AI customer support for small businesses: what to automate (and what not to)

In February 2024, Klarna announced that its AI assistant had handled 2.3 million conversations in its first month, two-thirds of its customer service chats, the work of about 700 full-time agents. Fifteen months later its chief executive told Bloomberg the company was hiring human agents again, saying "it's so critical that you are clear to your customer that there will always be a human if you want." Both things were true: the AI did handle the volume, and the business still needed people. That's the real lesson of AI customer support for a small business. The question isn't whether to use it. It's which part of the work to hand over.

This guide gives you a practical split: what to automate, what to keep human, what sits in between, how the bot should sound, the five numbers to watch, and what it costs.

What AI customer support is

AI customer support means an AI model answering your customers' questions, usually in a chat on your website or in your inbox, using your own knowledge: your pages, your policies, answers your team has written. When it can't help, it hands the conversation to a person.

It's different from the old scripted chatbot, which made people click through menus to reach a canned answer. An AI assistant reads the question in the customer's own words, in their language, and writes a reply. That flexibility is what makes it useful, and it's also what makes the rules around it matter. A scripted bot can only say what you typed. An AI one can say something you never approved, unless you set it up not to.

Some tools go further and let the AI take actions, such as issuing a refund or changing a booking. That's closer to an AI agent, and it carries more risk, as we'll see.

What the numbers say about customers and AI support

Before deciding what to automate, it helps to know what customers actually think. These are the most recent sourced figures we could find:

Read together, they say one thing: customers want fast answers at any hour, and they'll accept a bot that gives them, as long as it works and a person is one step away. A bad bot is expensive, because most people won't give it a second chance.

What to automate: the green list

Hand the bot questions that are frequent, factual and already answered somewhere you control.

A good starting point: take the ten questions your team answers most often, write one clear answer for each, and let the bot handle those first.

What not to automate: the red list

Keep these with people, at least for now:

• Refunds, credits and exceptions. A bot can explain the refund policy. Deciding whether this customer gets one is a judgement call with money attached.
• Complaints and upset customers. People who are angry want to be heard by someone who can fix it. A cheerful bot makes it worse.
• Anything about a specific account. Changing an address, a password or a payment method needs identity checks a chat window can't do.
• Discounts and negotiation. A bot that can be talked into a discount will be.
• Legal, medical, financial or safety questions. Point to a person or a professional.
• Anything your knowledge doesn't cover. This is the big one. A bot that guesses is worse than no bot.

The cautionary tale is Air Canada. In 2022 its website chatbot told a customer, who was booking flights after his grandmother died, that he could claim a bereavement discount retroactively, after buying full-price tickets. The airline's policy didn't allow that, and it refused. In February 2024 British Columbia's Civil Resolution Tribunal ordered Air Canada to compensate him and rejected its argument that it couldn't be held responsible for what its chatbot said. "It should be obvious to Air Canada that it is responsible for all the information on its website," the tribunal member wrote. Whatever your bot says, your business said.

The amber zone: AI drafts, a person sends

Between the two lists is work where AI saves real time but a person should still press the button:

• Email replies. The AI reads the message, finds the answer and drafts a reply; someone checks it and sends.
• Replies to tricky chats. The AI suggests a response inside the conversation; a team member edits or sends it with one click.
• Order or booking status. If your tool can look these up, start with a person approving before you let the bot answer on its own.
• Summaries for handoffs. The AI writes a two-line summary so the person taking over doesn't need to read the whole chat.

This "AI does the work, a person approves" pattern is called human in the loop, and for a small business it's often the best value: most of the time saved, almost none of the risk.

When a person must step in

Write your handoff rules down. A good set:

1. The customer asks for a person. Always one click away, never hidden behind three failed answers.
2. The bot tried and couldn't answer. It should say so plainly, not loop.
3. The topic is on your red list. Complaints, refunds, accounts, anything sensitive.
4. The customer is getting frustrated. Repeated questions, capital letters, "this is useless".

And make the handoff itself good:

• Tell them what happens next, honestly: "The team usually replies within an hour" beats "Someone will be with you shortly" when nobody will.
• Ask for an email if you don't have it, so a reply can reach them after they leave.
• Pass the whole conversation on, so the customer never repeats themselves. In Zendesk's 2026 survey, 74% of consumers said having to repeat information frustrates them.
• Alert your team, and escalate if nobody picks it up within a few minutes.
• Let the person take over completely. Once someone from your team has replied, the bot should stay out of that conversation until it's handed back.

