AI Email Reply Automation for Customer Support
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.
Step 1: Distil the email before any model sees it
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.
So every incoming email first becomes a few lines of structured facts, produced by code and a dedicated extraction pass:
- what the customer wants in this email;
- what we promised before, and what has already been done;
- what the customer has already sent (photos, videos, order number, invoice);
- verified order, shipping and warranty status;
- what is still missing, and where facts conflict.
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.
Step 2: Look things up instead of guessing
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".
Step 3: Draft every reply, with a proposed solution
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.
Step 4: Decide what sends itself
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:
- 25,576 replies were sent using an AI draft;
- 2,756 of them were sent by the AI on its own;
- 3,463 were approved by an agent without a single edit;
- 19,357 were edited by an agent and then sent.
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 human-in-the-loop AI agents.
Step 5: Measure what agents keep
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.
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.
What to set up first
- Connect order, shipping and warranty lookups before writing a single prompt.
- Build the fact-distilling step; never pass raw threads.
- Put all reply principles in one place that every step reads.
- Start with drafts for everything and automatic sending for nothing; open categories one at a time.
- Log AI draft and final sent text side by side from day one.
The flow of one email through the whole system, with all the numbers, is in our customer service case study. For the cost side, see custom AI agent development cost.