Customer replies eat a surprising amount of the working day, and most of them are variations on the same handful of questions. That makes them an obvious job for AI. It can draft a warm, clear reply in seconds, keep your tone consistent even when you are rushed, and take the routine load off you so you can deal with the messages that actually need a human.

But customer service is also where a careless setup does the most damage. A wrong price, an invented policy or a tone-deaf reply to an upset customer costs you trust, and sometimes money. This guide covers which model to use, where to draw the line on automation, and how to keep AI on-script.

Which model for replies

Good news: customer service rarely needs your most expensive model. Most replies are routine, so a fast, capable mid-tier model is usually plenty. In the Claude family, the best-balance option (Sonnet 5) handles this comfortably, and for high volumes of simple, near-identical replies the faster Haiku tier is cheaper still (both dated July 2026). OpenAI’s GPT and Google’s Gemini families offer the same kind of fast mid-tier option.

You want speed and consistency here more than raw brilliance, because you are answering the same sorts of questions many times, not writing your annual report. Save the top tier, Opus 4.8 or Fable 5 in the Claude family (dated July 2026), for the genuinely hard message: a serious complaint, a sensitive negotiation, a reply where getting the words exactly right really matters.

Draw the line: draft, do not send

This is the most important decision you will make, and it has nothing to do with which model you pick. Our honest advice is to keep a person in the loop and treat AI as the drafter, not the sender.

A sensible split looks like this:

  • Let AI draft, you approve. For the bulk of your replies, have AI write the response and you glance at it and hit send. You get most of the time saving and keep full control. This is the sweet spot for most small businesses.
  • Fully automate only the truly routine. An instant “thanks, we have got your message, we will be back within a day” or answering a simple opening-hours question can run without you. Keep these narrow and safe.
  • Always keep a human on the hard stuff. Complaints, refunds, anything about price or timing, anything from an unhappy customer. These need judgement AI does not have, and the cost of a wrong automatic reply is high.

The rule is simple: the more a mistake would cost you, the more a human belongs in the loop.

Stop it making things up

Left to guess, AI will happily invent a price, a delivery time or a returns policy that sounds completely plausible and is completely wrong. The fix is to feed it your real answers rather than let it improvise:

  • Give it your actual prices, policies and FAQs. Paste in your real returns policy, your genuine lead times, your standard answers. Now it is repeating your facts, not making up its own.
  • Tell it your tone. “Friendly, brief, we are a small local firm” produces a very different reply from the corporate default, and a much better one for most trades and small businesses.
  • Tell it what it must not do. Set clear limits, like “never quote a price, always say a human will confirm the figure”. A good instruction keeps it from wandering into territory where it might promise something you cannot honour.

Do this properly and AI becomes a reliable first responder that sounds like you. Skip it and it becomes a fluent source of wrong answers.

The rule that keeps you safe

AI drafts, you check. For routine replies a quick glance is enough. For anything sensitive, read every word before it goes, because a customer cannot tell the difference between your firm making a promise and an AI making one on your behalf. Either way, they will hold you to it.

Last reviewed: July 2026. The specific model names above are examples of each tier as it stood in July 2026. Model tiers move, so re-check the current line-up twice a year rather than chasing every release.

Your next step

Not sure whether your replies need the balanced model or the faster, cheaper one? Our Which AI model should I use? picker walks you through it in a few questions and gives you a starting point with a caveat. To compare the families side by side by speed, cost band and free-tier position, see the model comparison table.

The free resources shelf also has a one-page plain-English glossary to keep the jargon at bay. And to see reply-drafting set up safely on real customer messages, come to our AI Automation Masterclass in Manchester. Laptops open, plain English, no fluff. Tickets are normally £20. This one’s free, a limited-time offer to launch the series.