Triage and draft replies to customer emails with AI

Written by the editorial team · 12 November 2026 · Updated12 November 2026

An AI agent can sort incoming customer emails, summarise each one and draft a reply from your message history and knowledge base, with the source of each statement attached. A person reads and validates every reply before it goes out, and some mails should never be answered automatically at all: refunds, price commitments, legal matters and complaints. Measure quality on a sample of your own past emails, by comparing the agent's category and draft with what your team actually did.

A shared inbox fills up with mails of very different weight: a delivery question, a quote request, an angry complaint, an invoice query, an out-of-office reply. An email agent does two jobs on them. It sorts, and it prepares an answer. A person decides what leaves.

What does the agent read?

Three kinds of material, and the quality of the draft follows from them.

The first is the incoming message itself, with its attachments. The second is the history: earlier mails in the thread and earlier exchanges with the same customer, so the agent does not ask again for something the customer already sent. The third is your knowledge base, meaning the documents your team actually uses to answer: price lists, product sheets, terms and conditions, internal procedures, and above all previous replies that were judged good.

Give the agent access to the folders where these live and nothing else. An agent that can read the whole company drive will sooner or later quote a document that was never meant for customers.

What does the agent prepare?

For each mail, a short package your team can read in a few seconds:

• a category (information request, quote request, complaint, invoice question, appointment, other) and a priority;
• a two-line summary of what the customer wants;
• a draft reply in your tone, with the document each statement comes from, so the reader can check it;
• a list of what is missing, if the draft cannot be completed, for instance an order number;
• a suggested owner when the mail belongs to someone else on the team.

A draft that cites its source is checked faster than a draft that merely sounds right. When the knowledge base does not contain the answer, the right output is a short note saying so and a suggestion to ask a colleague, never a confident guess.

What does the person validate?

Everything, at the start. The person opens the draft, reads it against the cited source, edits it if needed and sends it from their own mailbox. This is our default: the agent prepares and a person in the company validates before anything is sent.

Once the team trusts a category, you can lighten the review for it, for example by letting the agent file and label mails without drafting. That decision belongs to you and is taken per category, with the quality measures below in hand.

What should an agent never send alone?

Draw the line before the first mail goes through. A reasonable list:

• anything that commits money: refunds, discounts, credit notes, price quotes;
• anything legal: contract terms, liability, data protection requests, threats of litigation;
• complaints from an unhappy customer, where tone matters as much as content;
• anything about health, safety or a security incident;
• mails from a person the agent has never seen, asking for personal or account information;
• any mail for which the knowledge base holds no source.

For these, the agent still summarises and proposes a draft, and a named person owns the answer.

How do you measure quality without invented figures?

Build the test from your own history. Take a sample of past emails, remove personal data if you need to share them, and keep for each one the category your team gave and the reply that was really sent. Then run the agent on the same mails and read the results side by side.

Four questions are worth tracking:

1. Does the category match what your team chose, and where it differs, who was right?
2. Does each statement in the draft appear in the cited source?
3. How often does your team send the draft unchanged, edit it lightly, or rewrite it? The edit trail is the most honest quality signal you will get after launch.
4. Did the agent hand over to a person every mail that belongs on the “never alone” list?

The last one deserves a separate count, because a single missed escalation costs more than several clumsy drafts. Rerun the same sample each time you change the knowledge base or the instructions, so you see the effect of the change.

And with us?

We publish this blog and we are a forward deployed agency: three Télécom Paris engineers specialised in AI who build custom agents, and who also take on machine learning and data projects. An email agent is one of the agents we build on request. For a network of garages, for instance, it could read appointment requests and the history of a customer, draft the reply and propose a slot, with the person at the front desk validating. The agent cites its sources and can be reached from Claude, ChatGPT or Copilot as well. Your mailbox is connected on demand, your documents are read in read-only mode on the folders you choose, and encryption keys are kept per company, with servers in France. The first call is 30 minutes, free and without commitment, and a first agent runs two weeks after it. More on automating mails with AI and on AI agents for companies.