingboai.com/en/ai-warranty-claim-handling.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How an AI agent handles warranty claims for physical products: reading photos, checking coverage, troubleshooting first, and routing the spare part to approval.</description><content:encoded><![CDATA[<p>A warranty claim is the longest conversation a hardware brand has with a customer. It starts with "it does not work", needs photos or a video, a coverage check, a diagnosis, a decision, a part or a repair, and a record of all of it. An AI agent can carry most of that, and should carry none of the decision about who pays. This is how warranty claims run in the support system described in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>, for an e-bike brand selling in eight languages.</p>

  <h2>Step 1: read what the customer already sent</h2>
  <p>The agent reads the message, the earlier thread, and the photos and videos attached. Before it asks for anything, it lists what the customer has already provided. A model once asked a customer to film a fault again when the video had already been sent earlier in the same thread. The rule since then: ask only for material the process needs and the customer has not given, and say what it is for and how to send it.</p>

  <h2>Step 2: check coverage from records</h2>
  <p>The agent looks up the order: purchase date, model, version, the store it was bought from. It reads the warranty terms that apply to that product and market from the knowledge base. If the frame or serial number is needed and missing, that is the one thing it asks for.</p>

  <h2>Step 3: give the fix first</h2>
  <p>Most claims begin as a fault report, and many faults have a known fix. The reply leads with what the customer can do now: the troubleshooting steps, the way to remove or reset the part, the video guide for that model. A customer wrote that the battery lock was jammed and the battery could not be removed, and an early draft answered "please contact your dealer first". The knowledge base held the removal method. Dealer visits, warranty checks and requests for material are later steps. The fix comes first.</p>

  <h2>Step 4: propose a solution from a fixed list</h2>
  <p>When troubleshooting does not solve it, the agent proposes one of the named solutions:</p>
  <ul>
    <li><strong>Reship a part.</strong> The agent names the part for the customer's exact model and version.</li>
    <li><strong>Local repair.</strong> The customer has the fault fixed at a nearby shop and is reimbursed up to a limit. The reply states the limit in the same email, so the customer does not have to write again to ask.</li>
    <li><strong>Reship a missing item,</strong> when something was left out of the box.</li>
    <li><strong>Return and refund,</strong> or compensation, where policy allows.</li>
    <li><strong>Stop using the product,</strong> when the fault makes it unsafe.</li>
  </ul>
  <p>The agent proposes. It does not decide liability. Whether a fault is covered, and any amount of money, is approved by a person.</p>

  <h2>Step 5: a person approves</h2>
  <p>The support agent sees the facts, the proposed solution and the draft reply together. They send it, edit it, or change the solution. Simple warranty questions, such as how long coverage lasts, are answered by the AI directly. A reshipment or a reimbursement is not.</p>

  <h2>Step 6: the part reaches the warehouse</h2>
  <p>A claim is not finished when the email is sent. Once the agent approves a reshipment, three things happen at that moment:</p>
  <ul>
    <li>The task enters the execution list for partial refunds and spare-part reshipments, which has a named owner.</li>
    <li>A solution card is posted to the after-sales team's group chat as one message: the text on top, and under it the photos or videos from the customer's most recent email that carried any, three at most. Screenshots the company sent earlier and logos from email signatures are left out.</li>
    <li>One line is written to the shared ledger.</li>
  </ul>
  <p>Nobody retypes the claim for the warehouse, and the person packing the part sees the customer's actual photo.</p>

  <h2>Record what was finally approved</h2>
  <p>The ledger, the task and the customer's promise all use the version a person approved, never the AI's first draft. We learned this from a check of 276 recorded compensations over 14 days: 38 did not match what the agent had finally sent. In 21 of them the final reply gave no money at all. When the system cannot confirm what a person changed, it marks the line "to be verified" and gives it to a person.</p>
  <p>Each decision is its own line. An early version kept one record per ticket and dropped a second reshipment on the same ticket without a trace.</p>

  <h2>Warranty in chat</h2>
  <p>The chat answers warranty terms and walks the customer through troubleshooting. When the customer needs a part, a repair or money, the chat does not agree to it. It creates a ticket with the order, the fault description and what has been tried, gives the customer the ticket number, and the email team continues from those facts.</p>

  <h2>What the AI sends on its own</h2>
  <p>Only replies with no decision in them: an answer about coverage, a troubleshooting guide, a request for the one missing detail. Liability, amounts and structural parts wait for a person in every case.</p>

  <h2>What to measure</h2>
  <ul>
    <li>The share of claims solved by troubleshooting, with no part shipped.</li>
    <li>How many times a customer was asked for something they had already sent. The target is zero.</li>
    <li>Whether the part named in the reply matches the part shipped.</li>
    <li>Time from approval to the task appearing in the execution list.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Brands that sell physical products with a warranty and spare parts: e-bikes, scooters, power tools, appliances, electronics, furniture with fittings. The more models and versions you sell, the more the lookup and the part matching save.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/ai-technical-support-troubleshooting.html">AI technical support for hardware</a>, <a href="https://ai.dingboai.com/en/ai-refund-request-handling.html">refund request handling</a> and <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">who approves what</a>. Send us your warranty policy and the free plan maps the claim flow within two working days.</p>]]></content:encoded></item>
<item><title>Role-Based AI Agents for Customer Service: Five Roles</title><link>https://ai.dingboai.com/en/role-based-ai-agents-customer-service.html</link><guid>https://ai.dingboai.com/en/role-based-ai-agents-customer-service.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How role-based AI agents split customer service into five jobs, what each role passes on, who owns each one, and what happens when that person is offline.</description><content:encoded><![CDATA[<p>One AI assistant that "handles support" is hard to manage, because nobody can say what it is responsible for. The system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> is organised the way a support team is: by role. Each role is a job with a scope, a list of things it passes on, and a named person who answers for it. This page describes the five roles as they run today for an e-bike brand with nine storefronts.</p>

  <h2>The five roles</h2>
  <ul>
    <li><strong>Pre-sales.</strong> Models, specifications, stock, prices, offers, questions before an order ships. It passes order questions to the orders role.</li>
    <li><strong>Orders and shipping.</strong> Tracking, chasing a late parcel, address changes, lost parcels, customs, invoices. It passes damage and faults to after-sales.</li>
    <li><strong>After-sales.</strong> Fault diagnosis, spare parts, repair shops, warranty checks, replacements. It passes refunds and angry cases on.</li>
    <li><strong>Returns and refunds.</strong> Returns, exchanges, refunds, compensation amounts. It passes legal threats and repeat complaints up.</li>
    <li><strong>Escalations.</strong> Escalated complaints, legal matters, marketplace reviews, dealers. This is the final stop, and it can send a case back to another role.</li>
  </ul>
  <p>A role is not a separate chatbot. All five share one knowledge base, one order lookup and one set of reply principles. What differs is the scope: which requests the role takes, which solutions it may propose, and where its cases go when they leave its scope.</p>

  <h2>What a role does with one case</h2>
  <p>Inside every role the work has the same shape. The AI distils the incoming message and the earlier thread into structured facts with sources, checks the order, the shipment and the knowledge base, proposes a solution from a fixed list, and writes the reply in the customer's language. The list of solutions is short and named: answer the facts, ask for missing information, troubleshoot, reassure about delivery time, reship a part, reship a missing item, arrange a local repair, change an address, check with the carrier, return and refund, compensate. A solution that is not on the list cannot be proposed.</p>

  <h2>Each role has an owner</h2>
  <p>Every role has a named person responsible for it. Approved work then goes to an execution list split by what a person has to do, and each part of that list has its own owner:</p>
  <ul>
    <li>partial refunds and spare-part reshipments,</li>
    <li>return and exchange pickups,</li>
    <li>delivery exception checks with the carrier,</li>
    <li>everything else that needs internal confirmation: cancellations before shipping, order changes, address changes.</li>
  </ul>
  <p>The split follows the action a person must take, not where the request came from. A refund agreed in chat and a refund agreed by email land in the same place.</p>

  <h2>When the owner is offline</h2>
  <p>When the responsible person is offline, the AI keeps checking facts and preparing plans and drafts. It does not carry out anything that needs that person's decision. When they start work, the queue is ready to review. Simple replies that need no decision keep going out: measured on 19 November 2026 over the previous 30 days, the AI sent 3,331 email replies on its own, and half of them left within 6.3 minutes of the customer's email.</p>

  <h2>Two buttons for the agent</h2>
  <p>An agent reviewing a draft has two choices. The first sends the reply and keeps the ticket open, for cases that are still a conversation. The second sends the reply and closes the ticket, and it asks one question first: does this case need someone to do something? If yes, the agent confirms the solution record, and the task goes to the execution list and to the team's group chat at that moment. If no, the agent picks the reason, such as remote troubleshooting, and the record is filed. Every close is tied to one record either way.</p>

  <h2>A second pair of eyes on what the AI did alone</h2>
  <p>Replies the AI sends on its own fall into two kinds. A plain answer is filed and can be spot-checked. A reply that carries a solution or asks someone to act goes to an "AI review" queue first, sorted by type: pre-sales questions, partial refunds and spare parts, cancellations, order changes, invoices, and product questions after delivery. A reviewer approves each one before anything is executed, writes an improvement note when something is wrong, and can pass the case to an agent.</p>

  <h2>The AI steps back when a person steps in</h2>
  <p>Once an agent has replied on a ticket, the AI stops sending on that ticket by itself. There are two exceptions, both narrow. When the customer accepts the solution the agent proposed, the AI confirms and closes the ticket in the agent's name. When the customer writes again after the ticket was closed, it counts as a new round and is judged from the start.</p>

  <h2>Why roles beat one general assistant</h2>
  <ul>
    <li><strong>Scope can be audited.</strong> A reviewer can read a case and say whether the role stayed inside its list of solutions.</li>
    <li><strong>Errors have an address.</strong> A wrong warranty answer belongs to after-sales and its owner, not to "the AI".</li>
    <li><strong>Rollout goes role by role.</strong> Pre-sales answers can go out automatically while refunds still wait for approval.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Brands with a support team of several people, more than one sales channel or country, and cases that pass between people before they close. If one person answers everything from one inbox, start with a single role, usually orders and shipping, and add roles as volume grows.</p>
  <p>The approval lines are in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>, and the handoff between roles and people in <a href="https://ai.dingboai.com/en/ai-customer-service-escalation.html">when AI hands over</a>. The free plan maps your team's roles within two working days.</p>]]></content:encoded></item>
<item><title>CSAT for AI Customer Service: What Low Scores Meant</title><link>https://ai.dingboai.com/en/csat-ai-customer-service.html</link><guid>https://ai.dingboai.com/en/csat-ai-customer-service.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>We read every low CSAT score over two days of AI and human support replies. Most were about when the survey was sent and real delays, not about AI wording.</description><content:encoded><![CDATA[<p>When AI starts sending replies, the first question from management is whether customers like it less. The usual answer compares one average score before and after. That hides what is going on. We read every rating that came in over two days, from midday on 17 November 2026, for an e-bike brand's email support: 217 ratings, 52 of them low, meaning two points or fewer out of five. Then we grouped each rating by the message the customer was actually rating.</p>

  <h2>The four groups</h2>
  <ul>
    <li><strong>Replies the AI sent on its own:</strong> 64 ratings, 7 low. That is 11 percent.</li>
    <li><strong>Tickets an agent closed, surveyed at closing:</strong> 107 ratings, 21 low. That is 20 percent.</li>
    <li><strong>Tickets where an agent replied, the customer was silent for seven days, and the survey went out then:</strong> 31 ratings, 19 low. That is 61 percent.</li>
    <li><strong>Notices the company sent first, before the customer had written:</strong> 15 ratings, 5 low. That is 33 percent.</li>
  </ul>

  <h2>What the groups say</h2>
  <p>The AI's own replies had the lowest share of low scores. This is not because the AI writes better than people. It is because of what it is allowed to send: simple answers with no pending action. The customer asked, got the answer, and was done. Agents handle the hard cases, where a refund or a part is involved, and those rate lower whoever writes the reply.</p>
  <p>The third group is the finding. Six in ten ratings were low when the survey arrived after seven days of silence. Reading them one by one, the customers were not rating the reply. They were still waiting for a part, a refund or a follow-up, and the survey reached them in the middle of that wait. The score measured an unfinished job.</p>
  <p>Read this way, most of the 52 low ratings were about two things: a survey sent before the matter was finished, and real delays. The wording of AI replies was a small part.</p>

  <h2>Rule 1: survey when the matter is finished</h2>
  <p>Our survey has two triggers: when a ticket closes, and when a ticket is still open and the customer has not replied for seven days. One exception was added after this review. When a reply promises that a colleague will do something, such as sending an invoice separately, issuing a voucher or shipping a part, the survey waits the full seven days even if the ticket is closed, so that the promised thing can arrive first.</p>
  <p>Two further changes follow from the numbers: do not survey a customer about a notice they never asked for, and do not survey after seven days if we still owe the customer something.</p>

  <h2>Rule 2: group by the message that was rated</h2>
  <p>A rating belongs to a specific reply. Record which reply the survey followed, who sent it, AI or agent, and how long after the reply the survey went out. Without those three fields, the average mixes easy questions with hard ones and blames or credits the wrong sender.</p>

  <h2>Rule 3: read the low ones</h2>
  <p>A low score is a customer telling you something specific. Reading these 52 found one defect no dashboard showed: a reply had gone out in Vietnamese to a customer who wrote in Italian. The message was short and in capital letters, and the language detector misread it. We compared detection on 2,704 recent emails before and after the fix: 20 changed, all of them corrections.</p>
  <p>Other low ratings pointed at replies that said a colleague would follow up "separately", where the follow-up was slow. The reply was accurate. The wait behind it was the problem.</p>

  <h2>Rule 4: do not judge AI by complaints alone</h2>
  <p>Agents and managers remember the bad AI reply someone forwarded. Count instead. In our checks in November 2026 the low share for AI-sent replies was 10 to 13 percent, and for agent replies 17 to 20 percent. A fair comparison also looks at what each side handles, because the AI gets the simple cases by design.</p>

  <h2>Rule 5: watch what the customer does next</h2>
  <p>Most customers never answer a survey. Every customer who is unhappy with a reply does something: writes again, repeats the question, or opens a chat. For replies the AI sends on its own, the system looks at the customer's next message and judges whether the reply settled the matter. To work, that check has to see the AI's reply and the customer's messages before and after it. A monitor that returns the same verdict for every case is not seeing what it is supposed to judge.</p>

  <h2>How to run this review yourself</h2>
  <ul>
    <li>Take every rating from a few days. Two days gave us 217.</li>
    <li>For each, find the reply it followed, the sender, and the gap between reply and survey.</li>
    <li>Group, count the low share in each group, then read every low one.</li>
    <li>Sort what you read into: wording, wrong fact, unfinished action, delay, wrong language, survey timing.</li>
    <li>Fix the largest group first, and count again a week later.</li>
  </ul>

  <h2>Who this fits</h2>
  <p>Any team that sends a satisfaction survey and is about to let AI send replies, or already has. The review needs no new tooling beyond the three fields in Rule 2.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/measure-ai-customer-service-quality.html">how to measure AI customer service quality</a> and <a href="https://ai.dingboai.com/en/after-hours-customer-service-ai.html">after-hours customer service with AI</a>. The free plan includes how we would measure quality on your tickets, within two working days.</p>]]></content:encoded></item>
<item><title>What Is CSAT? Score, Formula and Survey Timing</title><link>https://ai.dingboai.com/en/what-is-csat.html</link><guid>https://ai.dingboai.com/en/what-is-csat.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>CSAT is the customer satisfaction score: the share of customers who rate an interaction as satisfied. The formula, the question, and when to send the survey.</description><content:encoded><![CDATA[<p>CSAT stands for customer satisfaction score. It is the share of customers who say they were satisfied with an interaction, such as a support reply. You ask one question after the interaction, the customer picks a rating, and the score is the percentage of ratings that count as satisfied.</p>

  <h2>The question</h2>
  <p>One question, asked right after the interaction: "How satisfied were you with the help you received?" The customer answers on a scale of 1 to 5, where 1 is very dissatisfied and 5 is very satisfied. A comment box is optional and worth having.</p>

  <h2>The formula</h2>
  <p>CSAT = number of satisfied ratings ÷ total number of ratings × 100.</p>
  <p>On a five-point scale, ratings of 4 and 5 count as satisfied. If 200 customers answer and 160 give a 4 or a 5, CSAT is 160 ÷ 200 × 100 = 80 percent.</p>
  <p>The score is calculated only on the customers who answered, so always report the number of ratings beside the score.</p>

  <h2>CSAT, NPS and CES</h2>
  <ul>
    <li><strong>CSAT</strong> asks about one interaction: were you satisfied with this?</li>
    <li><strong>NPS,</strong> net promoter score, asks about the relationship: how likely are you to recommend us?</li>
    <li><strong>CES,</strong> customer effort score, asks how easy it was to get the matter solved.</li>
  </ul>
  <p>For a support team, CSAT is the direct one, because it attaches to a reply the team sent.</p>

  <h2>Look at the low share as well</h2>
  <p>The satisfied share tells you the general level. The low ratings, 1 and 2, tell you what to fix. We track the low share and read every low rating.</p>
  <p>In one review of an online brand's email support, we took every rating from two days starting at midday on 17 November 2026: 217 ratings, 52 of them low. Grouped by the message the customer was rating:</p>
  <ul>
    <li>Replies sent by AI alone: 64 ratings, 7 low, which is 11 percent.</li>
    <li>Tickets closed by an agent and surveyed at closing: 107 ratings, 21 low, 20 percent.</li>
    <li>Tickets surveyed after the customer had been silent for seven days: 31 ratings, 19 low, 61 percent.</li>
    <li>Notices the company sent before the customer had written: 15 ratings, 5 low, 33 percent.</li>
  </ul>
  <p>One overall number would have hidden all of this. The third group shows the main lesson.</p>

  <h2>When to send the survey</h2>
  <p>Survey timing changes the score. In the third group above, six ratings in ten were low. Those customers were not unhappy with the reply. They were still waiting for a part, a refund or a follow-up, and the survey arrived in the middle of the wait.</p>
  <p>Rules that follow:</p>
  <ul>
    <li>Send the survey when the matter is finished, not when the reply is sent.</li>
    <li>If the reply promised something that a colleague still has to do, wait until it can have arrived. Our system waits seven days in that case.</li>
    <li>Do not survey a customer about a message they did not ask for.</li>
    <li>Send one survey per matter.</li>
  </ul>

  <h2>What a rating is attached to</h2>
  <p>Store three things with every rating: which reply it followed, who sent that reply, and how long after the reply the survey went out. Without them you cannot tell whether a low score is about a person, a policy, a delay or the survey itself.</p>

