AI Email Replies for Customer Service: Drafts You Approve

The setup I recommend to a small business is simple to say and takes some care to build: AI reads each customer email, sorts it, and writes a draft reply from your own approved answers, and then a person on your team reads the draft, fixes what needs fixing and presses Send. This can reduce repetitive typing while keeping a person responsible for the reply; measure the time saved and the corrections needed during a small trial. And you keep the one thing you cannot afford to lose with customers, which is control over what leaves your company in your name.

The useful distinction is whether the software prepares a reply or sends it. For a first project, I recommend drafts with human approval.

What does an AI email responder do?

An AI email responder is software that reads an incoming message, works out what the customer wants, looks up the answer in the material you gave it, and writes a reply. Some tools send that reply automatically; others leave it as a draft.

You may already have the draft version without noticing. Google describes Gemini in Gmail as able to "create new emails or refine existing drafts", and Microsoft's own help page for Copilot in Outlook tells the user to review the message and edit the draft before selecting Send. These are personal writing helpers, though: they help one person write one email. Customer service needs more than that, because the question is not "can AI write an email" but "does every customer get a correct answer, on time, from the right person, and does somebody notice when the same complaint arrives for the fifth time this week?"

Why should a person approve every reply, at least at the start?

Ask yourself a few uncomfortable questions about your own inbox. Which of your answers to customers are written down anywhere at all? Who decides what you say when a delivery is late and the customer is angry? If an AI wrote a polite, confident and completely wrong reply about a refund, how would you find out, and how fast?

Few owners can answer all three on the spot, and that is fine, because writing those answers down is the first part of the work. Until they exist, an AI that sends on its own is guessing with your name on the email. A draft that a person approves is the safe middle: the AI does the typing and the looking up, your employee does the judging, and you start to see, draft by draft, where your written answers are missing or out of date.

There is also a security reason that is easy to miss. Customer emails are text from outside your company, and OWASP, the security community that publishes the best-known list of risks for AI applications, lists prompt injection as the first one in its 2025 list: input from external sources can change how the AI behaves. An email can contain instructions that a person would ignore and a model might follow. A human who approves every outgoing message is one of the controls OWASP itself names for actions that matter.

Which customer emails should you start with?

Not all of them. I would start with the emails that are frequent, boring and have one correct answer:

• order or job status questions ("where is my order", "when is the technician coming");
• opening hours, addresses, what to bring, how to pay;
• requests to change a date or a delivery address;
• simple questions about an invoice, where the answer is in your system.

I would keep away, for now, from complaints about money, anything with a legal flavour, cancellations of large contracts and any message where the customer is plainly upset. The AI can still sort these and put them at the top of the queue with a short summary, which already saves your team time, but a person writes the reply.

A practical way to choose: export two weeks of customer emails, read fifty of them, and mark each one with its type. You will usually see that a handful of types make up most of the volume. Those are your first set.

How do you set it up so it does not go wrong?

The order matters more than the tool.

1. Write down the answers first. A short document per email type: what we say, what we never say, which facts we look up and where. This is the material the AI drafts from. Without it, the AI invents a plausible tone and plausible facts.
2. Decide where the drafts appear. In the inbox your team already uses, or in your helpdesk, as a draft attached to the customer's message. A separate window adds friction, so test whether the team actually uses the tool.
3. Give every message an owner and a reply time. The draft is only useful if someone is responsible for approving it today.
4. Check the data question before you connect anything. Customer emails contain names, addresses and sometimes things that are none of the AI provider's business. Use a business plan where the provider does not train on your data by default (OpenAI, for example, states this for its business and API products), sign the provider's data processing agreement, and agree internally which details are stripped out before the text reaches the AI. Also check the lawful basis, customer information, retention, provider subprocessors and any international transfers. Start with redacted test messages and read-only access; drafting should not give the AI permission to change orders, addresses or payments.
5. Review the drafts that needed the most changes. After two weeks, look at the drafts your team rewrote the most. Each one points to a missing answer in step 1 or a type of email that should go back to a person.

Only after that would I talk about letting some replies go out without approval, and only for the most boring types, with a weekly sample check.

What does it cost and how long does it take?

The honest answer depends on where your emails live (a shared inbox, a helpdesk, a CRM) and how much of step 1 already exists. If your answers are written down and your inbox tool has an AI drafting feature or an open connection, a first set of email types may be running within a few weeks, depending on access and approvals. If nothing is written down, the writing is the real project and the AI part is the smaller half. I publish fixed starting prices for this kind of setup on the customer support service page, and the scope is agreed in writing before any work starts.

From our own system I run my own work on a system of AI agents: research, documents, reports and email drafts. The rule for email is the same one I suggest to you: an agent may prepare a draft, but the draft waits in my Drafts folder and nothing goes out without my yes. On this site I am setting up the assistant the same way: it will answer routine questions from an approved knowledge base and the public price list, and anything non-standard will come to me as a draft that I read and send myself. It is being set up now, so I describe the design here, not results.

Where this fits with the rest of your customer service

Email replies are one piece. If enquiries also arrive by phone, form and messenger, the bigger win is often to bring them into one place with a person responsible for each one, which I cover in the guide to enquiries, CRM and support. If your problem is speed rather than volume, read speed to lead, and for what happens after the first answer, lead follow-up automation. And if you are wondering how far to let AI act on its own, AI agents for small businesses goes into where to keep a person in the loop; the AI acceptable use policy template helps you write down what data your team may give to AI in the first place.

Talk it through

If you want to know which of your customer emails are worth drafting with AI, and which should stay fully human, we can look at it together in a free 30-minute call. Afterwards you get a short written outline with up to three tasks worth starting with.

Calls are in English.

Free 30-min call with Lidia

Sources

• Collaborate with Gemini in Gmail - Gmail Help (Google), accessed 18 Oct 2026.
• Draft an email message with Copilot in Outlook - Microsoft Support, accessed 18 Oct 2026.
• Enterprise privacy at OpenAI / Business data, accessed 18 Oct 2026.
• LLM01:2025 Prompt Injection - OWASP Top 10 for LLM Applications, accessed 18 Oct 2026.