The short answer: AI drafts good customer replies when you give it three things it can't know — what actually happened, what you've decided to offer, and how your business sounds — and keep two things out: personal details the reply doesn't need, and any decision about refunds or fault. A person reads it before it goes. Done that way, the reply you've been avoiding goes out before lunch.
Customer email is where most small businesses first try AI, and where most of them decide it "sounds robotic". It's a fair complaint about the default output and an unfair one about the tool. The generic reply is what you get when the tool has nothing to work from but the customer's message.
Why AI replies sound generic
Give a tool a complaint and nothing else, and it writes the average of every customer service email it has seen. It thanks them for reaching out. It apologises three times for the inconvenience. It restates their complaint back to them at length. It promises that their feedback is valued. Then, somewhere in paragraph three, it gets to whatever might actually happen next — usually a vague version, because it doesn't know.
The fix is entirely on the input side. Getting output you can actually use covers the general principle; for customer replies it comes down to a short list.
What to give it
- The customer's message, with details stripped out (more on that below).
- What happened, from your side, as dot points. The job notes, the timeline, what was done and what wasn't.
- The decision you've already made. "We'll re-attend Thursday morning at no charge." "No refund on the deposit, but we'll credit the call-out fee." The tool writes around a decision; it shouldn't make one.
- The channel and length. An email reply, a text, a response to a public review — each wants a different length.
- Two or three replies you were happy with, as examples of your voice. This does more for tone than any description of tone ever will.
Then one instruction that changes most replies: answer the question they actually asked in the first two lines.
What to keep out
The reply rarely needs the customer's full name, address, phone number, account number or anything about their health or circumstances. "The customer" or a first name is enough for the tool to write a good reply; you add the rest when you send it. Which account you're pasting into matters as much as what you paste — tiers and what happens to your data explains why, and your never list should name customer correspondence explicitly.
Before you hit send
- The first two lines answer the question they asked
- Every fact in it was in your notes, and is true
- The offer is the one you decided, not a more generous one
- It doesn't admit or deny fault in words you haven't chosen
- It sounds like someone at your business, not a call centre
Complaints are where it helps most
The reply that sits in the inbox for three days is almost never a hard one to write. It's an unpleasant one to start. That's where a draft helps most: it takes the emotional cost out of the first version, and you're editing instead of composing at the end of a long day.
Say a customer writes that your technician tracked mud through the house and didn't fix the leak. Your notes say: re-attend Thursday 8am, carpet clean booked at our cost, technician spoken to. A reply built from those notes opens with an apology for both problems and the Thursday time — in the first two lines — instead of three paragraphs about valuing their feedback.
Two refinements make complaint replies better. Ask the tool to remove anything defensive, and ask it to make the draft shorter. Almost every first draft of a complaint reply is too long, and length reads as excuse-making.
If you'd like your team's reply setup built on your own past emails, with the never list and the check written in, that's a half-day workshop.
See the team workshopMake it a setup, not a habit
Everything above — the voice examples, the sign-off, your standard offers, what to keep out, the first-two-lines rule — belongs in a reusable setup rather than in someone's memory. Staff then paste the customer's message and their notes and get a draft in the business's voice, whoever is on the desk. In ChatGPT that's a custom setup; in Claude it's a Project. The module is building reusable setups, and if you're choosing between tools, the rule of thumb applies — for this job either works, and the setup matters more than the brand.
The same approach carries across industries: quote follow-ups in trades and field services, client updates in cleaning and maintenance, and the volume of tenant and owner correspondence that buries property managers.
What stays with a person
The words can be drafted. Some decisions can't be: whether to refund, whether you were at fault, how to respond to a consumer guarantee claim under the Australian Consumer Law, anything with an insurance or legal angle, and anything clinical. A person decides those, and the tool writes the reply around the decision. Where there's an insurance or legal angle, the exact wording of an admission matters — that's a call for the owner, and sometimes for the insurer.
And every reply gets read before it's sent. How to check an AI draft in five minutes covers what to look for.