  <h2>What CSAT cannot tell you</h2>
  <ul>
    <li><strong>It counts only the customers who answer.</strong> A customer who gave up and left is not in the score.</li>
    <li><strong>It rates the outcome, not the reply.</strong> A correct "no" to a refund gets a low score.</li>
    <li><strong>It arrives late.</strong> By the time a pattern shows, many customers have already had the same reply.</li>
  </ul>
  <p>Pair it with a measure that covers every customer: what the customer does next. Do they write again with the same question? Do they open a chat? That signal exists for every reply, rated or not.</p>

  <h2>How to raise CSAT</h2>
  <p>Read the low ratings and sort them: wrong fact, unfinished action, delay, wrong language, wording, survey timing. Fix the largest group first, then count again. In our review, reading the 52 low ratings found a reply sent in the wrong language, a defect no dashboard had shown.</p>

  <h2>CSAT when AI sends replies</h2>
  <p>Compare AI-sent and agent-sent replies separately, and remember what each handles. An AI that sends only simple answers will score well, because simple answers satisfy. The full review is in <a href="https://ai.dingboai.com/en/csat-ai-customer-service.html">CSAT for AI customer service</a>, and the wider set of measures in <a href="https://ai.dingboai.com/en/measure-ai-customer-service-quality.html">how to measure AI customer service quality</a>.</p>
  <p>Our free plan includes how we would measure satisfaction on your tickets, within two working days.</p>]]></content:encoded></item>
<item><title>How an AI Support Agent Learns From Agent Edits</title><link>https://ai.dingboai.com/en/ai-learns-from-agent-edits.html</link><guid>https://ai.dingboai.com/en/ai-learns-from-agent-edits.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How to make an AI support agent learn from the edits human agents make: compare draft and sent reply, find repeated changes, verify by replay, store only rules.</description><content:encoded><![CDATA[<p>Every time a support agent edits an AI draft before sending it, they show the system how the company actually handles that case. In the 30 days to 19 November 2026, agents in the deployment from our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> edited 19,339 AI drafts. That is the most valuable training material the system has, and it is easy to lose. This is the loop we built to use it, and what it found.</p>

  <h2>Keep both versions</h2>
  <p>The loop needs two texts for every reply: what the AI drafted and what was sent. Many systems store one reply field, and the agent's edit overwrites the draft. From that day nobody can tell what the AI wrote. The AI's draft is stored in its own field and never overwritten.</p>
  <p>Before trusting any comparison, test it on replies the AI sent with no person involved. Draft and sent text must be identical there. If the comparison reports differences on those, the method is broken.</p>

  <h2>Five steps</h2>
  <ul>
    <li><strong>Mine the differences.</strong> For each edited reply, extract what the agent changed in substance: a different solution, an added fact, a removed promise. Group the changes by theme.</li>
    <li><strong>Propose.</strong> A theme that passes a threshold of repeated, independent cases is compared with the existing rules and classified: already covered, needs a revision, new, in conflict with a rule, or skip.</li>
    <li><strong>Verify.</strong> A candidate rule is tested by replaying real past cases with and without it.</li>
    <li><strong>Measure the gap.</strong> Where the practice is already in the rules, rerun the current system on the same cases and count how often it does what the agents did.</li>
    <li><strong>Diagnose.</strong> For the cases where it does not, find why.</li>
  </ul>

  <h2>What the first run found</h2>
  <p>The first run covered 300 complex after-sales replies sent by agents over seven days. Of the eight practices that passed the threshold, six were already covered by existing rules. The agents were not doing something the system had never been told. The system had been told, and did not do it consistently.</p>
  <p>That changes the work. The answer was not to add experience. It was to find out why existing rules were not followed, and the usual cause was that the same matter was described in several places with small differences.</p>

  <h2>An example: one rule in five places</h2>
  <p>When a customer has a fault repaired at a local shop, the company reimburses up to a limit. Agents kept adding the limit to drafts that had left it out. The rule existed in five places, among them the reply principles, a constraint entry, the solution descriptions and a procedure document, and they did not say the same thing. We made all five say one thing: state the limit in the same email. Replayed on the original tickets, the drafts then carried the limit, and ordinary tickets were unchanged.</p>

  <h2>Agent practice wins</h2>
  <p>When the knowledge base, a rule or a procedure document conflicts with what agents consistently do, the agents are right and the document is changed. "Consistently" is the condition: the practice shows up across several agents and many replies. One agent's one-off choice, or a reply that was simply wrong, does not count.</p>

  <h2>Store rules, never raw text</h2>
  <p>What goes into the experience library is a distilled rule that makes sense without the original ticket: in which situation, do what, with which exceptions. No order numbers, no names, no addresses, no sentences copied from a customer or an agent. Similar cases merge into one entry.</p>
  <p>We set this rule after an audit. Of 83 entries in the library, 34 had been saved as raw review comments, carrying order numbers, addresses, names, and in some cases a whole reply in English. Every new draft was searching through them. One group of entries said the opposite of the current policy. Every route that writes to the library now passes through a distilling step first.</p>

  <h2>Verify by replay, on frozen facts</h2>
  <p>A candidate change is tested on real past tickets. Two details decide whether the test means anything:</p>
  <ul>
    <li><strong>Freeze what changes over time.</strong> If the system can search internal team messages, a replay today finds the correction someone posted after the original draft. The old version then looks as good as the new one. Tools like that are frozen to the moment of the original draft.</li>
    <li><strong>Run each case twice.</strong> Model output varies from run to run. A change counts when it is clearly better across both runs of the same batch, compared side by side with the alternatives.</li>
  </ul>
  <p>A change that does not show a clear improvement is not deployed. Several of ours were dropped at this step.</p>

  <h2>Fix by subtraction</h2>
  <p>When the loop finds a gap, the first move is to delete: repeated explanations, rules that contradict each other, steps that add nothing. A new rule is added only if it is a general principle. A rule written for one ticket makes the next ticket worse. Constraints that record a past incident are kept, merged with similar ones, and their tests still have to pass.</p>

  <h2>What agents change most</h2>
  <p>In a review of 400 edited drafts: 53 percent of the edits removed a promise the agent did not want to make, 22 percent added stock status the draft had left out, and 10 percent replaced "we have no information" with an answer the knowledge base already held. Three numbers, three specific pieces of work.</p>

  <h2>Who this fits</h2>
  <p>Teams where people review AI output every day. The loop needs volume: thousands of edited replies a month give it enough repeated cases to separate a practice from an accident.</p>
  <p>Related: <a href="https://ai.dingboai.com/en/measure-ai-customer-service-quality.html">how to measure AI customer service quality</a> and <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>. The free plan describes how this loop would run on your tickets, within two working days.</p>]]></content:encoded></item>
<item><title>How to Write a Good Customer Service Email: 9 Rules</title><link>https://ai.dingboai.com/en/how-to-write-customer-service-email.html</link><guid>https://ai.dingboai.com/en/how-to-write-customer-service-email.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Nine rules for writing a good customer service email, drawn from what support agents changed in 400 drafts: answer first, give the fix, keep promises small.</description><content:encoded><![CDATA[<p>A good customer service email answers the customer's question in its first sentence, tells them what happens next, and promises only what will be done. Everything else is detail.</p>
  <p>The nine rules below come from a support operation where every reply starts as a draft and a trained agent corrects it before sending. We reviewed 400 of those corrected drafts. What the agents changed, again and again, is the list of what makes an email good.</p>

  <h2>1. Answer in the first sentence</h2>
  <p>The customer asked something. Say the answer before anything else.</p>
  <p><em>Weak:</em> "Thank you for contacting us. We understand how important your order is to you."<br><em>Better:</em> "Your order shipped on Monday and is at the carrier's depot in Lyon."</p>
  <p>Thanks and sympathy can follow. They cannot stand in for the answer.</p>

  <h2>2. Answer this email</h2>
  <p>In a long thread, the temptation is to summarise everything again. Reply to what the customer asked this time. Do not repeat a matter that was settled two emails ago, and do not skip the one question that is still open.</p>

  <h2>3. Give the fix before the process</h2>
  <p>When a customer asks how to solve a problem, give the steps first. Forms, warranty checks and requests for photos come afterwards. A customer wrote that a battery was stuck in its lock, and a draft told them to contact their dealer. The removal method was in the company's own knowledge base.</p>

  <h2>4. Promise only what will happen</h2>
  <p>This was the largest group of corrections: 53 percent of the edits removed a promise the agent did not want to make. "We will refund you today." "This will not happen again." A promise in an email is a commitment the company has to keep. Write what has been decided and what has been done. If a colleague still has to act, say who does what.</p>

  <h2>5. Say what you know, including stock and dates</h2>
  <p>22 percent of the edits added stock status the draft had left out. A customer asking about a part wants to know whether it is available. Put the concrete fact in: in stock, the date it ships, the tracking status. Leaving it out forces a second email.</p>

  <h2>6. Do not say "we have no information"</h2>
  <p>10 percent of the edits replaced "we have no information on this" with an answer that existed in the knowledge base. Before writing that you do not know, search again. If there is no definitive answer, give your assessment from the facts you have and say what it rests on.</p>
  <p><em>Weak:</em> "We cannot determine the cause."<br><em>Better:</em> "From the symptoms you describe, the charger is the likely cause. Please try the steps below, and if the light stays red, we will send a replacement charger."</p>

  <h2>7. Ask once, for what is missing, and say why</h2>
  <p>Read what the customer has already sent. Then ask only for what the process needs and does not have, explain what it is for, and say how to send it.</p>
  <p><em>Weak:</em> "Please send photos and your order details."<br><em>Better:</em> "To send the right part we need the frame number. It is stamped under the frame near the pedals. A photo of it is enough."</p>

  <h2>8. Keep the inside of the company inside</h2>
  <p>The customer does not need to know which team owns the case or what the internal status is called. "I have passed this to the warehouse team for verification" tells them nothing. "The replacement ships within three working days and you will get a tracking number by email" does. Never tell a customer to contact another department themselves.</p>

  <h2>9. Reply in the thread, in plain words</h2>
  <p>Answer on top of the customer's latest email, with the original subject line, so both sides see one conversation. Use short sentences and common words, so a customer reading in a second language understands them. Address the customer politely by name, and write in their language.</p>

  <h2>A structure that works</h2>
  <ul>
    <li><strong>Line 1:</strong> the answer or the decision.</li>
    <li><strong>Next:</strong> what happens now, and who does it.</li>
    <li><strong>Then:</strong> anything the customer must do, with the reason.</li>
    <li><strong>Last:</strong> one line of thanks or apology where it is due.</li>
  </ul>

  <h2>Check before sending</h2>
  <ul>
    <li>Does the first sentence answer the question?</li>
    <li>Do the amounts, dates and statuses match what was said earlier in the thread?</li>
    <li>Is every promise something that will be done?</li>
  </ul>

  <h2>The same rules, at volume</h2>
  <p>In the operation these rules come from, 36,789 customer emails arrived in the 30 days to 19 November 2026, and 26,451 replies were sent from an AI draft. Agents edited 19,339 of those drafts first. The rules above are what they edit toward. How the drafting works is in <a href="https://ai.dingboai.com/en/ai-reply-customer-emails-best-practices.html">AI replies to customer emails</a> and <a href="https://ai.dingboai.com/en/ai-email-reply-automation-customer-support.html">AI email reply automation</a>.</p>
  <p>If your team answers the same emails every day, the free plan shows which of them a draft would cover, within two working days.</p>]]></content:encoded></item>
<item><title>Ticketing System for Customer Support: What It Must Do</title><link>https://ai.dingboai.com/en/customer-support-ticketing-system.html</link><guid>https://ai.dingboai.com/en/customer-support-ticketing-system.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>What a ticketing system does for customer support: one ticket per matter, statuses, owners and tasks. What changes when AI drafts and sends part of the replies.</description><content:encoded><![CDATA[<p>A ticketing system turns each customer matter into one record, a ticket, that holds every message about it, shows who is responsible, and shows what state it is in. Without one, a support team works from a shared mailbox, where two people answer the same email and a third matter is answered by nobody.</p>
  <p>We built a ticket workbench for a support operation that AI and people run together. In the 30 days to 19 November 2026, it opened 8,608 tickets. This page covers what any ticketing system has to do, and what has to change when AI is part of the team.</p>

  <h2>What every ticketing system must do</h2>
  <ul>
    <li><strong>One matter, one ticket.</strong> A customer's follow-up email joins the existing ticket. It does not open a second one.</li>
    <li><strong>Every channel in one place.</strong> Email, chat and web forms arrive in the same queue.</li>
    <li><strong>An owner.</strong> Each ticket belongs to a named person.</li>
    <li><strong>A status.</strong> Waiting for us, waiting for the customer, closed.</li>
    <li><strong>The history.</strong> Everything said and done, in order, on one screen.</li>
    <li><strong>The customer's order beside the conversation.</strong> An agent should not open a second system to see what was bought.</li>
  </ul>

  <h2>Where tickets come from</h2>
  <p>Of the 8,608 tickets in our 30 days: 6,113 were opened by a customer email, 1,694 by the chat AI when a chat request needed a person, and 738 by notices the system sent to customers about their orders. The remaining 63 were opened by hand or recovered from the spam folder.</p>
  <p>The second number is specific to working with AI. The chat AI answers questions itself and creates a ticket only when a person has to act, with the facts already written in.</p>

  <h2>Merging and splitting correctly</h2>
  <p>Merging by subject line or sender alone goes wrong. Two rules we rely on:</p>
  <ul>
    <li>An email is added to a ticket when it belongs to the same thread.</li>
    <li>When a ticket was closed and the customer writes again, that is a new round on the same matter, with its own handling and its own record.</li>
  </ul>
  <p>Automated mail is filtered before it becomes a ticket. How mail is sorted is in <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a>.</p>

  <h2>What changes when AI drafts the replies</h2>
  <p>A classic ticket screen is built for typing. When AI has already gathered the facts and written a draft, the agent's job is to check and decide, and the screen should be built for that:</p>
  <ul>
    <li><strong>Facts first.</strong> The order, the tracking status and a short summary of the thread, above the draft.</li>
    <li><strong>The proposed solution, named.</strong> Reship a part, refund, reimburse a local repair. The agent can change it.</li>
    <li><strong>Two buttons.</strong> Approve and send, or approve, send and close. One question goes with them: does this need someone to carry out an action?</li>
    <li><strong>The AI draft kept.</strong> The draft is stored separately from what was sent, so the difference can be measured later.</li>
  </ul>
  <p>In our 30 days, agents sent 23,120 replies that began as an AI draft, 19,339 of them after editing.</p>

  <h2>Tickets are not tasks</h2>
  <p>A reply that says "we will send the part" creates work for someone who never sees the ticket. A ticketing system that stops at the reply loses that work. In ours, approving a reply that needs action creates a task in an execution list at that moment. The list is divided by type of work, such as reshipments, order changes and carrier checks, and each type has a named owner.</p>
  <p>Each approval also writes one line to a shared ledger. Every decision is its own line. An early version kept one record per ticket, and a second reshipment on the same ticket was dropped without a trace.</p>

  <h2>Closing is an end state, not a gate</h2>
  <p>No process waits for a ticket to close. A task is created when the decision is approved. A record is filed when the reply is sent. If follow-up steps wait for closing, anything on a ticket that stays open is never done.</p>

  <h2>A review queue for what AI sent alone</h2>
  <p>When AI sends simple replies without a person, those replies need their own place to be checked. Ours is a queue sorted by six types of reply, where a reviewer checks them. For a reply that involves an action, the task is created only after the reviewer passes it. In the 30 days, the AI sent 3,331 replies this way.</p>

  <h2>Notes and system records apart</h2>
  <p>The notes field on a ticket holds what people wrote by hand. What the system records, such as "task created" or "survey sent", goes to the ticket's activity log. Mixed together, the one note a colleague left gets buried under machine lines.</p>

  <h2>Satisfaction surveys belong to the ticket</h2>
  <p>The survey goes out when the ticket closes. When to send it is covered in <a href="https://ai.dingboai.com/en/what-is-csat.html">what is CSAT</a>.</p>

  <h2>Choosing a ticketing system</h2>
  <ul>
    <li>Does it merge by thread?</li>
    <li>Can it show order data beside the conversation?</li>
    <li>Does it have an API that an AI agent can read tickets from and write drafts to?</li>
    <li>Can a reply create a task for another team?</li>
    <li>Can you export every ticket with its history?</li>
  </ul>
  <p>If you already use a helpdesk, an AI agent can be added to it: see <a href="https://ai.dingboai.com/en/ai-agent-zendesk-freshdesk-gorgias.html">adding an AI agent to a helpdesk</a>. If you work from a shared mailbox, see <a href="https://ai.dingboai.com/en/ai-shared-inbox-support-email.html">AI for a shared support inbox</a>. Tell us what you use today and the free plan shows where AI fits, within two working days.</p>]]></content:encoded></item>
<item><title>AI Customer Service Outsourcing: How It Works</title><link>https://ai.dingboai.com/en/ai-customer-service-outsourcing.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-outsourcing.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>AI customer service outsourcing explained: what the provider runs, what your team keeps, where data stays, how a pilot starts, and paying per completed task.</description><content:encoded><![CDATA[<p>AI customer service outsourcing means a provider builds and runs an AI agent that does your support work inside your own systems, and you pay for the work it completes. It is a third option next to the two that companies already know: hiring an outsourced team, and buying software to run yourself. This is how the arrangement works in practice, based on the support operation in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>Three ways to get support work done</h2>
  <ul>
    <li><strong>Traditional outsourcing.</strong> People at another company answer your customers. You pay per agent or per hour. Quality depends on training and turnover, and every new language or time zone means more people.</li>
    <li><strong>Software you run yourself.</strong> You buy a helpdesk or chatbot subscription. You pay per seat. Your team configures it, writes the knowledge base, connects the systems and maintains all of it.</li>
    <li><strong>AI outsourcing.</strong> The provider connects an AI agent to your store, inbox and order system, runs it, and keeps it accurate as your rules change. Your team reviews and decides. You pay per completed task.</li>
  </ul>
  <p>The third option moves the engineering and the daily upkeep to the provider, and leaves the decisions with you.</p>

  <h2>What the AI takes over</h2>
  <p>In the operation we run, the AI does the reading, checking and writing:</p>
  <ul>
    <li>It answers chat questions about products, stock, shipping status, usage, troubleshooting and published policy. In the last 30 days that was 15,445 conversations, and agents stepped into the chat window zero times.</li>
    <li>It reads every incoming email, looks up the order and the shipment, searches the knowledge base, proposes a solution and writes a reply in the customer's language. 36,294 emails came in over those 30 days.</li>
    <li>It sends simple replies that need no action on its own: 2,756 in those 30 days.</li>
    <li>It sorts mail, closes automated messages, merges emails into tickets and chases overdue cases.</li>
  </ul>

