Support that answers at 3 a.m.

Use AI customer service for the questions you answer every day, such as opening hours, shipping, order status and setup steps. Keep people on refunds, complaints and anything unusual. Write the knowledge base before you switch anything on, because an AI agent can only be as good as the answers you give it, and always offer an easy way to reach a person.

Done this way, a small business can answer common questions at 3 a.m. without a night shift. Done badly, the bot invents a policy that the business then has to honor.

01What AI customer service handles well

The best questions to automate are frequent, have one correct answer, and cost little if the answer is slightly off. The worst are rare, emotional or involve money. Sort your questions like this:

02What to keep human, and why it matters

Customers forgive a slow reply more easily than a runaround. The questions that matter most to them, such as “where is my refund?” or “why was I charged twice?”, are exactly the ones where a bot’s wrong answer does the most damage.

Klarna is a well-known test of how far to go. In early 2024, it announced that its AI assistant was already doing a large share of its support work:

Fifteen months later, it changed course. In May 2025, its CEO said cost had weighed too heavily in the plan and “what you end up having is lower quality.” Klarna announced plans to hire people again, and he said customers should always know they can reach a human.

The lesson for a small business: automate the routine so your people have time for the conversations that keep customers.

03You own what your bot says

A support bot speaks for your business, legally as well as practically. In February 2024, a Canadian tribunal ordered Air Canada to compensate a customer after its chatbot gave wrong advice about bereavement fares. The airline argued the chatbot was a separate entity. The tribunal said the company is responsible for all information on its website, whether it comes from a static page or a chatbot.

Our guide to AI for small business tells that story in full. The practical rule is simple: if your bot can say it, assume you will have to stand behind it.

Disclosure is becoming law too. In the EU, since August 2, 2026, people must be told when they are interacting with an AI system, unless it is obvious. Elsewhere, it is simply good practice. A line such as “I’m the virtual assistant for [business]; a person can take over at any time” costs nothing.

04Build the knowledge base first

AI support agents answer from your content: they search your articles for the relevant passage and answer from it. Our explainer on RAG shows how that works. The consequence is simple. When an article is missing or out of date, the bot either says it does not know or, worse, fills the gap with a plausible guess.

1. 01

   Find your top questions

   Export your last three months of emails, chats or tickets. Remove names, emails and order numbers, then ask AI to group them into the questions customers actually ask, with a count for each.

2. 02

   Write one short article per question

   Give the answer, the policy behind it, the exceptions, and when to contact a person. Plain sentences beat long pages, because the bot quotes what it finds.

3. 03

   List what the bot must never answer

   Refunds, legal threats, medical or safety questions, account security and anything you decide case by case. These go straight to a person.

4. 04

   Test with real past questions

   Before launch, ask the bot 30 to 50 questions from your old tickets and compare its answers with what a good agent said. Fix the articles, then test again.

5. 05

   Launch small and read everything

   Start on one channel, such as website chat. For the first two weeks, read every AI conversation. You will find missing articles within days.

This prompt does the sorting in step one:

Below are [number] customer messages from the last three months, with personal details removed.
Group them by the underlying question, in the customer's words, most frequent first.
For each group, give the count, one example message, and whether the answer is: a fixed fact, a policy, a lookup (such as order status), or a judgment call.
Flag any group where customers sound angry or confused.
Messages:
[paste]

And this one turns each answer into an article the bot can use:

Write a help article answering: [question].
Use only these facts: [paste your policy, prices, steps].
Structure: a one-sentence answer, the steps or details, the exceptions, and when to contact us.
Under 200 words, plain language, no marketing. If a fact is missing, write [CHECK] instead of guessing.

05Design the handoff to a person

The handoff is where AI support earns trust or loses it. A customer stuck in a loop with a bot remembers the loop, not the answers.

If your tool lets you write standing instructions for the bot, start from something like this and adapt it:

You are the virtual assistant for [business]. Say so in your first message.
Answer only from our help articles. If they do not cover the question, say you are not sure and offer a person.
Never promise refunds, discounts, credits, delivery dates or exceptions.
Hand the conversation to a person when the customer asks for one, mentions a refund, a complaint, a legal issue or a safety problem, sounds upset, or when you have failed to help twice.
When handing off, summarize the issue so the customer does not have to repeat it, and say when a person will reply.

Then check your setup against the list below.

06AI customer service tools and what they cost

The five tools below all include an AI agent, and four of them charge per resolved conversation on top of the subscription. Prices below are as listed on each vendor’s pricing page in September 2026. Some show annual-billing rates, and all change often.

The fair way to compare is cost per resolved question. Take last month’s ticket count, estimate the share the bot could answer from your articles, and multiply by the price per resolution. Then compare that with the time your team spends on the same questions today.

07How to measure it honestly

Vendors measure success in resolutions, and each defines the word its own way. “No further help requested” can mean a happy customer. It can also mean someone who gave up and phoned you instead. So track what happens after the chat, not just in it.

Once a week, read 20 AI conversations picked at random. Tag each one as right, incomplete or wrong. Fix the articles behind the wrong ones. It is a small weekly habit, and it catches problems before customers complain in public.

08FAQ

Next, see how phone bots work in voice AI explained, or set ground rules for your team with our AI use policy template.