Tone: how your AI support should sound

Your bot is the first voice many customers hear from you. Some rules that work:

• Say it's an AI. In the EU, Article 50 of the AI Act has required since 2 August 2026 that people are told they're talking to an AI system unless it's obvious. Everywhere else, it's simply honest. A name and a friendly face are fine; claiming to be a person isn't.
• Short answers first. Two or three sentences, then a link to the page with the detail.
• Show where the answer came from. A link to the pricing or returns page builds trust and lets people check.
• Admit what it doesn't know. "I don't have that information, but I can get someone from the team to reply" is a good answer.
• Match the customer's language. If they write in Spanish, answer in Spanish.
• Sound like your business, not like a call-centre script. Write a few lines of tone guidance: formal or casual, what to recommend, what never to say.

Here's the difference in practice:

How to measure AI customer support

Track five numbers, weekly at first:

Pair the numbers with reading. Twenty minutes a week reading real conversations will tell you more than any dashboard, especially in the first month. If a measure suddenly worsens, look for a change on your site: a new price the bot hasn't read yet is the usual culprit.

What AI customer support costs

Prices are structured in very different ways, so compare on your own volume. Three examples, checked on their pricing pages:

A worked example: say your site gets 600 chats a month and the bot finishes 60% of them on its own. Per-resolution pricing at $0.99 comes to about $356 a month for those 360 conversations, before seats. Per-reply pricing depends on how long the chats are, so check a month of real conversations before you compare. Per-resolution pricing ties what you pay to results; per-reply pricing ties it to usage, which is cheaper for simple questions but means you pay for chats that end in a handoff too.

Two honest notes. Intercom and Zendesk are complete helpdesks, with much more than the AI chat, which a small business may or may not need. And the biggest cost is rarely the software: it's the hours you put into writing good answers and reviewing conversations. Budget an afternoon to set up and twenty minutes a week after that.

A 30-day rollout plan

• Week 1: collect. Pull the last month of questions from your inbox, contact form and social messages. Find your top ten.
• Week 2: write. Write a clear answer for each, check your website pages are current, and set your handoff rules and tone.
• Week 3: soft launch. Put the chat live, test it like a stranger, and read every conversation.
• Week 4: improve. Write answers for everything it couldn't handle, check the five numbers, and decide what to hand over next.

How crew bot fits

Our own support chat is built around the split above. You put one line on your site, and a bot you choose for that site answers from your pages, your team's answers and your documents, in the visitor's language, with its sources underneath. It never invents prices, dates, discounts or promises. When it can't answer, or the visitor wants a person, it takes their email and the conversation waits for your team in crew bot, with an alert, and an email to the team if nobody replies within two minutes. The visitor sees the reply live, or by email if they've left.

The support bot has no tools by design: it can't touch your systems, payments or files, so anything a visitor types can at worst produce an odd reply. Other bots in your crew can suggest a reply in a conversation as a draft that a person sends with one click, which is the amber zone above. The questions the bot couldn't answer are listed, each with a button to write the answer. If you'd also like help with email, crew bot's inbox drafts the replies that need you and sends nothing until you say so. Setup details are on /support, and our step-by-step guide shows how to add an AI chat to your website on any platform.

FAQ

Can AI replace my customer support team?

Not entirely, and it shouldn't try. AI handles repetitive, factual questions well and at any hour. People are still needed for judgement calls, complaints, exceptions and anything involving money or accounts. In Gartner's 2026 survey, 87% of customers said companies using AI for service must offer a way to reach a human.

How much does AI customer support cost for a small business?

It depends on the pricing model and your volume. Some tools charge per resolved conversation (Fin charges $0.99 per outcome; Zendesk $1.50 to $2.00 per automated resolution beyond those included), others per reply or through a monthly plan. Work it out on a month of your real conversations, and set a spending cap.

What questions should a support chatbot not answer?

Refunds and exceptions, complaints, changes to a specific customer's account, discounts, legal, medical or financial advice, and anything its knowledge doesn't cover. For those, it should say so and hand over to a person.

Do I have to tell customers they're talking to an AI?

In the EU, yes: since 2 August 2026, Article 50 of the AI Act requires that people are told they're interacting with an AI system unless it's obvious. Elsewhere it's still the honest default, and customers trust a bot more when it's upfront.

What are the disadvantages of AI customer service?

It can give wrong answers if its knowledge is out of date or it's allowed to guess, it frustrates people when it hides the way to a human, and a bad first experience puts customers off: Gartner found only 27% would try a chatbot again after a negative one. Good knowledge, a strict "don't guess" rule and an easy handoff fix most of this.