  <h2>What your team keeps</h2>
  <p>Every decision that costs money or moves goods stays with your people: refunds, compensation, reshipments, replacements, returns, cancellations, and anything outside policy. The AI prepares the plan and the reply, and a person approves, edits or changes it. Above a threshold you set, a second person confirms.</p>
  <p>This is the difference from handing your inbox to an outside team. Nobody outside your company decides what your customers are given. In the same 30 days, agents approved 3,463 AI drafts without an edit and edited 19,357 before sending. The line between the two sides is described in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>.</p>

  <h2>Where your data stays</h2>
  <p>At go-live the system is installed on your own server. Customer data stays inside your network. The provider maintains the system there. This answers the first question most companies ask about outsourcing anything that touches customer records.</p>

  <h2>How it starts</h2>
  <ul>
    <li><strong>A free plan.</strong> You describe the work. Within two working days you receive a written plan: which systems get connected, what the AI handles, what stays with your people.</li>
    <li><strong>A pilot on real work.</strong> The agent is connected to your existing systems and runs real cases for 2 to 4 weeks. You see every record it handles, and your team reviews all of them at first.</li>
    <li><strong>Pay per completed task.</strong> The pilot shows how much the agent handled and how accurately. The unit price is agreed on those numbers. After that you pay monthly for completed work and can stop at any time.</li>
  </ul>
  <p>There are no seat fees and no module fees. A conversation answered, an email reply sent or a document produced is one unit. Work that waits for your decision is not billed until it is finished.</p>

  <h2>What to check in the agreement</h2>
  <ul>
    <li><strong>The definition of a completed task.</strong> A sent reply, not a draft. An executed action, not a proposed one.</li>
    <li><strong>The source of the count.</strong> Records in your own system that you can open one by one.</li>
    <li><strong>The approval lines.</strong> A written list of what the AI does directly, what a person approves and what needs a second sign-off, which you can move.</li>
    <li><strong>Failed actions.</strong> Every result is written back to the ticket, and a failed action is reported, never silent.</li>
    <li><strong>Who updates the answers.</strong> Policies and products change. Keeping the knowledge base and the reply rules current is part of the service.</li>
    <li><strong>The exit.</strong> Monthly billing, no minimum term, and the system already on your server.</li>
  </ul>

  <h2>When it is not a fit</h2>
  <p>Low volume is not worth outsourcing to an AI. Work that happens a few times a month saves little and still needs upkeep. Work that depends on knowledge held only in an experienced employee's head, and written in no system, should wait until the facts are recorded, because without them the AI can only guess.</p>

  <h2>Beyond customer service</h2>
  <p>The same arrangement applies to any work that happens every day, follows rules and lives in a system: a support email desk, a weekly voice-of-customer report built from chat and email history, or contracts, notices and certificates generated in bulk from a spreadsheet. Examples are in <a href="https://ai.dingboai.com/en/ai-agent-use-cases.html">AI agent use cases running in production</a>.</p>
  <p>The pricing logic is compared with other models in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>. Tell us which work takes the most hours on your team and the free plan arrives within two working days.</p>]]></content:encoded></item>
<item><title>AI Agent Pricing Models: Seat, Usage, Project, Outcome</title><link>https://ai.dingboai.com/en/ai-agent-pricing-models.html</link><guid>https://ai.dingboai.com/en/ai-agent-pricing-models.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>AI agent pricing models compared: per seat, per usage, fixed project and outcome-based. Who carries the risk in each, and what you must count to pay per result.</description><content:encoded><![CDATA[<p>Every AI agent pricing model answers one question: who pays when the agent does not do the job? Per seat, per usage, fixed project and outcome-based pricing give four different answers. We sell AI agents on outcome-based pricing, and the system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> runs that way, so this comparison comes from the vendor side of the table.</p>

  <h2>Per seat</h2>
  <p>You pay a monthly fee for each person who uses the software. The price is easy to budget and has nothing to do with how much work the AI finishes. A team of ten pays the same in a month when the AI drafts 20,000 replies as in a month when it drafts none that anyone sends.</p>
  <p>Seat pricing fits tools that help a person work faster. It fits badly when the point of the agent is to take the work over, because the better the agent gets, the fewer seats you need, and the vendor's income falls as its product improves.</p>

  <h2>Per usage</h2>
  <p>You pay for what the system consumes: model tokens, messages, API calls or conversations started. Cost follows activity, which is fair for infrastructure. The weakness is that activity is not results. A reply that is drafted three times and thrown away costs three times as much as one that is drafted once and sent.</p>
  <p>Usage also depends on engineering choices you do not see. In one of our cases a long email thread sent to the model as raw text used 37,604 input tokens. After the thread was distilled into structured facts first, the same case used 4,722, and the answer got better. Under usage pricing the first design earns the vendor eight times more than the second.</p>

  <h2>Fixed project price</h2>
  <p>You agree a scope and a price, the vendor builds it, and you own or license the result. You know the cost before you start. The risk sits in the scope: it is frozen on the day of signing, and a real business process keeps changing. Return rules change, a new carrier is added, a product line launches. Each change after delivery is a new quote.</p>
  <p>A fixed price also pays for delivery, not for use. The vendor is paid when the system passes acceptance, whether or not your team is still using it six months later.</p>

  <h2>Outcome-based: pay per completed task</h2>
  <p>You pay for each unit of work the agent finishes: a conversation answered, an email reply sent, a document produced. Work that is not finished is not billed. A refund that waits for your decision is not billed until it is done.</p>
  <p>The vendor carries the risk. If the agent handles little, the invoice is small. That puts the vendor's attention where yours is: on how much real work gets through and how good it is.</p>
  <p>This model has one hard requirement. "Completed" must be something both sides can count from the same records. In customer service that is straightforward, because every conversation and every sent email is already a record in a system. Over 30 days the system in our case study produced these countable units: 15,445 chat conversations answered, and 25,576 email replies sent using an AI draft, of which 2,756 were sent by the AI on its own, 3,463 were approved by an agent without an edit and 19,357 were edited by an agent and then sent.</p>

  <h2>What you must define before paying per outcome</h2>
  <ul>
    <li><strong>The unit.</strong> One conversation, one sent reply, one document. Drafts are not units: one email can be drafted several times, and only the reply that goes out is work done.</li>
    <li><strong>What counts as completed.</strong> For an email, the reply was sent. For an action such as a refund or a reshipment, the action was carried out and written back to the ticket.</li>
    <li><strong>Who approves.</strong> Anything that moves money or goods is approved by your people. A unit is complete after that approval, not before.</li>
    <li><strong>Where the count comes from.</strong> Your own system's records, visible to you line by line, not a number on the vendor's dashboard.</li>
    <li><strong>How to stop.</strong> Monthly billing with no minimum term keeps the model honest.</li>
  </ul>

  <h2>How the unit price gets set</h2>
  <p>A unit price quoted before any work has been seen is a guess. We set it after a pilot: the agent is connected to the client's existing systems and runs real cases for 2 to 4 weeks, the client's team reviews every record, and the pilot shows how much the agent handled and how accurately. The unit price is agreed on those numbers. The steps of that pilot are in <a href="https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html">how to run an AI agent pilot</a>.</p>

  <h2>Which model fits which situation</h2>
  <ul>
    <li>A tool your staff use while they keep doing the work themselves: per seat.</li>
    <li>Raw model access for a team that builds its own system: per usage.</li>
    <li>A one-off build with a process that will not change: fixed project.</li>
    <li>Daily, repetitive work that lives in a system and can be counted: outcome-based.</li>
  </ul>
  <p>For what sits behind a development quote, see <a href="https://ai.dingboai.com/en/custom-ai-agent-development-cost.html">custom AI agent development cost</a>. For turning the agent's output into a number your finance team accepts, see <a href="https://ai.dingboai.com/en/ai-agent-roi.html">how to calculate AI agent ROI</a>. Our free plan states the unit we would count for your work within two working days.</p>]]></content:encoded></item>
<item><title>AI Agent ROI: How to Calculate It From Real Work</title><link>https://ai.dingboai.com/en/ai-agent-roi.html</link><guid>https://ai.dingboai.com/en/ai-agent-roi.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How to calculate AI agent ROI from production data: count output that was actually used, convert it to hours, and track three ratios. With real numbers.</description><content:encoded><![CDATA[<p>"It saves us five people" is the most common way to state AI agent ROI, and the one that falls apart first. Headcount did not drop, order volume grew, and the claim no longer matches the payroll. Or headcount did drop, and the remaining people spend their day fixing the agent's mistakes. ROI has to be measured on production data: how much work the agent actually took, and how good that work was.</p>

  <h2>Measure output, not headcount</h2>
  <p>We use one method with every client: count the work the AI finished <strong>that was actually used</strong>, and convert it to the time a person needs for the same work. The result answers "how many hours would this take if people did all of it". It does not answer "how many people did we cut", and it should not try to. When a business is growing, the usual result of an AI agent is that the same team absorbs more volume.</p>
  <p>So the figure we publish is always hours of output, and we always say it is a conversion.</p>

  <h2>A worked example with real numbers</h2>
  <p>These numbers come from the production database of an e-bike brand selling through nine online storefronts in eight languages, counted on 17 November 2026 for the previous 30 days. The full picture is in the <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>
  <ul>
    <li><strong>Chat:</strong> 15,445 conversations containing 36,309 customer messages, all answered by AI. Agents stepped into the chat window zero times.</li>
    <li><strong>Email:</strong> 36,294 customer emails came in. 25,576 replies were sent using an AI draft.</li>
  </ul>
  <p>The conversion: 36,309 chat messages at 3 minutes each is 1,815 hours. 25,576 email replies at 5 minutes each is 2,131 hours. Together that is 3,900 hours of work in 30 days.</p>
  <p>Two choices in that calculation matter more than the arithmetic.</p>
  <p>First, the email line counts replies that were sent, not drafts that were written. The same system produced 204,910 drafts since June, and 113,538 replies were sent using a draft over the same period. One email can be drafted several times, so counting drafts inflates the result. A draft nobody sent is cost, not output.</p>
  <p>Second, the minutes per unit are yours to set, and they must be written next to the result. Three minutes per chat message and five per email are the assumptions used here. Anyone who reads the figure can replace them and redo the sum.</p>

  <h2>Three ratios that say whether the hours are real</h2>
  <p>A total can look impressive while the team quietly redoes the work. These three ratios show what is behind it.</p>
  <ul>
    <li><strong>Adoption:</strong> how much of what the AI produced was used. Low adoption means the AI is doing work nobody wants.</li>
    <li><strong>Sent without edits:</strong> of the 25,576 replies above, 2,756 were sent by the AI on its own and 3,463 were approved by an agent without changing a word. The other 19,357 were edited first. Editing takes less time than writing, and it is still time. This ratio tells you how much review effort remains.</li>
    <li><strong>Reasons for edits:</strong> what people changed. In a review of 400 edited drafts, 53 percent of the edits removed promises the agent did not want to make, 22 percent added stock status the draft had left out, and 10 percent replaced "we have no information" with an answer the knowledge base already contained. Each of those is a specific fix, and each fix moves the second ratio. More on this in <a href="https://ai.dingboai.com/en/measure-ai-customer-service-quality.html">how to measure AI customer service quality</a>.</li>
  </ul>

  <h2>Record a baseline before launch</h2>
  <p>Most projects cannot show their ROI afterwards because nobody wrote down the starting point. Before the agent goes live, record four numbers: cases handled per day, average handling time per case, the size of the backlog, and time to first reply. After launch, measure the same four the same way. The comparison then rests on records, not on impressions.</p>

  <h2>What to put on the cost side</h2>
  <ul>
    <li>Building and connecting: integrations with your order, shipping, ticket and mail systems.</li>
    <li>Running: model usage, hosting, and the work of keeping answers up to date as your rules change.</li>
    <li>Review time: the hours your team spends approving and editing. Take it from the edit ratio, not from a guess.</li>
  </ul>
  <p>ROI is the value of the hours of output, minus those three lines, divided by those three lines. Under <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">outcome-based pricing</a> the first two lines become one number, the price per completed unit multiplied by the units, which makes the sum short.</p>

  <h2>Why the calculation and the invoice should match</h2>
  <p>Because completed work can be counted this precisely, we charge by it. The plan is free. The agent is connected to your existing systems and runs real cases for 2 to 4 weeks, with your team reviewing every record. The pilot produces the volume and the three ratios above, the unit price is agreed on those numbers, and after that you pay monthly for completed work and can stop at any time. If the agent finishes little, we are paid little.</p>
  <p>If the numbers stay low after launch, the causes are in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>]]></content:encoded></item>
<item><title>AI Agent Pilot: Running a Proof of Concept on Real Work</title><link>https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html</link><guid>https://ai.dingboai.com/en/ai-agent-pilot-proof-of-concept.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>How to run an AI agent pilot that proves something: one job, a baseline, real historical cases, full human review, and three numbers that decide go or no-go.</description><content:encoded><![CDATA[<p>An AI agent demo proves that the agent can answer questions someone chose for it. A pilot has to prove something harder: that the agent can do one of your jobs, on your data, at a quality your team accepts. Every client of ours starts with a 2 to 4 week pilot on real work before paying anything. This is how we run it.</p>

  <h2>Pick one job, not a platform</h2>
  <p>A pilot that tries to cover "customer service" or "operations" ends with a slide deck. Pick one job with a clear start and finish: replying to order status emails, answering pre-sales chat, generating certificates from a spreadsheet. The job should happen every day, follow rules that can be written down, and leave a record in a system. A list of jobs that meet these conditions is in <a href="https://ai.dingboai.com/en/ai-agent-use-cases.html">AI agent use cases running in production</a>.</p>

  <h2>Record the baseline first</h2>
  <p>Before the agent touches anything, write down four numbers for the job: cases handled per day, average handling time per case, the size of the backlog, and time to first reply. Without them, the pilot ends in opinions. With them, the same four numbers measured the same way after the pilot give the answer.</p>

  <h2>Connect read-only</h2>
  <p>In a pilot the agent reads your systems and writes nothing into them. It looks up orders, shipments, tickets and documents. It does not edit an order, issue a refund or send anything to a customer without a person. This keeps the pilot safe to run on live work, and it is also how the finished system treats lookups: reading is the agent's job, changing things is a person's decision.</p>
  <p>Each lookup must return one of four results: found, confirmed that none exists, found part of it, or could not check. An agent that cannot tell "could not check" from "does not exist" gives confident wrong answers, and a pilot is the place to find that out.</p>

  <h2>Replay real history before going live</h2>
  <p>Take a few hundred real cases from the last months, including the long threads and the angry ones, and run them through the agent. Compare each output with what your team actually sent. Test questions written for the pilot hide the real problems. History shows them.</p>
  <p>One example from our own replays: a ticket with 25 back-and-forth messages, fed to the model as raw text, produced a reply that contradicted a refund the company had already promised. Distilling the thread into structured facts before the model saw it fixed the contradiction and cut the input from 37,604 tokens to 4,722. No invented test case would have found that.</p>

  <h2>Then run live, with a person reviewing everything</h2>
  <p>After the replay, the agent works on incoming cases and your team reviews every output before it goes anywhere. Nothing is sent automatically during the pilot. Reviewers do one of three things with each output: send it as it is, edit it and send it, or discard it. Those three actions are the pilot's raw data, so the review screen must record them.</p>

  <h2>The three numbers that decide</h2>
  <ul>
    <li><strong>Volume handled.</strong> How many cases the agent produced a usable output for, out of all cases of that job.</li>
    <li><strong>Sent without edits.</strong> The share your team sent unchanged. This shows how much review effort remains after launch.</li>
    <li><strong>Reasons for edits.</strong> What people changed, grouped. In a review of 400 edited drafts from a live system, 53 percent of the edits removed promises the agent did not want to make, 22 percent added stock status, and 10 percent replaced "we have no information" with an answer the knowledge base already held. Each group points to one fix.</li>
  </ul>
  <p>Fix the largest group, replay the same history, and confirm that the group shrank and nothing else got worse. A pilot that goes through this loop two or three times shows a trend, and a trend is worth more than one good week.</p>

  <h2>Set the go or no-go rule on day one</h2>
  <p>Agree before the pilot starts what result means "go". A rule written on the first day, in terms of the three numbers, cannot be bent later by enthusiasm or by disappointment. Write the stop rule as well: the result at which both sides agree the job is not ready for an agent.</p>

  <h2>What a pilot should end with</h2>
  <ul>
    <li>The baseline and the same four numbers after the pilot.</li>
    <li>Volume handled, the unedited share, and the edit reasons with what was fixed.</li>
    <li>A written line between what the AI does directly and what a person approves. See <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>.</li>
    <li>A price. We set the price per completed task on the pilot's numbers, which is why we do not quote one earlier. The reasoning is in <a href="https://ai.dingboai.com/en/ai-agent-pricing-models.html">AI agent pricing models</a>.</li>
  </ul>

  <h2>Why 2 to 4 weeks</h2>
  <p>Two weeks is the shortest period that covers a replay, a live run and one round of fixes. Four weeks covers a second round and a full monthly cycle of the business, such as a promotion or a billing run. A pilot that needs longer than that is a job that was scoped too wide.</p>
  <p>The first step is free: describe the job in one line, and within two working days you receive a written plan naming the systems to connect, what the AI handles and what stays with your people.</p>]]></content:encoded></item>
<item><title>AI Agent Use Cases: 9 Jobs Running in Production</title><link>https://ai.dingboai.com/en/ai-agent-use-cases.html</link><guid>https://ai.dingboai.com/en/ai-agent-use-cases.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Nine AI agent use cases running in production today, with volumes: chat answers, order lookups, email drafts, ticket handoff, triage, reports and documents.</description><content:encoded><![CDATA[<p>Lists of AI agent use cases tend to describe what an agent could do. This one lists nine jobs that AI agents do every day in systems we built and run, with the volumes where we have counted them. Most of the numbers come from one deployment, an e-bike brand selling through nine online storefronts in eight languages, described in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>What makes a job a good use case</h2>
  <p>Four conditions, and a job needs all of them:</p>
  <ul>
    <li>It happens every day, in volume.</li>
    <li>It follows rules that can be written down.</li>
    <li>The facts it needs live in a system the agent can read.</li>
    <li>Its results can be counted and checked.</li>
  </ul>
  <p>Each job below meets all four.</p>

  <h2>1. Answering customer chat</h2>
  <p>The agent answers questions about products, stock, shipping, usage, troubleshooting and published policy, in the customer's language. Volume in the last 30 days: 15,445 conversations with 36,309 customer messages. Agents stepped into the chat window zero times.</p>

