Everybody has been on the wrong end of an AI customer service experience. The chatbot that answers a question you did not ask, the six clicks to find a phone number, the email reply that clearly did not read your email.

Which makes it worth being precise about what actually works, because the useful version of this is real, and it does not look like any of that.

Start with drafting, not answering

The highest-value and lowest-risk starting point is AI that drafts a reply for a person to review and send. Not AI that replies.

The person still reads the customer's message, still owns the answer and still presses send. What disappears is the typing, the hunting for the standard paragraph about delivery and the fifteen minutes of not starting because the message is awkward.

Do this for a quarter and you get two things: a real time saving, and several hundred examples of where the AI draft was wrong. That evidence is what tells you whether anything can safely be automated further. Skipping it and going straight to an autonomous bot is how businesses end up switching one off three months later.

What to automate end to end

A narrow list, and it is narrow on purpose. Automate answers that are factual, stable and verifiable:

  • Opening hours, locations and how to find you
  • Order and delivery status, pulled from the actual system rather than guessed
  • Product specifications and compatibility, grounded in your own documentation
  • Lead times and current availability
  • How to do a routine self-service task — reset a password, download an invoice, book a slot

What these have in common is that there is a correct answer, it lives somewhere the tool can read and being wrong is embarrassing rather than expensive.

What never to automate

Equally firm, and worth writing down where the person configuring the tool will see it:

  • Complaints. Somebody who is annoyed and gets a bot is now more annoyed and has a story.
  • Anything about money already paid. Refunds, disputed charges, billing errors.
  • Cancellations. The last conversation before somebody leaves is the one most worth having a person in.
  • Vulnerability or safeguarding. Bereavement, illness, financial difficulty, anything involving a child.
  • The second contact about the same issue. A returning customer is telling you the first answer failed.
The rule that matters most is the last one. Getting it wrong once is a support failure. Getting it wrong twice, automatically, is a review.

Grounding is the whole game

An assistant that answers from general knowledge will tell your customer something reasonable about businesses like yours. An assistant grounded in your own documentation tells them something true about you.

That means the content underneath it matters more than the tool on top. Your help centre articles, specifications, terms and lead times need to be current, in one place and written clearly, because the assistant will quote whatever it finds.

This has a pleasant side effect. Every business we have taken through this ends up with a much better self-service knowledge base as a by-product, and a meaningful number of customers stop contacting support at all because they can now find the answer themselves.

Keeping your own voice

Three things prevent the slide into generic corporate warmth:

Give it real examples. Take four replies your best person actually wrote — including a slightly awkward one handled well — and use them as the reference for tone. This does more than any amount of instruction about being friendly and professional.

Name the words you do not use. An explicit banned list works better than a description of your voice. If your business does not say reach out, seamless or we apologise for any inconvenience caused, say so.

Keep it short. Most robotic-sounding output is simply too long. Three sentences that answer the question read as human. Nine sentences that circle it read as a machine, whoever wrote them.

Always show the exit

Every automated interaction needs a visible, one-click route to a person, on every screen, without a challenge about whether the customer has tried the help articles.

Hiding the exit is the single decision that converts a helpful assistant into a hated one. It also produces the outcome the business least wants, which is customers who give up quietly rather than complaining and simply do not come back.

Measure the right thing

Deflection rate is a tempting metric and a misleading one, because a customer who gave up counts as deflected.

Measure instead: how many automated conversations ended without a follow-up contact within seven days, how many escalated to a person and were then resolved quickly and what your customer satisfaction looks like split between automated and human handling. If the split shows a gap, that gap is the honest cost of the automation.

Where to start

Draft-and-review for a quarter. Tidy your help content while that runs. Then automate the five factual categories and nothing else, with a visible route to a person on every screen. Most small businesses find that is where the benefit plateaus anyway, and the further steps carry all the risk.

If you would like a hand connecting an assistant to your actual systems so it answers from real order data rather than guesses, that is system integration work and we do a lot of it. Tell us what your customers ask most and we will tell you which of it is worth automating.