  <h2>2. Looking up orders and shipments</h2>
  <p>The agent finds the order by number or checkout email, confirms the match, reads the carrier's tracking status and explains it in plain words. It reads and never writes: no order edits, no cancellations. Details are in <a href="https://ai.dingboai.com/en/ai-chatbot-order-tracking.html">AI chatbot order tracking</a>.</p>

  <h2>3. Drafting a reply to every customer email</h2>
  <p>For each incoming email the agent distils the thread into structured facts, checks the order, the shipment and the knowledge base, proposes a solution and writes the reply. Volume: 36,294 customer emails in 30 days, and 204,910 drafts since June. Agents approved 3,463 of the last 30 days' replies without an edit and edited 19,357 before sending.</p>

  <h2>4. Sending simple replies on its own</h2>
  <p>Replies that involve no money, no promise and no policy exception go out without waiting for a person: tracking answers, links to manuals, a thank-you that closes a finished conversation. Volume: 2,756 emails in 30 days. The agent stops doing this on a ticket once a person is following it.</p>

  <h2>5. Sorting the inbox</h2>
  <p>Across several support mailboxes the agent sorts incoming mail by what must happen next, closes automated messages, merges emails from the same customer into one ticket and chases overdue cases. The method is in <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a>.</p>

  <h2>6. Turning chat requests into tickets</h2>
  <p>When a chat customer needs an action, such as a refund, a return, a cancellation, an address change or an invoice, the agent does not agree to it. It creates a ticket carrying the facts already established, gives the customer the ticket number and the follow-up email address, and the email team takes it from there.</p>

  <h2>7. Preparing decisions for approval</h2>
  <p>For refunds, reshipments, replacements and returns, the agent prepares the plan: what happened, what the policy says, what it proposes, and the reply that would go out. A person approves, edits or changes the decision. An agent can answer "make it €30" and the plan is redone.</p>

  <h2>8. Recording what was approved</h2>
  <p>Once a person approves a solution, the agent writes it into the execution ledger with an owner and posts a short card to the team's group chat, so the people who carry it out see the task without anyone retyping it. Solutions the AI handled on its own wait in a review queue and are executed after a reviewer approves them. The ledger records the version a person finally approved, never the AI's first draft.</p>

  <h2>9. Reports and documents</h2>
  <p>Two jobs outside the inbox run the same way. A weekly voice-of-customer report reads the chat and email history and pulls out the week's pain points, selling points and unanswered questions. A bulk document job takes a spreadsheet and a template and produces one contract, notice or certificate per row as Word or PDF, sends each one and marks it as sent.</p>

  <h2>What the nine have in common</h2>
  <p>In every job the agent does the reading, checking and writing, and a person keeps any decision that costs money or moves goods. Every result is written back to a record, and a failed action is reported. That is why each job can be counted, and why we charge for these jobs by completed work.</p>
  <p>Converted to working time, the chat and email jobs alone equal 3,900 hours of work in 30 days, at 3 minutes per chat message and 5 minutes per reply sent. The method behind that figure is in <a href="https://ai.dingboai.com/en/ai-agent-roi.html">how to calculate AI agent ROI</a>.</p>

  <h2>Jobs that are not good use cases</h2>
  <ul>
    <li><strong>Low-volume work.</strong> A task done a few times a month saves little once automated and still needs upkeep.</li>
    <li><strong>Work with no system behind it.</strong> If the facts live only in an experienced employee's head, the agent can only guess.</li>
    <li><strong>Work with no countable result.</strong> If nobody can say whether a unit was done correctly, nobody can improve it or pay for it fairly.</li>
  </ul>

  <h2>How to pick your first one</h2>
  <p>List the jobs your team repeats daily. For each, write down the system the facts live in and how you would count a finished unit. The job with the highest volume and the clearest count goes first. If your job is not on this page but meets the four conditions, it is done the same way. For the choice between an agent and a fixed workflow, see <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>.</p>
  <p>Send us the job in one line and the free plan maps it within two working days.</p>]]></content:encoded></item>
<item><title>How to Build an AI Agent That Survives Real Work</title><link>https://ai.dingboai.com/en/how-to-build-an-ai-agent.html</link><guid>https://ai.dingboai.com/en/how-to-build-an-ai-agent.html</guid><pubDate>Thu, 19 Nov 2026 08:00:00 +0000</pubDate><description>Seven steps to build an AI agent for real work: pick the job, connect read-only tools, prepare facts, write rules once, set approval lines, replay, measure.</description><content:encoded><![CDATA[<p>Building an AI agent that answers a demo question takes an afternoon: a model, a prompt, one tool. Building one that handles thousands of real cases a month is a different project, and most of the work is outside the model. These are the seven steps we follow, taken from running agents on live customer service. The volumes are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>Step 1: pick one job and write down what "done" means</h2>
  <p>Not "automate support". One job: "reply to emails asking where an order is". Then define a finished case: the customer has the latest tracking status and the delivery time the policy states. If you cannot say what done means, you cannot test the agent.</p>
  <p>Pick a job that repeats daily, has real volume, and has a system behind it to look things up in. How to choose is in <a href="https://ai.dingboai.com/en/ai-agent-use-cases.html">AI agent use cases</a>.</p>

  <h2>Step 2: connect tools, read-only first</h2>
  <p>List what a person looks up to do this job: the order, the tracking page, the policy document. Each becomes a tool the model can call. Start with tools that only read. An agent that can read everything and change nothing cannot damage your data while you learn how it behaves.</p>
  <p>Design each tool to return a small, fixed answer. Fulfilment status should come back as one of a few fixed values. We once passed it as a sentence of description, and the code read every description as "shipped". Over three days, 25 address-change replies were held back, and 12 of those orders had not shipped.</p>
  <p>Separate "there is nothing" from "the lookup failed". An empty result and an error need different handling.</p>

  <h2>Step 3: prepare the facts before the model sees them</h2>
  <p>This step decides the quality more than the choice of model. Do not hand the model the whole email thread, the whole order or the whole document. Turn them into the few facts this case needs: what the customer wants now, what was promised before, what has been done, what is missing.</p>
  <p>We learned this from two cases. In a thread of 25 messages, the model did not notice that the company had already promised a full refund, and contradicted it. On an email about spare parts, sending only the order lines the email concerned cut the input from 37,604 tokens to 4,722, and all four parts in the reply were then correct.</p>

  <h2>Step 4: write the rules once</h2>
  <p>How the agent should think and speak is written in one text, and every part of the system refers to it. When the same rule exists in five places, the copies drift apart and the model follows whichever it read last. We found a reimbursement rule in five places that did not agree. Agents kept correcting the drafts until all five said the same thing.</p>
  <p>Write principles, not ready-made sentences. A model copies an example sentence word for word into every reply.</p>

  <h2>Step 5: set the approval lines</h2>
  <p>Decide, action by action, what the agent does alone and what a person approves. Our starting lines: the agent answers questions of fact and published policy alone. Anything that costs money or sends goods is prepared by the agent and approved by a person. The detail is in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">who approves what</a>.</p>
  <p>Record every approved action from the version the person approved, not from the agent's first proposal.</p>

  <h2>Step 6: replay real cases before launch</h2>
  <p>Take real past cases and run the agent on them without sending anything. Compare its output with what your people did. Three details matter:</p>
  <ul>
    <li>Run each case twice. Model output varies between runs.</li>
    <li>Freeze anything that changes over time. If the agent can search internal messages, a replay today finds corrections posted after the original case.</li>
    <li>Make the replay unable to write or send. Check that by trying a write and seeing it fail.</li>
  </ul>
  <p>A small set is enough to see the pattern. We look at five to ten cases before running more.</p>

  <h2>Step 7: launch with review, then measure what people change</h2>
  <p>At launch, a person reviews every output. Store the agent's draft and the version that was sent in separate fields. The difference between them is your measurement: in a review of 400 edited drafts, 53 percent of the edits removed a promise, 22 percent added stock status, and 10 percent replaced "we have no information" with an answer the knowledge base already held.</p>
  <p>Each number named a fix. Only when a type of reply goes out unchanged, case after case, do you let the agent send that type alone. In the 30 days to 19 November 2026, our email agent sent 3,331 replies alone and people sent 23,120 after review.</p>

  <h2>What to do when it goes wrong</h2>
  <p>Remove before you add. Look for the repeated explanation, the two rules that contradict each other, the step that adds nothing. A new rule written for one case makes the next case worse. This is described in <a href="https://ai.dingboai.com/en/ai-learns-from-agent-edits.html">how an AI agent learns from agent edits</a>.</p>

  <h2>Build it yourself or hire it out</h2>
  <p>The seven steps need someone who knows the job, someone who can connect the systems, and someone who watches the output every week. If you have all three, build. If not, the options are in <a href="https://ai.dingboai.com/en/ai-customer-service-system-build-vs-buy.html">build, open source or buy</a>. Our free plan lays out the seven steps for your job within two working days.</p>]]></content:encoded></item>
<item><title>Fiido: AI Customer Service Case Study</title><link>https://ai.dingboai.com/en/case-study-ai-customer-service.html</link><guid>https://ai.dingboai.com/en/case-study-ai-customer-service.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>How Fiido uses AI for ecommerce customer service across 9 storefronts and 8 languages: real email volumes, human approvals and the full workflow.</description><content:encoded><![CDATA[<header class="case-hero">
  <div class="eyebrow">Fiido / AI customer service case study</div>
  <p class="client">E-bike brand · Europe &amp; the US · 9 storefronts · 8 languages</p>
  <h1>AI for ecommerce customer service.<br><em>People keep the decisions.</em></h1>
  <p class="outcome">The same team absorbed the volume while AI handled the reading, checking and writing, and people moved to reviewing and deciding.</p>
  <div class="case-context"><span>Website &amp; app chat · live January 2026</span><span>4 support mailboxes · live June 2026</span><span>22 support agent accounts</span></div>
  <nav class="case-jumps" aria-label="In this case study"><a href="#workflow">Follow one email ↓</a><a href="#decisions">Who decides what ↓</a><a href="#volume">See the volume ↓</a></nav>
  <div class="period"><b>Last 30 days</b><span>18 Oct–17 Nov 2026 · counted on 17 Nov 2026</span></div>
  <dl class="case-stats">
    <div><dt>Chat conversations</dt><dd>15,445</dd><dd class="stat-detail">36,309 customer messages answered by AI. Agents stepped into the chat window 0 times.</dd></div>
    <div><dt>Customer emails received</dt><dd>36,294</dd><dd class="stat-detail">Inbound customer email across the support operation.</dd></div>
    <div><dt>Replies sent using an AI draft</dt><dd>25,576</dd><dd class="stat-detail">2,756 sent by AI on its own · 3,463 approved without an edit · 19,357 edited, then sent.</dd></div>
  </dl>
  <div class="output"><div><strong>3,900</strong><span>hours of work in 30 days</span></div><p>36,309 chat messages × 3 minutes + 25,576 emails × 5 minutes equals <b>3,900 hours of work in 30 days</b>. This is output, not headcount cut: the same team absorbed the volume and moved to reviewing and deciding.</p></div>
  <p class="source">Source: Fiido production database, read-only count on 17 Nov 2026 at 14:02 Beijing time.</p>
</header>

<section class="case-section" id="workflow">
  <div class="eyebrow">01 / From incoming email to recorded action</div>
  <h2>One email. Six steps. A person keeps the decision.</h2>
  <p class="section-intro">Each support role is a job the AI does end to end. It reads, checks and writes; a person decides whether to send a reply and whether to pay out. These are the steps behind Fiido’s AI agent customer service examples.</p>
  <ol class="email-flow">
    <li><span class="step-no">01</span><h3>Read &amp; distil</h3><p>Before any model sees an incoming email, it becomes a few lines of structured facts: this request, earlier promises, completed actions, material already supplied, verified order and shipping status, missing information and conflicts. Each line has a source and date. Earlier messages travel as a summary, never the raw thread.</p></li>
    <li><span class="step-no">02</span><h3>Check</h3><p>The AI checks the order, shipment and knowledge base: product facts, published policies, troubleshooting steps and video guides. It also consults a library of past mistakes and their correct handling.</p></li>
    <li><span class="step-no">03</span><h3>Decide a solution</h3><p>The AI proposes the handling: answer a question, troubleshoot, reship a part, offer a partial refund, arrange a return, cancel an order or check logistics. The proposal goes to the agent with the draft.</p></li>
    <li><span class="step-no">04</span><h3>Write</h3><p>The reply uses the customer’s language and one shared set of reply principles: answer the question in the first sentence, give the fix first, give a judgement even without a final answer, and never pass the customer around.</p></li>
    <li class="human-step"><span class="step-no">05 / Human decision</span><h3>Human decides</h3><p>The agent sees both draft and solution, then sends it as is, edits it or changes the decision. Simple replies that need no action go out directly from the AI.</p></li>
    <li><span class="step-no">06</span><h3>Execute &amp; record</h3><p>Once a person approves a refund, reshipment, return pickup, logistics check or cancellation, the solution enters the execution ledger with an owner, and a short card goes to the team’s group chat. AI-handled solutions wait in the “AI review” queue until a reviewer approves execution. Every reply is archived in the complaint record.</p></li>
  </ol>
</section>

<section class="case-section" id="decisions">
  <div class="eyebrow">02 / Authority stays explicit</div>
  <h2>Who decides what</h2>
  <div class="decision-lanes">
    <div class="lane direct"><span class="lane-label">AI</span><h3>AI does directly</h3><p>Answers facts and published policy, troubleshoots, looks up orders and shipping, drafts every email and sends simple replies.</p></div>
    <div class="lane approval"><span class="lane-label">Person</span><h3>A person approves</h3><p>Any money, any reshipment or replacement, returns, cancellations and anything outside policy.</p></div>
    <div class="lane review"><span class="lane-label">Reviewer</span><h3>A second person confirms</h3><p>AI-handled solutions wait for reviewer approval before execution. Preparing a solution and carrying it out are separate steps.</p></div>
  </div>
</section>

<section class="case-section">
  <div class="eyebrow">03 / Two channels, different jobs</div>
  <h2>Chat vs email</h2>
  <div class="channel-grid">
    <div class="channel"><span class="channel-tag">Website &amp; app</span><h3>Chat answers facts</h3><p>Replies go out instantly and cannot be taken back. The chat AI answers product questions, stock, shipping status, use and troubleshooting questions, and published policy.</p><p>Invoices, refunds, compensation, reshipments, exchanges, cancellations and address changes become tickets. The customer gets a ticket number and follow-up email address; the email team handles the action.</p></div>
    <div class="channel"><span class="channel-tag">4 support mailboxes</span><h3>Email prepares the action</h3><p>The AI distils the request, checks facts, proposes a solution and drafts a reply in the customer’s language. The agent sees the draft and solution together and decides what to send and approve.</p><p>Simple replies need no action and are sent directly. Approved solutions enter the execution ledger; AI-handled solutions wait for reviewer approval. Every reply goes into the complaint record.</p></div>
  </div>
</section>

<section class="case-section" id="volume">
  <div class="eyebrow">04 / The operating volume</div>
  <h2>Customer email, month by month</h2>
  <div class="volume-grid">
    <figure class="email-chart"><figcaption>Inbound customer emails · 2026</figcaption>
      <ul class="bar-chart">
        <li><span>Jun</span><div class="bar-track"><i style="width:92.1%"></i></div><b>41,246</b></li>
        <li><span>Jul</span><div class="bar-track"><i style="width:98.9%"></i></div><b>44,283</b></li>
        <li><span>Aug</span><div class="bar-track"><i style="width:91.3%"></i></div><b>40,886</b></li>
        <li><span>Sep</span><div class="bar-track"><i style="width:100%"></i></div><b>44,779</b></li>
        <li><span>Oct</span><div class="bar-track"><i style="width:86.5%"></i></div><b>38,717</b></li>
        <li class="partial"><span>Nov*</span><div class="bar-track"><i style="width:43.8%"></i></div><b>19,633</b></li>
      </ul><p class="chart-note">* November is partial: to 17 Nov.</p>
    </figure>
    <div class="since-june"><h3>Since June 2026</h3><p>1 Jun–17 Nov 2026</p><dl class="totals"><div><dt>Customer emails</dt><dd>229,544</dd></div><div><dt>AI drafts</dt><dd>204,910</dd></div><div><dt>Replies sent using an AI draft</dt><dd>113,538</dd></div><div><dt>Tickets</dt><dd>59,757</dd></div></dl></div>
  </div>
</section>

<section class="case-section">
  <div class="eyebrow">05 / Lessons from the work</div>
  <h2>What made it work</h2>
  <ol class="lessons">
    <li><h3>Distil before the model</h3><p>A ticket with 25 back-and-forth messages caused a contradiction when fed raw. Structured facts fixed it. In one case, input fell from 37,604 to 4,722 tokens.</p></li>
    <li><h3>Keep one set of reply rules</h3><p>Every step uses the same reply principles, instead of separate copies that drift apart.</p></li>
    <li><h3>Fix by subtraction</h3><p>When something breaks, delete repeated or conflicting rules first. Add only general principles.</p></li>
    <li><h3>Record the human-final version</h3><p>Ledgers and promises use what a person finally approved, never the AI’s first draft.</p></li>
    <li><h3>Supply the fact behind the judgement</h3><p>Give the model the fact used to spot the error, such as today’s date, instead of adding a rule.</p></li>
  </ol>
</section>
<section class="case-cta"><div class="eyebrow">Put AI to work in a role</div><h2>Which job should AI take on in your team?</h2><p>Start with a free plan. Pay by completed work, not by seat or project.</p><a class="btn" href="#lead">Get a free plan →</a></section>]]></content:encoded></item>
<item><title>Shopify AI Customer Service Agent: What to Connect</title><link>https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html</link><guid>https://ai.dingboai.com/en/shopify-ai-customer-service-agent.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>What a Shopify AI customer service agent needs beyond a chat widget: order lookups, carrier tracking, a knowledge base, ticket handoff and multi-store support.</description><content:encoded><![CDATA[<p>A chat widget that answers from your FAQ is the easy part of AI customer service on Shopify. The questions that fill a support inbox, such as where an order is, whether an item is in stock, why a part does not fit, or how to return something, need the agent to reach your orders, your carriers and your policies. This is what we connected for an e-bike brand selling through nine Shopify storefronts in eight languages, whose AI handled 15,445 chat conversations and drafted replies to 36,294 customer emails in the last 30 days. Details are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>.</p>

  <h2>1. Orders, read-only</h2>
  <p>The agent looks orders up in Shopify by order number or by the email used at checkout, and confirms the match before showing any details. It reads; it does not edit orders, cancel them or issue refunds. Changes to an order are actions, and actions go to a person.</p>
  <p>With several storefronts, the lookup must search all of them and keep track of which store an order came from, because shipping options, prices and policies differ by market. A customer of the German store and a customer of the UK store asking the same question can need different answers.</p>

  <h2>2. Shipping status from carriers</h2>
  <p>Shopify knows a tracking number; the carrier knows where the parcel is. The agent reads tracking status from the tracking service and the carriers' public tracking pages and explains it in plain words. It never writes into tracking or carrier systems, not even to start tracking a shipment.</p>

  <h2>3. Product facts by storefront</h2>
  <p>Specifications can differ between versions of the same product and between markets. The knowledge base names the version for each fact, and where versions differ, the product page of the store the customer bought from is the reference. Stock status matters too: in a review of 400 drafts that agents edited, 22 percent of the edits added stock status the draft had left out.</p>

  <h2>4. A knowledge base that is searched every time</h2>
  <p>Policies, troubleshooting steps, video guides and parts compatibility live in one knowledge base. The agent searches it before every reply, with one short query per question in the message, and the model decides how to use what it finds. See <a href="https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html">what goes into an AI customer service knowledge base</a>.</p>

  <h2>5. A ticket handoff from chat</h2>
  <p>On Shopify, most chat requests that need a person are about money or goods: refunds, returns, exchanges, cancellations, address changes, invoices. The chat agent does not agree to any of these. It creates a ticket in the support system, gives the customer the ticket number and the follow-up email address, and the email team handles it with a person approving each action. Agents stepped into the chat window zero times in the last 30 days.</p>

  <h2>6. Email, not only chat</h2>
  <p>For most Shopify stores, email carries more support volume than chat, and email is where refunds and returns get settled. The same agent drafts a reply for every customer email, with the order, shipping and knowledge base facts already checked, and proposes a solution for the agent to approve. 25,576 replies in the last 30 days were sent using an AI draft.</p>

  <h2>7. One language layer for every store</h2>
  <p>Each storefront writes to customers in its own language. The agent works from facts in one working language and writes the reply in the customer's language, so one team can review replies for every store. More on this in <a href="https://ai.dingboai.com/en/multilingual-ai-customer-service.html">multilingual AI customer service</a>.</p>

  <h2>What to set up first</h2>
  <ul>
    <li>Read-only order lookup across every storefront.</li>
    <li>Carrier tracking, read-only.</li>
    <li>A knowledge base with product facts by version and market.</li>
    <li>The chat-to-ticket handoff, before chat goes live.</li>
    <li>AI drafts for every email, with automatic sending switched off until the edit rate in a category drops.</li>
  </ul>
  <p>The free plan maps this to your stores and support tools in two working days.</p>]]></content:encoded></item>
<item><title>AI Email Reply Automation for Customer Support</title><link>https://ai.dingboai.com/en/ai-email-reply-automation-customer-support.html</link><guid>https://ai.dingboai.com/en/ai-email-reply-automation-customer-support.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>How to automate customer support email replies with AI: distil each email into facts, draft every reply, choose what sends itself, and measure what agents keep.</description><content:encoded><![CDATA[<p>Most AI email tools stop at "suggest a reply". The hard part of support email is not writing sentences; it is knowing what this customer already asked, what you already promised, and what their order actually looks like. Here is how we run AI email replies for a cross-border e-commerce brand that received 229,544 customer emails between June and mid-November 2026, and what the numbers say about which parts to automate.</p>

  <h2>Step 1: Distil the email before any model sees it</h2>
  <p>The easy approach is to paste the whole thread, the ticket history and a stack of knowledge-base pages into the model. It then misses what matters. In one ticket with 25 back-and-forth messages, we had already promised a full refund once the return reached the warehouse, and the model writing the next reply contradicted it. In another, a case file of hundreds of thousands of characters led the model to ask the customer to re-film a video they had already sent.</p>
  <p>So every incoming email first becomes a few lines of structured facts, produced by code and a dedicated extraction pass:</p>
  <ul>
    <li>what the customer wants in this email;</li>
    <li>what we promised before, and what has already been done;</li>
    <li>what the customer has already sent (photos, videos, order number, invoice);</li>
    <li>verified order, shipping and warranty status;</li>
    <li>what is still missing, and where facts conflict.</li>
  </ul>
  <p>Each line carries its source and date. Earlier messages travel as a summary, never as the raw thread. On one order this cut the model's input from 37,604 tokens to 4,722, and every item in the reply came out right.</p>

  <h2>Step 2: Look things up instead of guessing</h2>
  <p>With the facts in hand, the AI checks the order, the shipment and the knowledge base: product facts, published policies, troubleshooting steps and video guides. It also searches a library of past mistakes and their correct handling, so a lesson learned on one ticket applies to the next similar one. When the customer asks how to fix something, the reply leads with the fix from the knowledge base, not with "please contact your dealer".</p>

  <h2>Step 3: Draft every reply, with a proposed solution</h2>
  <p>The AI drafts a reply for every customer email, in the customer's language, and proposes how to handle the case: answer, troubleshoot, reship a part, partial refund, return, cancellation or logistics check. Since June it has written 204,910 drafts. All drafting steps follow one shared set of reply principles, kept in a single place in the code: answer in the first sentence, give the fix first, give a judgement even without a final answer, never pass the customer around. When the team changes how they want to talk to customers, they change it once and every step follows.</p>

  <h2>Step 4: Decide what sends itself</h2>
  <p>A reply that needs no action, such as a thank-you, a status answer or a policy question, can go out on its own. A reply that commits money or goods waits for a person. In the last 30 days the split looked like this:</p>
  <ul>
    <li>25,576 replies were sent using an AI draft;</li>
    <li>2,756 of them were sent by the AI on its own;</li>
    <li>3,463 were approved by an agent without a single edit;</li>
    <li>19,357 were edited by an agent and then sent.</li>
  </ul>
  <p>Widen the automatic share one category at a time, and only after the agents' edits in that category have dropped. When the AI also proposes a solution on a case it handled alone, the solution waits for a reviewer before anything is executed; we describe that loop in <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>.</p>

  <h2>Step 5: Measure what agents keep</h2>
  <p>The number that matters is not how many drafts the AI writes. It is how many drafts the team actually sends, and how much they change them. Count three groups: sent by the AI, sent unchanged after approval, and edited before sending. Then read the edits. In one review of 400 edited drafts, 53% of edits removed promises the agent did not want to make, 22% added stock status, and 10% replaced "we have no information" where the information did exist. Each pattern pointed to a fix in the material or the rules the AI receives, not to a different model.</p>
  <p>To size the work, we count 5 minutes per email an agent would otherwise write and 3 minutes per chat message. For the last 30 days that equals 3,900 hours of output. It is an output figure: the same team absorbed the volume and moved to reviewing and deciding.</p>

  <h2>What to set up first</h2>
  <ul>
    <li>Connect order, shipping and warranty lookups before writing a single prompt.</li>
    <li>Build the fact-distilling step; never pass raw threads.</li>
    <li>Put all reply principles in one place that every step reads.</li>
    <li>Start with drafts for everything and automatic sending for nothing; open categories one at a time.</li>
    <li>Log AI draft and final sent text side by side from day one.</li>
  </ul>
  <p>The flow of one email through the whole system, with all the numbers, is in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>. For the cost side, see <a href="https://ai.dingboai.com/en/custom-ai-agent-development-cost.html">custom AI agent development cost</a>.</p>]]></content:encoded></item>
<item><title>Multilingual AI Customer Service: 8 Languages, One Team</title><link>https://ai.dingboai.com/en/multilingual-ai-customer-service.html</link><guid>https://ai.dingboai.com/en/multilingual-ai-customer-service.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>How one support team answers customers in eight languages with AI: one knowledge base, replies in the customer&#x27;s language, and people reviewing in their own.</description><content:encoded><![CDATA[<p>An e-bike brand selling through nine storefronts receives customer messages in German, Italian, French, Spanish, Dutch, Polish and English, and the support team writes Chinese and English. Hiring native speakers for every market was never realistic. This is how the AI system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> handles eight languages with one team: 15,445 chat conversations and 36,294 customer emails in the last 30 days.</p>

  <h2>Reply in the customer's language, think in one language</h2>
  <p>The model reads the email in whatever language it arrives in. The facts it works from are written in one working language: what the customer wants, the order and shipping status, what was promised before. The reply is then written in the customer's language. Keeping the facts in one language means the rules, the knowledge base and the reviewers all work from the same version of the truth, and only the final text changes language.</p>

  <h2>One knowledge base, searched across languages</h2>
  <p>Product facts, policies, troubleshooting steps and video guides live in one knowledge base maintained by the team. A German customer's question about a battery lock has to find an entry written in Chinese or English. Two things made this work:</p>
  <ul>
    <li><strong>Search with short questions in the knowledge base's language.</strong> The system rewrites each question in the email as a short, single-topic query in the working language. In our tests the phrasing of the query decided the result far more than the search engine did: one short question put the right FAQ entry first.</li>
    <li><strong>Search every time.</strong> The system retrieves before every draft, rather than letting the model decide whether it needs to. The model then decides how to use what it found.</li>
  </ul>

  <h2>Reviewers work in their own language</h2>
  <p>An agent who reads Chinese cannot check a Polish draft by reading it. So the review screen shows the agent the facts, the proposed solution and the draft, with a translation into the agent's language. The agent approves the substance, the solution and the facts, and the system keeps the customer-facing text in the customer's language. In the last 30 days agents sent 3,463 AI drafts without a single edit and edited 19,357 before sending; edits concentrate on substance such as promises and stock status, not on grammar.</p>

  <h2>Chat answers instantly, in every language</h2>
  <p>Chat cannot wait for a reviewer, so the chat AI answers facts and published policy directly in the customer's language: product questions, stock, shipping status, how to use and troubleshoot. Anything that needs a person to act becomes a ticket, and the customer receives the ticket number and the follow-up email address in their own language. Agents stepped into the chat window zero times in the last 30 days.</p>

  <h2>Language rules belong in one place</h2>
  <p>Language bugs hide in small rules. A customer typing in Chinese to a Philippine storefront still expects an English reply; a customer who explicitly asks for another language should get it. Rules like these live in one shared set of reply principles that every step reads, so the email drafter, the chat and the final check never disagree about which language to use.</p>

  <h2>Where multilingual systems break</h2>
  <ul>
    <li><strong>Punctuation in extractors.</strong> A full-width colon in Chinese text once made an extractor drop a whole email. Accept both full-width and half-width forms.</li>
    <li><strong>Translated knowledge drifting apart.</strong> Keep one source per fact and translate at reply time instead of maintaining eight copies.</li>
    <li><strong>Literal translation of internal wording.</strong> Internal labels and status names must be rewritten for the customer, not translated word for word.</li>
  </ul>

  <h2>Getting started</h2>
  <p>Start with the language that brings the most volume, connect the order and shipping lookups, put the knowledge base in one working language and turn on drafts for every email before any automatic sending. For the review loop itself see <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>. The free plan maps this to your markets and languages in two working days.</p>]]></content:encoded></item>
<item><title>AI Chatbot Order Tracking: Connecting Orders and Shipping</title><link>https://ai.dingboai.com/en/ai-chatbot-order-tracking.html</link><guid>https://ai.dingboai.com/en/ai-chatbot-order-tracking.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>How an AI chatbot answers where-is-my-order questions: reading the order number safely, read-only carrier lookups, and what it hands to the email team.</description><content:encoded><![CDATA[<p>"Where is my order?" is the most common question an online shop receives, and the one an AI chatbot should answer completely on its own. Doing that well takes three connections: finding the right order, reading the shipment status, and knowing what to hand over when the answer needs a person. This is how the chat in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> does it across 15,445 conversations in the last 30 days, with agents stepping into the chat window zero times.</p>

  <h2>Find the right order</h2>
  <p>The chatbot asks for the order number or the email used at checkout and looks the order up in the shop system. The extraction step that picks the order number out of free text is where most errors start:</p>
  <ul>
    <li><strong>Other numbers look like order numbers.</strong> A customer once included a police report reference, and an early pattern took its first part as an order number and linked the case to someone else's order. Test any extraction pattern against a large sample of real messages before release.</li>
    <li><strong>Punctuation differs by language.</strong> A full-width colon in Asian-language text once caused a whole message to be dropped. Accept both forms.</li>
    <li><strong>Confirm before you reveal.</strong> Match the order to the email or name the customer gives before showing any details.</li>
  </ul>

  <h2>Read shipping status, never write it</h2>
  <p>Shipment status comes from the carrier-tracking service and the carriers' public tracking pages. The chatbot only reads. It never pushes tracking numbers into a tracking platform or a carrier system, not even to start tracking a shipment. A chatbot that can write into logistics systems can create shipments, duplicates and alerts that someone has to clean up.</p>
  <p>The answer to the customer is the status in plain words: where the parcel is, when it last moved, and what happens next. Internal status codes are rewritten for the customer, never passed through.</p>

  <h2>Answer with tools, or not at all</h2>
  <p>A fast answer without a lookup is worse than a slower correct one. Our chat does not answer order questions from memory or from the conversation alone. When the lookup is slow, the customer waits; the system allows up to 45 seconds before treating a turn as failed. The team's rule: better a timeout than a wrong status.</p>

  <h2>Hand over what needs a person</h2>
  <p>Some tracking questions turn into requests: the parcel is lost, the address must change, the order should be cancelled, a refund is wanted. Chat replies cannot be taken back, so the chatbot does not agree to any of these. It creates a ticket with the facts it has already collected, gives the customer the ticket number and the email address for follow-up, and the email team handles the rest with a person approving each action.</p>

  <h2>Write the status the way a customer reads it</h2>
  <p>Carrier statuses are written for logistics staff: "arrived at hub", "exception", "returned to sender initiated". Passed through as they are, they create a second question for every answer. The chatbot turns each status into what it means for the customer: whether the parcel is moving, whether anything is needed from them, and what the next step is. An "exception" with no movement for several days is different from a delivery attempt that failed because nobody was home, and the reply says which one it is.</p>
  <p>Statuses that need a person to look into them, such as a parcel that has stopped moving, become a logistics check ticket so the email team can contact the carrier. Normal transit statuses are answered in chat and the conversation ends there.</p>

  <h2>Questions the chatbot answers on its own</h2>
  <ul>
    <li>Has my order shipped, and with which carrier?</li>
    <li>Where is my parcel now, and when did it last move?</li>
    <li>What does this tracking status mean?</li>
    <li>What is the shipping policy for my country?</li>
  </ul>

  <h2>Questions it turns into tickets</h2>
  <ul>
    <li>My parcel is lost or damaged.</li>
    <li>Please change my address or cancel my order.</li>
    <li>I want a refund, a replacement or compensation.</li>
  </ul>
  <p>For how those tickets are sorted and approved, see <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a> and <a href="https://ai.dingboai.com/en/ai-refund-request-handling.html">AI refund request handling</a>. The free plan maps your order and shipping systems in two working days.</p>]]></content:encoded></item>
<item><title>AI Refund Request Handling: What the AI May Promise</title><link>https://ai.dingboai.com/en/ai-refund-request-handling.html</link><guid>https://ai.dingboai.com/en/ai-refund-request-handling.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>How to let AI handle refund requests without paying out by mistake: who proposes, who approves, what chat must never promise, and which version gets recorded.</description><content:encoded><![CDATA[<p>Refund requests are where AI customer service saves the most time and where a mistake costs real money. The AI can read the request, check the order, find the policy and write a clear reply in seconds. The question is which of those steps carries a promise. Here is how the support system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> splits the work.</p>

  <h2>The AI proposes, a person decides</h2>
  <p>For every refund request by email, the AI does the full preparation: it distils what the customer wants and what has already happened, checks the order and shipment, finds the relevant policy and proposes a solution, such as a partial refund, a replacement part, a return or a full refund after the return arrives. It writes the reply around that solution. A person then decides: send as is, edit, or change the solution. Money never moves on the AI's proposal alone.</p>

  <h2>Give the fix first</h2>
  <p>Many refund requests are really problem reports. A customer who writes "the battery lock is stuck, I want my money back" usually wants the bike to work. The reply leads with what the customer can do now, taken from the knowledge base and troubleshooting steps, and handles the refund question alongside. In one review of 400 edited drafts, more than half of the edits removed promises the agent did not want to make. Leading with the fix leaves fewer promises to remove.</p>

  <h2>Chat never promises a refund</h2>
  <p>A chat reply goes out instantly and cannot be taken back. So the chat AI does not agree to refunds, does not quote amounts and does not offer compensation. It explains the published refund policy, creates a ticket, and gives the customer the ticket number and the email address where the team will follow up. The email team then handles the request with the approval step above. We made this a hard rule after a chat reply agreed to reissue an invoice on its own.</p>

  <h2>Record the amount the person approved</h2>
  <p>This is the failure that stays invisible until someone reconciles the books. When the AI proposes 30 euros and the agent replies with a replacement part instead, the compensation ledger must record the replacement, not the 30 euros. For a period, our ledger took amounts from the AI's proposal. Comparing AI drafts with the replies actually sent over 14 days, 38 of 276 cases did not match: 21 sent replies gave no money at all, and 13 changed the amount or the remedy.</p>
  <p>The rule now: the ledger, the execution step, the warehouse card and any promise use only the version a person confirmed. When the system cannot tell what the person changed, it marks the record "to check" and hands it to a person instead of filling in the AI's value.</p>

  <h2>Drafts with amounts are still drafted</h2>
  <p>It is tempting to block the AI from drafting anything that mentions money. Do not. The amount affects whether a reply may be sent automatically and who must approve it; it does not stop the draft. Agents get a complete reply to review, and the approval step is where the amount is decided.</p>

  <h2>A checklist for refund handling</h2>
  <ul>
    <li>The AI prepares every refund case end to end and proposes a solution.</li>
    <li>A person approves every solution that moves money or goods.</li>
    <li>Chat explains the policy and opens a ticket; it never promises.</li>
    <li>Ledgers and cards follow the human-final version.</li>
    <li>Every decision is recorded, including a second decision on the same ticket.</li>
  </ul>
  <p>For the wider approval design see <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>, and for how tickets are sorted by action see <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a>. The free plan maps your refund flow in two working days.</p>]]></content:encoded></item>
<item><title>AI Ticket Triage: Sort Support Email by What Must Happen</title><link>https://ai.dingboai.com/en/ai-ticket-triage.html</link><guid>https://ai.dingboai.com/en/ai-ticket-triage.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>AI ticket triage that sorts support email by the action it needs, not by topic: pure replies, solutions that need approval, and work for the warehouse team.</description><content:encoded><![CDATA[<p>Most ticket triage sorts by topic: shipping, warranty, billing. That tells you who might care about a ticket, not what has to happen next. In the support system from our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>, which created 59,757 tickets between June and mid-November 2026, triage sorts by the action each ticket needs. That one change decides who touches the ticket, whether a reply can go out on its own, and what reaches the warehouse.</p>

  <h2>The first split: pure reply or solution</h2>
  <p>Every ticket the AI handles lands in one of two groups:</p>
  <ul>
    <li><strong>Pure reply.</strong> The customer needs an answer and nothing else: a status, a policy, a troubleshooting step. The reply is the whole job.</li>
    <li><strong>Reply with a solution.</strong> Something has to happen in the world: money moves, a part ships, a pickup is booked, an order is cancelled.</li>
  </ul>
  <p>Pure replies are archived when they are sent. Solutions go on to a person and, after approval, into the execution ledger. Keeping the split strict matters more than any clever category list: a refund hidden inside a "pure reply" is the most expensive mistake triage can make.</p>

  <h2>Why topic categories are not enough</h2>
  <p>Take one email: "My battery charger stopped working and I need the bike for work on Monday." By topic it is a warranty ticket. By action it can be three different things: a troubleshooting reply if the fault is a loose plug, a replacement charger shipped from the warehouse, or a logistics check if a replacement is already on its way. The topic label is the same in all three cases; the owner, the approval step and the cost are not. Sorting by action puts the ticket in front of the person who has to do something about it.</p>

  <h2>The second split: what kind of action</h2>
  <p>Solutions are sorted into the jobs the operations team actually does:</p>
  <ul>
    <li>partial refund or replacement part;</li>
    <li>return pickup;</li>
    <li>logistics check with the carrier;</li>
    <li>cancellation before delivery;</li>
    <li>other, which in practice means order or address changes.</li>
  </ul>
  <p>Each category maps to one owner and one execution list, and each item carries the person who raised it, so everyone can filter to their own. An AI-raised item is labelled as the AI's.</p>

  <h2>Approval before execution</h2>
  <p>When an agent approves a solution, it goes straight to the execution list and a short card goes to the team's group chat. When the AI handled a ticket on its own, the reply can go out, but the solution waits in an AI review queue sorted by the same categories. A reviewer approves or corrects it, and only then does it reach execution. Whatever is recorded uses the version a person confirmed, never the AI's first proposal.</p>

  <h2>Do not let triage block the work</h2>
  <p>Two rules kept triage from becoming a bottleneck:</p>
  <ul>
    <li><strong>Closing a ticket is an end state, not a gate.</strong> Recording, review and execution happen when a decision is made, whether or not the ticket is closed.</li>
    <li><strong>Every decision is recorded.</strong> A ticket with two decisions gets two records. An early version allowed one record per ticket and silently dropped a second reshipment, with no ledger line, no task and no card.</li>
  </ul>

  <h2>Sort by meaning, not by keywords</h2>
  <p>Keyword rules break on real email. Customers write "refund" when they want a replacement and "broken" when a part is missing. The triage step reads the distilled facts of the ticket, what the customer wants now and what has been done, and assigns the action category from those. Category labels and their meanings are fixed in one place, and the classifier's output format is a strict contract the rest of the system depends on.</p>

  <h2>What to measure</h2>
  <ul>
    <li>How many AI-raised solutions reviewers change, by category. A high rate means the AI is missing a fact for that category.</li>
    <li>How many solutions reach execution without a matching ledger line. This should be zero.</li>
    <li>Time from approval to execution, by category.</li>
  </ul>
  <p>For how replies themselves get drafted and approved, see <a href="https://ai.dingboai.com/en/ai-email-reply-automation-customer-support.html">AI email reply automation</a>. The free plan maps your ticket types to action categories in two working days.</p>]]></content:encoded></item>
<item><title>AI Customer Service Escalation: When AI Hands Over</title><link>https://ai.dingboai.com/en/ai-customer-service-escalation.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-escalation.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>An AI customer service escalation process that works: escalate by what the request needs, hand over the facts, and never let chat promise what a person must do.</description><content:encoded><![CDATA[<p>Most escalation rules for AI customer service are written around the AI's confidence: when it is unsure, pass the customer to a person. That rule fails in both directions. The AI is often confidently wrong about a refund, and often unsure about a troubleshooting question it could answer with one more lookup. In the support system from our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a>, escalation is decided by what the request needs, not by how sure the AI feels.</p>

  <h2>Escalate by action, not by confidence</h2>
  <p>The test is one question: does answering this require someone to do something in the world? If the customer needs information, such as product facts, published policy, an order or shipping status, or troubleshooting steps, the AI answers. If the customer needs an action, such as an invoice, a refund, compensation, a reshipment, an exchange, a cancellation, an order change or an address change, the AI escalates.</p>
  <p>This rule is easy to audit. Anyone can read a conversation and say whether it asked for information or for an action. "The AI was unsure" cannot be audited at all.</p>

  <h2>Chat and email escalate differently</h2>
  <p>A chat reply goes out the moment it is written and cannot be taken back. An email draft can wait for a person. So the two channels escalate in different ways:</p>
  <ul>
    <li><strong>Chat:</strong> when a request needs an action, the AI does not agree to it, does not suggest a remedy and does not quote an amount. It creates a ticket, tells the customer the ticket number and the email address where the team will follow up, and the conversation continues on facts only.</li>
    <li><strong>Email:</strong> every email gets a full AI draft with a proposed solution. Escalation here means approval: a person decides on anything that moves money or goods before the reply is sent.</li>
  </ul>
  <p>We made the chat rule absolute after chat replies agreed to reissue invoices on their own. In the last 30 days the chat handled 15,445 conversations and agents stepped into the chat window zero times; the requests that needed people reached them as tickets.</p>

  <h2>Hand over the facts, not the transcript</h2>
  <p>An escalation that makes the agent reread the whole conversation wastes the time the AI just saved. The ticket the chat creates carries what the AI already established: what the customer wants, the order number, the verified order and shipping status, what the customer has already sent, and what is still missing. The agent starts from a decision, not from a transcript.</p>
  <p>The same applies inside email. When a thread has many messages, earlier messages travel as a structured summary with sources and dates, never as raw text. We learned this when a model missed a refund promise buried in a 25-message thread.</p>

  <h2>Do not escalate to avoid answering</h2>
  <p>"Let me transfer you to a colleague" and "we are checking with the relevant team" are escalations that give the customer nothing. Our rule for every reply is that it answers what can be answered now and gives the customer a way forward, even when part of the request goes to a person. When the facts point to a likely cause, the reply says what that cause is and what to do next. Passing the customer around is treated as a failure, not as caution.</p>

  <h2>When a person is already on the case</h2>
  <p>Once an agent is following a ticket, the AI stops sending replies on that ticket by itself, apart from a short thank-you that closes a finished conversation. Two voices answering the same customer is worse than one slow one.</p>

  <h2>Do not let escalation block the work</h2>
  <p>Escalation adds a person, not a gate. Closing a ticket is never a precondition for anything else: an approved solution goes to the execution list when it is approved, whether or not the ticket is closed. And every decision is recorded on its own. An early version allowed one record per ticket and silently dropped a second reshipment on the same ticket.</p>

  <h2>An escalation checklist</h2>
  <ul>
    <li>Write down the actions that need a person. Everything else the AI answers.</li>
    <li>In chat, turn those actions into tickets with a number and a follow-up address.</li>
    <li>In email, draft everything and route actions to approval.</li>
    <li>Send the facts with the escalation, with sources.</li>
    <li>Stop automatic replies once a person is on the case.</li>
  </ul>
  <p>For the approval step itself, see <a href="https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html">human-in-the-loop AI agents</a>; for how escalated work is sorted, see <a href="https://ai.dingboai.com/en/ai-ticket-triage.html">AI ticket triage</a>. The free plan maps your escalation rules in two working days.</p>]]></content:encoded></item>
<item><title>Human-in-the-Loop AI Agents: Who Approves What</title><link>https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html</link><guid>https://ai.dingboai.com/en/human-in-the-loop-ai-agents.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Where the person sits in a human-in-the-loop AI agent: what the AI does alone, what a person approves and what a second person confirms, from live support work.</description><content:encoded><![CDATA[<p>"Human in the loop" is easy to say and easy to get wrong. Put the person on every step and the AI saves nobody any time. Put the person nowhere and the first wrong refund lands on your bank statement. The useful question is narrower: <strong>which decisions does a person own, and what does the AI hand them at that moment?</strong> This is how we split it on a live customer service system that handled 36,294 customer emails and 15,445 chat conversations in the last 30 days.</p>

  <h2>Draw the line at consequences, not at difficulty</h2>
  <p>Teams often hand the "easy" questions to the AI and keep the "hard" ones for people. That is the wrong axis. A hard troubleshooting question with ten steps is safe for the AI to answer, because a wrong step costs the customer one more message. A one-line "yes, we will refund you" is dangerous, because it moves money and creates a promise the company has to keep.</p>
  <p>So we split by what the action changes in the world:</p>
  <ul>
    <li><strong>The AI does directly:</strong> answering product facts and published policy, troubleshooting, order and shipping lookups, drafting every email reply, and sending simple replies that need no action.</li>
    <li><strong>A person approves:</strong> anything that moves money, any reshipment or replacement, returns, cancellations, and anything outside policy.</li>
    <li><strong>A second person confirms:</strong> solutions the AI handled on its own, before anything is executed.</li>
  </ul>
  <p>With that split, the AI does the reading, checking and writing on every case, and the person spends their time on the one decision that carries weight.</p>

  <h2>Give the approver a decision, not a pile of text</h2>
  <p>An approval step only works if the approver can decide in seconds. If they have to reread a twenty-message thread to check the AI, the loop becomes the bottleneck and people start clicking "approve" without reading. That is worse than no loop.</p>
  <p>Our approver sees three things: the reply the AI wrote, the solution it proposes (for example "reship the charger, no refund"), and the short list of facts it used, each with its source: what the customer asked this time, what was promised before, what has been done, the verified order and shipping status, and what is missing. The approver then does one of three things: send as is, edit and send, or change the decision. In the last 30 days, 3,463 AI drafts went out without a single edit and 19,357 were edited first. The edited ones are the most valuable signal we have, because each edit shows where the AI's picture of the business differs from the team's.</p>

  <h2>Chat and email need different loops</h2>
  <p>An email draft waits for a person. A chat reply goes out the moment it is written and cannot be taken back. So the same "AI proposes, person decides" pattern cannot be copied into chat.</p>
  <p>Our chat AI answers facts and published policy only. When a customer asks for anything that needs a person to act, such as an invoice, a refund, compensation, a reshipment, an exchange, a cancellation or an address change, the chat AI does not agree to it and does not quote an amount. It opens a ticket, gives the customer the ticket number and the follow-up email address, and the email team handles it with the approval loop above. Agents stepped into the chat window zero times in the last 30 days.</p>

  <h2>Record only what the person finally approved</h2>
  <p>This is the mistake that costs real money, and it is invisible until someone reconciles the books. When the AI drafts "we will refund 30 euros" and the agent changes it to "we will send a replacement part", every downstream record has to follow the agent's version: the execution ledger, the task card sent to the warehouse team, and the promise to the customer.</p>
  <p>We learned this the hard way. For a stretch of days the compensation ledger took its amounts from the AI's proposal instead of the agent's final reply. When we compared AI draft against final sent text over 14 days, 38 of 276 cases did not match: 21 final replies gave no money at all, and 13 changed the amount or the remedy. The fix was a rule that now applies everywhere: <strong>registration, execution and promises use only the version a person confirmed.</strong> When the system cannot tell what the person changed, it marks the record "to check" and hands it to a person; it never fills the gap with the AI's value.</p>

  <h2>Let the second check catch what the first one waved through</h2>
  <p>Some cases the AI resolves fully on its own: a reply plus a solution such as reshipping a small part. Those replies go out, but the solution does not execute yet. It waits in an "AI review" queue, sorted by type (partial refund or part reship, return pickup, logistics check, cancellation before delivery, other). A reviewer approves or corrects each one, and only then does it reach the execution ledger and the team's group chat. Pure replies with nothing to execute are archived directly.</p>
  <p>This keeps the speed where it is cheap, in the reply, and the human check where it matters, in the action.</p>

  <h2>When the AI is wrong, give it the fact you used</h2>
  <p>The reflex after a bad answer is to add a rule. Ask a different question first: how did I know it was wrong? Once the AI answered a question about a promotion incorrectly. We only knew the right answer because we knew today's date, and the model had not been given today's date. Giving it the date fixed the problem without a single new rule. Most approval loops improve fastest this way: each correction a person makes points to a fact the AI was missing, and connecting that fact removes the need for the correction next time. For the opposite failure, rules piling up until they contradict each other, see <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>

  <h2>A checklist for your own loop</h2>
  <ul>
    <li>List every action the AI can trigger and mark the ones that move money or goods. Those need a person.</li>
    <li>Show the approver the decision and its facts, with sources, on one screen.</li>
    <li>Treat edits as data: review them weekly and fix the missing fact behind each pattern.</li>
    <li>Keep chat to facts and policy; turn requests for action into tickets.</li>
    <li>Make every ledger and promise follow the human-final version.</li>
  </ul>
  <p>The full numbers and the flow of one email through this system are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>. If you want the same split on your own operation, the free plan maps it to your work in two working days.</p>]]></content:encoded></item>
<item><title>AI Replies to Customer Emails: 8 Best Practices</title><link>https://ai.dingboai.com/en/ai-reply-customer-emails-best-practices.html</link><guid>https://ai.dingboai.com/en/ai-reply-customer-emails-best-practices.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Eight best practices for AI replies to customer emails, from 204,910 drafts: answer first, give the fix, keep promises consistent, reply in the thread.</description><content:encoded><![CDATA[<p>Between June and mid-November 2026, the AI in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a> drafted 204,910 replies to customer emails for an e-bike brand selling in eight languages. Every draft that an agent edited before sending told us something about what a good AI reply looks like. These are the eight practices that came out of that work.</p>

  <h2>1. Answer the question in the first sentence</h2>
  <p>Customers read the first line and decide whether the email helped. If they asked for a link, the first sentence gives the link. If they asked what the box contains, the first sentence lists it. Background, apologies and context come after the answer, if at all. Drafts that opened with a summary of the customer's own message were the ones agents rewrote most often.</p>

  <h2>2. Answer this email, not the whole history</h2>
  <p>The AI reads the earlier thread as a structured summary so it does not contradict past promises. But the reply answers what the customer asked in this email. Old issues that were already answered are not repeated; open issues that the customer raises again are answered, not skipped.</p>

  <h2>3. Give the fix before the process</h2>
  <p>When a customer asks how to solve a problem, the reply leads with what they can do now: the troubleshooting steps, how to free a stuck part, how to keep using the product in the meantime, and the likely cause. Warranty checks, dealer visits and requests for photos come after. A reply that starts with "please contact your dealer" leaves the customer exactly where they were.</p>

  <h2>4. Give a judgement, even without a final answer</h2>
  <p>Some questions have no confirmed answer yet. The reply still says what the known facts point to and what the customer should do next. Refusing to give any view, or deferring entirely to "the official position", reads as indifference. The judgement must rest on real facts; it never invents a date, an amount or a status.</p>

  <h2>5. Keep promises and numbers consistent</h2>
  <p>An amount, a date or a status that the company has already stated must not change between emails without a reason. The facts given to the model include what was promised before and when, so a new reply builds on the old one. In a review of 400 edited drafts, 53 percent of edits removed promises the agent did not want to make, which is why promises with money attached always wait for a person.</p>

  <h2>6. Ask only for what is missing, and say why</h2>
  <p>Before asking for photos, an order number or an invoice, the AI checks what the customer has already sent. When something is genuinely missing, the reply asks for that one thing, says what it is for and how to send it. Asking a customer to resend a video they already sent is one of the fastest ways to lose their patience; we saw a model do exactly that after reading a very long case file.</p>

  <h2>7. Keep internal wording out</h2>
  <p>Status codes, team names, internal SKUs and process labels stay inside. The reply describes the result from the customer's side: the parcel has reached the carrier in your country, the replacement part is being packed, the refund has been issued to the original payment method. All of these reply rules live in one shared set of principles that every drafting step reads, so changing the tone of replies is a change in one place.</p>

  <h2>8. Reply in the thread</h2>
  <p>A reply belongs under the customer's latest email, with the original subject line, so the customer's mail program keeps the conversation together. Starting a new thread for an existing conversation makes the customer search for context. New threads are only for first contact, such as a reply to a web form.</p>

  <h2>Measure what agents keep</h2>
  <p>The test of these practices is how many drafts agents send without changes. In the last 30 days, 3,463 drafts went out unchanged after approval, 19,357 were edited first and 2,756 were sent by the AI on its own. Each recurring edit points to a missing fact or a rule to change in the material the AI receives. For the full drafting flow see <a href="https://ai.dingboai.com/en/ai-email-reply-automation-customer-support.html">AI email reply automation</a>.</p>]]></content:encoded></item>
<item><title>AI Customer Service Knowledge Base: What Goes In</title><link>https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-knowledge-base.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>What to put in a knowledge base for AI customer service, how to search it so the right answer comes first, and how to correct it without breaking the sync.</description><content:encoded><![CDATA[<p>The knowledge base decides most of what an AI customer service system can say. A strong model with a thin knowledge base writes polite replies that say nothing; an ordinary model with a good one answers the question. This is what the knowledge base behind our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> contains, how it is searched and how it is kept correct.</p>

  <h2>What goes in</h2>
  <ul>
    <li><strong>Product facts by model and version.</strong> Specifications differ between versions of the same product, so each fact names the version. Where versions differ, the storefront's product page is the reference.</li>
    <li><strong>Published policies, word for word.</strong> Shipping, returns, warranty, invoices. The AI may quote these directly.</li>
    <li><strong>Troubleshooting steps.</strong> The full sequence for each common fault: what to check, how to free a stuck part, how to use the product temporarily, what the likely cause is. These are the answers customers want first.</li>
    <li><strong>Video guides.</strong> A link to a short repair or setup video often replaces three paragraphs.</li>
    <li><strong>Parts and accessories.</strong> Which part fits which model, so a replacement proposal names the right item.</li>
  </ul>

  <h2>What stays out</h2>
  <p>Internal status names, team assignments and raw customer messages stay out. Lessons from individual tickets go into a separate library of past mistakes and their correct handling, written as general rules without names, order numbers or quotes. That library is searched by situation, so a lesson from one ticket applies to the next similar one without copying anyone's words.</p>

  <h2>Search before every draft</h2>
  <p>Do not let the model decide whether it needs the knowledge base. The system searches before every draft and passes the hits to the model; the model decides how to use them. In our tests, what decided retrieval quality was the query. One short question about one thing, phrased the way the FAQ is phrased, put the right entry first. Long queries that mixed several topics buried it.</p>
  <p>So each email is broken into its separate questions, each question becomes one short query, and each query is searched on its own.</p>

  <h2>Never cut content while processing it</h2>
  <p>Every filter between the knowledge base and the model is a place where answers can disappear. We once added a step that removed video links from retrieved text by cutting at sentence boundaries. Because steps were written as one long sentence joined by commas, it deleted whole troubleshooting sequences, parts, causes and next steps, in a quarter of a day's drafts. The rule since: a filter removes only its target, and any change to a processing step is checked against a real knowledge entry before and after, word for word.</p>

  <h2>Correct the copy, keep the sync</h2>
  <p>The team maintains the source documents, and the system keeps a synced copy. When a source entry is wrong, waiting for the document owner can take days. The system keeps local corrections that are applied when an entry is read, so the next sync does not overwrite them. When the source document is fixed, the correction is removed.</p>

  <h2>Keep one version of each fact</h2>
  <p>When the same fact lives in several places, such as a product sheet, an FAQ, a reply guide and a translated copy for each market, the copies drift apart and the AI quotes whichever one it finds first. Keep each fact in one entry, give it the version it applies to, and let the reply step translate it for the customer. When a fact changes, one edit updates every reply in every language, and reviewers never have to ask which copy is the current one.</p>

  <h2>Signs your knowledge base is the problem</h2>
  <ul>
    <li>Agents keep replacing "we have no information" with an answer they know. In one review this was 10% of all edits.</li>
    <li>Replies send customers to a dealer or a form instead of giving the fix.</li>
    <li>The same question gets different answers on different days.</li>
  </ul>
  <p>For retrieval failures in more depth, see <a href="https://ai.dingboai.com/en/rag-not-finding-answers.html">RAG not finding answers</a>. The free plan reviews your current knowledge base against real tickets in two working days.</p>]]></content:encoded></item>
<item><title>AI Customer Service Hallucination: Real Examples and Fixes</title><link>https://ai.dingboai.com/en/ai-customer-service-hallucination.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-hallucination.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Six hallucinations we caught in a live AI customer service system, from invented shop names to wrong order numbers, and the fix that stopped each one.</description><content:encoded><![CDATA[<p>A hallucination in customer service is not an abstract model problem. It is a reply that promises a refund nobody approved, quotes a policy that does not exist, or tells a customer to send a video they already sent. Below are six we caught in a production system that answers customer email and chat for an e-bike brand in eight languages, what caused each one, and the change that stopped it. In almost every case the cause was the input, not the model.</p>

  <h2>1. The model contradicted a promise buried in a long thread</h2>
  <p>A ticket had 25 back-and-forth messages. Somewhere in the middle, the team had promised a full refund once the return reached the warehouse. The model writing the next reply did not see it and wrote something that contradicted the promise. A stronger model given the same thread got it right, which told us what was missing: the model had all the material and could not find what mattered in it.</p>
  <p><strong>Fix:</strong> no model receives a raw thread any more. Code and a dedicated extraction step turn the history into a short list of facts, each with source and date: what the customer wants now, what we promised, what has been done, what the customer has already sent, what is missing, and what conflicts. On one order this cut the input from 37,604 tokens to 4,722 and every item in the reply came out right.</p>

  <h2>2. Rules that fought each other produced invented shop names</h2>
  <p>On a content pipeline, the prompt said "always name the place in Chinese and English", "always give a number when you mention a cost" and "never invent a business name". When the source had no shop names or prices, the first two rules forced the model to write something and the third forbade it. The model picked the first two and invented names.</p>
  <p><strong>Fix:</strong> rewrite the forcing rules to "write it when the material has it; otherwise leave it out". The invented names disappeared. The general lesson: when a model invents, look for a rule that demands output the material cannot supply.</p>

  <h2>3. A wrong answer about a promotion, caused by a missing date</h2>
  <p>A customer asked whether a promotion still applied. The AI answered wrongly. When we asked ourselves how we knew it was wrong, the answer was simple: we knew today's date and the model did not. Giving the model the current date fixed it without a single new rule.</p>
  <p><strong>Fix:</strong> after every mistake, write down the fact you used to spot it, then make sure the model receives that fact. This beats adding "be careful about promotions" to the prompt.</p>

  <h2>4. A police report number became an order number</h2>
  <p>A customer included a police report reference in their email. The order-number extractor matched the first part of it as an order number and linked the ticket to another customer's order. The AI then answered with someone else's shipping status.</p>
  <p><strong>Fix:</strong> tighten the pattern, then scan production emails for every string with the same shape before shipping the new pattern. The same scan found a second problem: a full-width colon used in Chinese text made one extractor drop an entire email. Extractors now accept both full-width and half-width punctuation.</p>

  <h2>5. "We have no information" when the information existed</h2>
  <p>In a review of 400 drafts that agents edited before sending, 10% of the edits replaced "we do not have this information" with an answer the knowledge base already contained. The model was not lying about the world; it had not been shown the right entry.</p>
  <p><strong>Fix:</strong> the system now searches the knowledge base before every draft, one short query per question in the email, and passes the hits to the model. Retrieval quality depended mostly on the query: a short, single-topic question put the right FAQ entry first.</p>

  <h2>6. An automatic send reused an old judgement</h2>
  <p>When a customer replied to an older thread, an automatic send went out with a draft written for the previous message. The reply was fluent and wrong for the new email.</p>
  <p><strong>Fix:</strong> anything sent automatically is generated fresh for the email it answers. Earlier drafts and earlier cases can inform the judgement; they are never sent as they are.</p>

  <h2>What these have in common</h2>
  <ul>
    <li>Give the model distilled facts with sources instead of raw material.</li>
    <li>Remove rules that force output the material cannot support.</li>
    <li>Supply the fact a person used to judge, such as today's date.</li>
    <li>Test extractors against real production text before release.</li>
    <li>Retrieve before every draft, with short questions.</li>
    <li>Generate automatic replies fresh, every time.</li>
  </ul>
  <p>The same system and its numbers are in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>. For the wider pattern of agents breaking in production, see <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>. If you want this applied to your own support desk, the free plan maps it to your work in two working days.</p>]]></content:encoded></item>
<item><title>How to Measure AI Customer Service Quality</title><link>https://ai.dingboai.com/en/measure-ai-customer-service-quality.html</link><guid>https://ai.dingboai.com/en/measure-ai-customer-service-quality.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Measure AI customer service by what your team keeps: sent unchanged, edited, sent by the AI, and why agents edit. The metrics we track and what each told us.</description><content:encoded><![CDATA[<p>Vendors report AI customer service quality as resolution rates and satisfaction scores. Those numbers are easy to inflate and hard to act on. The measure that changed how we built the system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">case study</a> was simpler: of the replies the AI drafts, what does the team actually keep, and what do they change?</p>

  <h2>Count three groups of sent replies</h2>
  <p>Every reply that goes out with an AI draft falls into one of three groups. In the last 30 days:</p>
  <ul>
    <li><strong>Sent by the AI on its own:</strong> 2,756.</li>
    <li><strong>Approved by an agent without a single edit:</strong> 3,463.</li>
    <li><strong>Edited by an agent before sending:</strong> 19,357.</li>
  </ul>
  <p>Together that is 25,576 replies sent using an AI draft. The edited group is large, and that is useful rather than embarrassing: it is the richest source of information about what the AI gets wrong.</p>

  <h2>Store the draft and the final text side by side</h2>
  <p>None of this works unless the system keeps the AI's original draft and the text actually sent as two separate fields. We learned this the hard way: one field was overwritten with the final text at send time, so for a while every comparison showed the AI "matching" the agent perfectly. Keep both from day one.</p>

  <h2>Read the edits, then classify them</h2>
  <p>We classified 400 edited drafts by what the agent changed:</p>
  <ul>
    <li>53 percent removed promises the agent did not want to make.</li>
    <li>22 percent added stock status the draft had left out.</li>
    <li>10 percent replaced "we have no information" with an answer the knowledge base already contained.</li>
  </ul>
  <p>Each group pointed to a fix in what the AI receives, not to a different model. Over-promising pointed to rules that pushed the model to offer solutions without checking what had been agreed. Missing stock status pointed to a lookup the AI did not have. "No information" pointed to retrieval.</p>

  <h2>Separate wording edits from substance edits</h2>
  <p>Agents also change greetings, sign-offs and word choice. Those edits are real but cheap. An edit that changes the solution, the amount, the facts or a promise is expensive, because the draft would have misled the customer or cost money. Weight the two differently, and spend improvement time on substance.</p>

  <h2>Check the records, not only the replies</h2>
  <p>A reply can be right while the record behind it is wrong. Comparing AI drafts with final replies over 14 days, 38 of 276 compensation records did not match what was actually sent, because the ledger had taken amounts from the AI's proposal. 21 of those final replies gave no money at all. Measure whether ledgers, task cards and execution steps follow the version a person approved.</p>

  <h2>Use edit rates to decide what may send itself</h2>
  <p>Automatic sending should widen one category at a time, and only where agents have stopped changing drafts in that category. Before opening a new category, count the drafts already waiting in it: they were written under the old rules and should not go out automatically. One automatic send of an older draft taught us that.</p>

  <h2>Translate volume into hours, honestly</h2>
  <p>To size the work, we count 5 minutes for each email an agent would otherwise write and 3 minutes for each chat message. For the last 30 days that equals 3,900 hours of output. It is an output figure, not a saving claim: the same team absorbed the volume and moved to reviewing and deciding.</p>

  <h2>A measurement checklist</h2>
  <ul>
    <li>Keep the AI draft and the sent text in separate fields.</li>
    <li>Report sent-by-AI, sent-unchanged and edited every week.</li>
    <li>Classify a sample of edits by what changed and fix the input behind the largest group.</li>
    <li>Audit that records follow the human-approved version.</li>
    <li>Open automatic sending only where edits have dropped.</li>
  </ul>
  <p>For the hallucinations those edits uncovered, see <a href="https://ai.dingboai.com/en/ai-customer-service-hallucination.html">AI customer service hallucination examples</a>.</p>]]></content:encoded></item>
<item><title>AI Customer Service System: Build, Open Source or Buy</title><link>https://ai.dingboai.com/en/ai-customer-service-system-build-vs-buy.html</link><guid>https://ai.dingboai.com/en/ai-customer-service-system-build-vs-buy.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Buying AI customer service software, building on open source or hiring a custom team: the real cost of each, and the four things every option still needs.</description><content:encoded><![CDATA[<p>Most teams asking this question have already tried once. They connected a model, loaded some documents, and the demo looked great. Then real customers arrived and it fell apart. The reason is the same every time: you do not need a bot that can chat. You need a customer service system that can <strong>get work done</strong>.</p>

  <h2>Ask the right question first: chat or work</h2>
  <p>A system that chats takes a week to build. Connect a model, feed it the common questions, and the tone is pleasant enough. A system that works has to look up where this order is right now, decide whether this return fits your policy, draft the reply, and hand the case to a person when it is unsure. The hard part was never making it talk. The hard part is making sure it holds the real facts, and knows when to stop talking.</p>
  <p>Once that line is clear, choosing a path makes sense. Until then, every comparison of AI customer service software compares demos.</p>

  <h2>Three paths and what each one really costs</h2>
  <h3>Buy ready-made AI customer service software</h3>
  <p>This is the fastest way to start, and it suits teams whose cases are standard and who do not need deep links into their own systems. The cost is that your decision logic lives inside someone else's templates. Connecting your own store back end, your ERP or a home-built ticket system costs extra or is not supported at all. Also read the contract for where your data is stored and for how long. Do not assume.</p>
  <h3>Build on open source</h3>
  <p>The software is free. The cost is all in people. Installing it takes a day or two. The time goes into three things after that: connecting your system interfaces, asking the business for every decision rule and writing it down as something the AI can follow, and having someone who can investigate when something goes wrong. If nobody on your team can own this for the long term, open source turns into the most expensive path.</p>
  <h3>Hire a team to build it for you</h3>
  <p>Whether this is expensive depends on how deep the team digs into your business rules. A team that only builds from the documents you hand over is selling you an open-source install with a markup. The real value is a team that sits with you and turns rules such as "when can a customer return this" and "from which day does the warranty count" into rules the system executes, and keeps correcting them after launch.</p>

  <h2>Four things you need whichever path you choose</h2>
  <ol>
    <li><strong>Connect your systems.</strong> Orders, shipping, stock and tickets have to be looked up live, not remembered from training text. Without this step everything after it is a demo.</li>
    <li><strong>A failed lookup must never come back empty.</strong> To a model, an empty result means "confirmed: nothing there", and it will tell the customer "you have no such order". Every lookup has to separate four states: found, confirmed not there, only partly found, and the lookup itself failed.</li>
    <li><strong>Start with a person approving every reply.</strong> The AI drafts, a person clicks send. The system is useful from day one, and a mistake costs one more edit, not a wrong promise to a customer.</li>
    <li><strong>Accept changes by replaying real history.</strong> A test set you write yourself only covers the cases you already thought of. Before any change goes live, replay a few hundred real conversations and compare old and new output. If it does not pass, it does not ship.</li>
  </ol>
  <p>The second and fourth points are where most projects break after launch. We go through the full list of failure causes in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>

  <h2>What a quote is really made of</h2>
  <p>Do not focus on the license fee or the model usage bill. That is the smallest part. The real cost sits in interface integration, working out the business rules, regression testing, and ongoing correction after launch. The last one is not a one-time cost. A quote that leaves these items out is either selling a demo or leaving those costs for a later invoice. We break the items down in <a href="https://ai.dingboai.com/en/custom-ai-agent-development-cost.html">custom AI agent development cost</a>.</p>
  <p>Our own model is to bill for the work the system actually takes over. The plan is free, we connect to real work for a trial run, and you settle on the numbers you see.</p>

  <h2>Real numbers to judge the scale</h2>
  <p>One of our clients is a cross-border e-commerce company. Over the last 30 days, counted directly from the production database on 2026-11-17, the AI held 15,445 chat conversations with customers, and agents stepped into the chat window zero times. On email, 36,294 customer emails came in, and 25,576 replies went out using an AI draft. The AI sent 2,756 of those on its own; agents approved the rest, many after editing.</p>
  <p>Converted at 3 minutes per chat message and 5 minutes per email, that output equals 3,900 hours of work per 30 days. <strong>This is an output figure, not a headcount cut.</strong> The same team absorbed the volume and moved to reviewing and deciding.</p>

  <h2>How to choose</h2>
  <ul>
    <li>Standard questions, no deep system links needed: buy the software.</li>
    <li>A strong in-house engineer who will own it for years: open source.</li>
    <li>Your rules are specific and your systems are your own: build it with a team that digs into those rules and stays after launch.</li>
  </ul>
  <p>Whichever you pick, check it against the four points above before you sign. The full flow of how one email moves through a live system is in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>. If you want to see which path fits your team, the free plan maps it to your work in two working days.</p>]]></content:encoded></item>
<item><title>RAG Not Finding Answers? Five Knowledge Base Problems</title><link>https://ai.dingboai.com/en/rag-not-finding-answers.html</link><guid>https://ai.dingboai.com/en/rag-not-finding-answers.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Documents uploaded, dashboard all green, yet the RAG chat cannot find the answer. Five causes we traced in production, and a fixed three-step acceptance check.</description><content:encoded><![CDATA[<p>This is the most common failure after a RAG knowledge base goes live, and the hardest to track down, because <strong>it throws no error</strong>. The upload shows success, the admin panel is green, and the chat still cannot find the answer. Below are five causes we traced on production systems, ordered from the outside in.</p>

  <h2>1. The disk filled up and the vector store went read-only</h2>
  <p>This is the outermost cause and the one people skip. Something else on the machine fills the disk, the vector database switches itself to read-only, and new content never gets written. The upload screen still says success. The symptom: old content is found, nothing added this week is.</p>
  <p>Check the disk level before anything else. It saves hours. On a self-hosted setup, set a disk alert at 80%, before this happens, not after.</p>

  <h2>2. The document is published, but the knowledge base is not attached to the app</h2>
  <p>These are two separate switches, and it is easy to flip only one. The document itself has to be in a published state, and the knowledge base has to be inside that app's retrieval scope. Do only the first, and the admin panel is all green while the content is never searched.</p>

  <h2>3. After a few turns, query rewriting pulls the question off course</h2>
  <p>Watch for this in any app that rewrites the user's question before searching. A few turns into a conversation, the rewriting step gets pulled by the earlier history and turns the user's words into a different question. Retrieval then misses. The typical symptom: the first turn finds the answer, and the same question asked a few turns later finds nothing.</p>
  <p>Log the rewritten query, run real multi-turn conversations through it and read the logs. The drift is obvious on sight.</p>

  <h2>4. Chunking cut the answer in half</h2>
  <p>A complete answer spans two chunks. Retrieval hits the first half, the second half never comes back, and the model answers half the question. This one is the sneakiest, because the system does answer, just incompletely, and the business side concludes the model is not smart enough.</p>
  <p>The fix is to chunk by meaning and by heading, so a complete entry is never split. For FAQ content the most stable form is one question and one answer per section. After changing how you chunk, re-import the content. Changing the setting without re-running the import does nothing.</p>

  <h2>5. Old and new versions sit in the knowledge base together</h2>
  <p>The old document was never taken down, the new policy went in beside it, and retrieval returns both. Which one the model follows is a coin toss. This is worse than not finding anything: the answer looks correct, and the customer receives a policy that no longer applies.</p>
  <p>The fix is not on the retrieval side. It is on the content side. Every piece of knowledge needs a clear status, live or retired, and an update has to retire the old version, not upload a second copy beside it.</p>

  <h2>Why RAG fails quietly</h2>
  <p>All five causes share one trait: every dashboard says things are fine. Upload status, indexing status and app status are each green on their own, while the path from a real question to the right paragraph is broken somewhere between them. That is why checking the admin panel tells you nothing. The only test that counts is asking the question the way a user asks it and reading what came back.</p>
  <p>The same pattern shows up beyond retrieval. When a lookup fails and the model receives an empty result, it reads "nothing found" as a confirmed fact and answers with confidence. We cover that and other production failures in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>

  <h2>Make acceptance a fixed three-step check</h2>
  <ol>
    <li>Ask <strong>a real question inside the app</strong> and confirm the new content is in the retrieved results. Do not trust the check mark in the knowledge base admin.</li>
    <li>Print the retrieved text and read it. Confirm it hit the complete passage, not half of it.</li>
    <li>Every time you add knowledge, re-run questions the system already answered correctly, and confirm the new content did not push the right answers out.</li>
  </ol>
  <p>These three steps take ten minutes. They save you from finding out two weeks after launch that customers have been receiving an outdated policy the whole time.</p>

  <h2>Where this matters most</h2>
  <p>In customer service, a knowledge base gap turns straight into a wrong answer to a customer. On the live system in our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>, the AI searches the knowledge base for product facts, published policies, troubleshooting steps and video guides on every email before drafting a reply, so retrieval quality decides reply quality. If your knowledge base is live and still not answering, the free plan maps where it breaks in two working days.</p>]]></content:encoded></item>
<item><title>Private LLM Deployment: When On-Premise Is Worth It</title><link>https://ai.dingboai.com/en/private-llm-deployment.html</link><guid>https://ai.dingboai.com/en/private-llm-deployment.html</guid><pubDate>Tue, 17 Nov 2026 08:00:00 +0000</pubDate><description>Before a private LLM deployment, decide whether you need a private model or private data. Most businesses need the second, which costs far less.</description><content:encoded><![CDATA[<p>When a client opens with "we need a private deployment", our first question is: do you want to keep the model private, or the data? Teams treat the two as one thing. Their cost differs by an order of magnitude.</p>

  <h2>Private data and a private model are two different projects</h2>
  <p><strong>Private data</strong> means your orders, customer records and internal documents stay on your own servers. Search and storage never leave. Only the small piece of content needed for the current request is sent to the model with that request. Most businesses can afford this, and it already solves the problem they were worried about.</p>
  <p><strong>A private model</strong> means the model itself runs on your own machines. That takes GPUs, a server room, and someone who manages versions and capacity for the long term. It is a different scale of project.</p>
  <p>In our experience, eight out of ten clients who ask for a private deployment want the first one.</p>

  <h2>When an on-premise LLM is the right call</h2>
  <p>Bring the model in-house when one of these is true:</p>
  <ul>
    <li>A contract or an industry regulator states that data must not leave the country or must not be processed by a third party.</li>
    <li>The content itself is the core asset: unpublished formulas, engineering drawings, source code.</li>
    <li>Your usage is so high that the API bill exceeds the full cost of running your own. <strong>Do this math for real</strong>, item by item. The items are listed below.</li>
  </ul>
  <p>If none of the three applies, a self-hosted LLM for business adds cost and work without removing a real risk.</p>

  <h2>On-premise LLM vs cloud: the costs beyond GPUs</h2>
  <p>The GPU is the most visible cost, and it is not the largest one. Add these:</p>
  <ul>
    <li>Power and server room space.</li>
    <li>Someone on call.</li>
    <li>Re-testing your whole business flow every time the model is upgraded.</li>
    <li>Scaling out when concurrent use grows.</li>
    <li>No vendor to fall back on when something breaks. This one turns very expensive on the day it happens.</li>
  </ul>
  <p>Convert these into people and time, and the total comes out higher than the API bill unless your volume is very large. That is the honest answer to "what does a private LLM deployment cost": the hardware quote is the start of the bill, not the bill.</p>

  <h2>The middle path we recommend most</h2>
  <p>Run the main flow through an API, and keep all fact lookup and data storage on your own machines. For the small slice of content that is truly confidential, use a small local model to handle it separately. The cost stays under control, and the most sensitive data has never left your network. For most companies this is far more practical than building the full stack on day one.</p>
  <p>For how fact lookup and reply drafting run on a live system, see our <a href="https://ai.dingboai.com/en/case-study-ai-customer-service.html">customer service case study</a>.</p>

  <h2>Two things to set up first, whichever path you take</h2>
  <p><strong>A disk-level alert.</strong> The first thing to configure on a private setup is not the model, it is the disk. When the disk fills, the vector database quietly switches to read-only, new knowledge never gets written, and the front end shows no error at all. Alert at 80%.</p>
  <p><strong>A three-part backup.</strong> Back up the database (app settings, prompts, knowledge base metadata), the vector database's data volume, and the keys and encryption salt in your configuration files. Lose the last one and the database alone cannot be restored. Missing any one of the three means you have no backup.</p>

  <h2>Where the real work is</h2>
  <p>Hosting does not decide whether the system is useful. What decides it is whether the model receives the right facts for each request and whether its output gets checked before it reaches a customer. A private model fed raw, unfiltered material fails in the same ways a cloud model does; we list those failures in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>, and the trade-offs in system design in <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>. If you are weighing a private deployment, the free plan tells you which of the two you need in two working days.</p>]]></content:encoded></item>
<item><title>Custom AI Agent Development Cost: Where the Money Goes</title><link>https://ai.dingboai.com/en/custom-ai-agent-development-cost.html</link><guid>https://ai.dingboai.com/en/custom-ai-agent-development-cost.html</guid><pubDate>Tue, 10 Nov 2026 08:00:00 +0000</pubDate><description>Quotes for a custom AI agent range from a few hundred dollars to six figures because they build different things. What each tier gets you and what to ask first.</description><content:encoded><![CDATA[<p>Ask what a custom AI agent costs and you will get quotes that are hundreds of times apart. That is not because someone is making prices up. The quotes are for different things. Break "an AI agent" into its parts and it becomes clear where the money goes.</p>

  <h2>What the cheap tier actually builds</h2>
  <p>At the low end you usually get a prompt written on an existing platform, a few documents uploaded, and a chat box. It can answer general questions the documents already cover: product descriptions, FAQs, how-to steps. If that is what you need, it is not a rip-off. The platforms are good at this, and you can often build it yourself.</p>
  <p>What it cannot do is touch your business systems. It cannot check the status of one specific order, it does not know this customer already wrote three times, and it cannot change an address or open a ticket for a person. Most of the labour companies actually want to remove sits exactly there.</p>

  <h2>Where the expensive part goes</h2>
  <ul>
    <li><strong>Connecting systems.</strong> Orders, shipping, inventory, tickets, mailboxes: each one needs an integration, permissions, and handling for failed lookups. A lookup has to tell the model whether it found the record, confirmed there is none, found part of it, or could not check. Otherwise the model treats a failed lookup as "does not exist" and gives a confident wrong answer. This is usually the largest piece of work.</li>
    <li><strong>Turning raw material into facts.</strong> You cannot pass a whole email thread or a full ticket history to the model. It first has to be distilled into a few lines of key facts, each with its source. This step sets the ceiling on quality; the reasons are in <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</li>
    <li><strong>Writing down how your business answers.</strong> Which things the AI may answer directly, which must be logged for a person, and how to talk to customers all have to be pulled out of how your team already works. That comes from reading your real case history, not from prompt-writing tricks.</li>
    <li><strong>Acceptance testing.</strong> Replay a few hundred real historical cases and compare outputs one by one. Test questions you invent yourself hide the real problems.</li>
    <li><strong>Running costs after launch.</strong> Model calls are billed by usage. Business rules change and the answers have to change with them, and every change needs another replay. Quotes often leave this out, yet over a year it is one of the biggest lines.</li>
  </ul>

  <h2>When you should not spend much at all</h2>
  <p>Low-volume work is not worth it. A task that happens a few times a month saves little once automated and still needs someone to maintain it. Work where the information needed to decide lives only in an experienced employee's head, and nowhere in a system, should also wait. Without the facts, the AI can only guess. Neither case turns out well, whatever the quote.</p>

  <h2>Pricing models you will see</h2>
  <ul>
    <li><strong>Per day or per hour.</strong> You pay for effort. Simple to compare, but the risk that it never works is yours.</li>
    <li><strong>Fixed project price.</strong> You know the cost up front, but the scope is frozen at signing, and real processes rarely stay frozen.</li>
    <li><strong>Subscription or per seat.</strong> Common for platforms. Fine for the cheap tier, less useful when most of the value is in your own integrations.</li>
    <li><strong>Outcome-based (per completed task).</strong> You pay for work that actually got done. The vendor carries the risk, so this only works when "done" can be counted and checked.</li>
  </ul>

  <h2>Five questions to ask before you sign</h2>
  <ul>
    <li>Will it connect to our existing systems? Which ones, and what does the AI say when a lookup fails?</li>
    <li>How will it be tested before launch? Will you replay our real historical data?</li>
    <li>Which things will the AI do on its own, and which will always be left for a person to approve?</li>
    <li>When rules or the business change after launch, who updates it, and how is that charged?</li>
    <li>What is the billing unit: days, the project, or the volume of work actually completed?</li>
  </ul>

  <h2>How we charge</h2>
  <p>We bill for completed work, not for days and not as a fixed package. The plan is free. We then connect to your existing systems and run real work for two to four weeks; you can see every record the AI handled, and your people review all of it at first. Once the pilot shows how much it handled and how accurately, we agree a unit price together, and after that you pay monthly for completed volume and can stop at any time. Your up-front risk stays small, and if we cannot deliver results, we do not get paid.</p>
  <p>We are a small team based in Shenzhen, China, and we work with clients remotely. Before deciding between an agent and a plain workflow, read <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>; it often changes the scope, and therefore the cost.</p>]]></content:encoded></item>
<item><title>AI Agent vs Workflow: How to Choose for Real Work</title><link>https://ai.dingboai.com/en/ai-agent-vs-workflow.html</link><guid>https://ai.dingboai.com/en/ai-agent-vs-workflow.html</guid><pubDate>Tue, 10 Nov 2026 08:00:00 +0000</pubDate><description>One test decides it: if you can draw the next step in advance, build a workflow. Why we keep agents inside single steps, and the tool design that matters more.</description><content:encoded><![CDATA[<p>This is the question we get most often when scoping a project, and the one teams most often get wrong. Choosing badly rarely stops the system from running. It stops you from maintaining it once it does.</p>

  <h2>The one-line test</h2>
  <blockquote><p>If you can draw the next step before you start, build a workflow. Use an agent only when the next step depends on what the previous step turned up.</p></blockquote>
  <p>"Customer email comes in, classify it, look up the order, draft a reply from the right policy, send it to a person for review." You can draw that on paper today, so it is a workflow. "Take this pile of messy supplier files and turn it into a comparison table." You have to look at the files before you know what to do next, so that is a job for an agent.</p>
  <p>Most business processes that are worth automating are the first kind. They happen every day, they follow a known route, and the company already has an opinion about what the right answer looks like. That is exactly why they are worth automating, and exactly why they do not need an agent deciding the route.</p>

  <h2>Where workflows win</h2>
  <ul>
    <li><strong>Reproducible.</strong> The same input gives the same path today and next month. When a customer complains about one reply, you can rerun that exact case and see what happened.</li>
    <li><strong>Regression-testable.</strong> Change one step, replay a few hundred real historical cases, and see precisely which ones changed. You cannot do this cleanly with a free-running agent, because it takes a different path every time.</li>
    <li><strong>Easy to debug.</strong> You can see which step broke. You are not reading a long chain of model reasoning and guessing what it was thinking.</li>
  </ul>

  <h2>What you pay for making everything an agent</h2>
  <p>First, unstable results. Ask the same question twice, get two different handlings, and the business team decides the system is broken while you cannot reproduce the problem. Second, unpredictable cost: an agent that wanders can burn dozens of model calls on one task. Third, debugging by guesswork: when something goes wrong in production, what you have is a natural-language transcript, not a call trace.</p>
  <p>None of these show up in a demo. All three show up once real volume arrives.</p>

  <h2>What we do: a graph outside, agents only inside single steps</h2>
  <p>Our customer service system is built as a graph. First it works out what the customer is asking this turn. Then code decides whether to look up the order, whether to search the knowledge base, and whether the case needs a person. Then it writes the reply. Classifying, retrieving, rewriting and drafting are separate steps, and the route between them is decided by code, not by the model.</p>
  <p>Only the small part that genuinely needs exploration gets an agent that can call tools on its own, and that agent has a hard cap on the number of steps. The whole stays controllable, and the part that needs flexibility gets it.</p>
  <p>This is also why "agentic workflow" is a useful phrase. It is not a third architecture. It is a workflow where some nodes are allowed to think, and the boundaries around them are explicit.</p>

  <h2>Tool design matters more than the framework</h2>
  <p>Whichever you pick, results come down to the handful of tools the model can call. Two lessons cost us real mistakes:</p>
  <ul>
    <li><strong>Describe when to call a tool, not just what it does.</strong> "Looks up an order" is not enough. "Call this when the customer mentions an order number, an email address used at checkout, or asks where a parcel is" is what changes behaviour.</li>
    <li><strong>Make a failed lookup look different from an empty result.</strong> If the order lookup times out and returns an empty list, the model reads that as "no such order" and confidently tells the customer it does not exist. A lookup needs four distinct answers: found, confirmed not found, partly found, and could not check. This kind of bug looks like a model hallucination. It is actually a code problem, and fixing it removed more wrong answers for us than rewriting prompts ever did.</li>
  </ul>

  <h2>A quick way to decide on your own project</h2>
  <p>Write the process down as numbered steps. If every step and every branch can be written before the work starts, it is a workflow, and putting an agent in charge only adds cost and risk. If one or two steps are "figure out what this is and what to do with it", put an agent inside those steps only, give it the tools it needs, cap its steps, and keep everything around it deterministic. If you cannot write the steps at all, the problem is usually not the architecture: the process itself is not defined yet, and no agent will define it for you.</p>
  <p>For the problems that show up after launch, whichever architecture you choose, see <a href="https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html">why AI agents fail in production</a>.</p>]]></content:encoded></item>
<item><title>Why AI Agents Fail in Production: 5 Causes and Fixes</title><link>https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html</link><guid>https://ai.dingboai.com/en/why-ai-agents-fail-in-production.html</guid><pubDate>Tue, 10 Nov 2026 08:00:00 +0000</pubDate><description>Great in the demo, unreliable on real work. Five causes we hit running AI agents on live customer service, what each looked like, and the fix that worked.</description><content:encoded><![CDATA[<p>Everything works in the demo, then real work arrives and results become hit and miss. That is the complaint we hear most. Switching to a more expensive model often does not help, because the problem is usually not the model. We run AI on live customer service email and chat, in several languages, and each of the five causes below is one we hit there, with what it looked like and the fix that actually worked.</p>

  <h2>1. Passing raw material to the model as one big pile</h2>
  <p>The easy approach is to give the model the whole email thread, the full ticket history and every retrieved document, and let it work it out. It then misses what matters in the long material. In one ticket with more than twenty back-and-forth messages, we had already promised the customer a full refund once the return reached the warehouse. The model writing the next reply did not notice and contradicted it. In another, the case history ran to hundreds of thousands of characters, and the model asked the customer to re-film a video they had already sent.</p>
  <p>A stronger model given the same pile got it right, which tells you what was missing: not material, but distillation. The fix is a step before the model, done by code or a dedicated extraction pass, that turns the material into a few lines of structured facts. What the customer wants this time, what we promised before, what has been done, what the customer has already sent, verified order and shipping status, what is still missing, and where things conflict. Every line carries its source and date. Original text appears only as a one- or two-sentence quote, never as whole messages.</p>

  <h2>2. Adding a rule after every mistake until the rules fight each other</h2>
  <p>Prompts usually grow like this: one case goes wrong, add a sentence; another goes wrong, add another. A few months later, several of the dozens of rules contradict each other, and the model can only satisfy one side, or it makes something up.</p>
  <p>It happened on our own content pipeline. The prompt said "always give the place name in Chinese and English", "always give a number when mentioning a cost", and "never invent a business name". When the source material had no shop names or prices, the first two rules forced the model to write something and the third forbade inventing it. It chose to invent. Changing the first two to "write it only if the material has it; otherwise tell the reader how to look it up" made the invented shop names disappear. Another prompt that had grown past 9,000 characters gave noticeably different outputs for the same input; cutting the duplicates and contradictions and restructuring it into short, separate rules at just over 3,000 characters made it stable.</p>
  <p>So when something breaks, look for what to delete first: repeated explanations, conflicting rules, unnecessary steps. If you must add something, add a general principle, not a patch for one case.</p>

  <h2>3. The model is missing the one fact you used to spot the error</h2>
  <p>When the AI gets something wrong, the reflex is to add a rule. It works better to ask: how did I know it was wrong, and what information did I use? Once the AI answered a question about a promotion incorrectly. Going back over it, we realised we only knew the right answer because we knew today's date, and the model had not been given today's date. Giving it the date fixed it, without a single new rule.</p>
  <p>Most "the AI's judgement is off" problems are like this: the fact needed to decide is in a person's hands, not the model's. Connecting that fact beats teaching the model to guess. For lookups, also watch how failure is reported, as covered in <a href="https://ai.dingboai.com/en/ai-agent-vs-workflow.html">AI agent vs workflow</a>.</p>

  <h2>4. A processing step silently deletes business content</h2>
  <p>Material usually passes through filtering, trimming, link removal and compression before the model sees it. When these steps go wrong they do not throw errors; replies just get worse. We once stripped video links out of knowledge-base content with a filter that split text at full stops. The source wrote the part, the action, the cause and the fix as one long sentence joined by commas, so the whole troubleshooting procedure was deleted along with the link. Customers asking how to fix a problem got a reply that only said to contact their dealer.</p>
  <p>Truncation does the same thing: cut long text at a character limit and the sentence you lose is often the one that mattered. The fix is unglamorous. Before changing any step that feeds the model, run one real source text through it and compare before and after to make sure no business content was lost. Summarise long content; never truncate it.</p>

  <h2>5. No human where a human should decide, or checkpoints everywhere</h2>
  <p>Both extremes are common. One jumps straight to full automation, and the AI promises refunds or compensation on a channel nobody reviews. The other piles blocks onto drafts that a person reviews anyway, so the AI refuses to write whenever it is unsure, and the person ends up writing from scratch.</p>
  <p>We now split it by whether the output can be taken back. If a person sees it first, such as an email draft waiting for review, rules act as reminders: the draft is still produced and the person decides what to change. If it cannot be taken back once sent, such as unreviewed live chat, the AI never commits the company to anything a person has to carry out, like refunds, replacements, order changes or invoices. It logs a ticket for a person and answers only facts and published policy. Hard blocks belong only at the step where something is actually sent or executed.</p>

  <h2>How to tell which one your project has</h2>
  <p>Take the latest batch of cases the AI got wrong and ask of each: did it have the facts it needed to get it right (causes 1 and 3)? Do two rules in the prompt conflict on this point (cause 2)? Compared with the original, what is missing from what the model actually received (cause 4)? Could this mistake be taken back (cause 5)? Print exactly what the model received and read it. Most problems are found on the spot. If you are still pricing a project, <a href="https://ai.dingboai.com/en/custom-ai-agent-development-cost.html">where the money goes in custom AI agent development</a> covers which of these work items a quote should include.</p>]]></content:encoded></item>
</channel></rss>
