How to use AI for customer service in a small business: a practical setup guide

Use AI for customer service by giving it a written answer sheet (prices, hours, policies and your most common questions), letting it draft replies to routine messages while a person checks and sends, and handing anything emotional, financial or unusual straight to a person. Tell customers when they are talking to AI, keep a human one step away, and keep personal details out of tools you haven't checked. Read the conversations every week and add whatever the AI got wrong to the answer sheet.

Most small businesses don't have a customer service team. They have an owner with a phone, a WhatsApp chat that never stops, an inbox with forty unread emails and a contact form that sends messages somewhere nobody checks on a Saturday. AI looks like the obvious fix: something that answers the easy questions so you can get on with the work. The worry is just as obvious. What if it tells someone the wrong price, or annoys a regular who wanted to talk to you?

Both feelings are reasonable, and the answer is in how you set it up rather than which tool you buy. I'm a user researcher at KahuLogic, where we build websites and marketing tools for small businesses. I'll mention our own tool once, near the end; everything else works whichever tools you use.

Start with your questions, not the tool

The best first step costs nothing: find out what customers actually ask you. Scroll back through a month of WhatsApp chats, emails, Instagram messages and form enquiries, and jot down each question in a few words. You'll usually find the same handful repeated over and over. Are you open on public holidays? How much is a standard clean? Do you have parking? Can I move my appointment?

Those repeat questions are where AI earns its keep. The rare, tricky ones are where it causes trouble. Gartner's survey numbers above aren't a reason to avoid AI, in my view; they're a description of what customers fear: being stuck with a bot, unable to reach anyone, and being told something wrong. If you design around those three fears, most of the risk goes away.

What AI should answer, and what stays with you

AI is good at answering questions whose answers you have already written down. It is bad at judgement, empathy and anything that involves an exception. Split your messages along that line before you switch anything on.

Picture a physio clinic. The AI can tell someone the clinic opens at 7 am, that the first appointment is longer, and where to park. It shouldn't be telling someone whether their knee pain sounds serious, or agreeing to waive a late cancellation fee. The first set saves the front desk hours. The second set is where a wrong answer costs you a customer or worse.

Write the answer sheet your AI works from

Every AI customer service tool, from a free chatbot to an expensive help desk, is only as good as the information you give it. If your prices live in your head, the AI will guess, and guesses sound confident. So the most valuable hour you'll spend is writing a single answer sheet. Keep it as a plain document you can update in a minute.

• Your services, each in one or two lines, and what you don't offer (people ask that more than you'd think).
• Prices exactly as you publish them, including what's extra: call-out fees, weekend rates, deposits.
• Opening hours, public holiday hours and how far ahead people can book.
• Location, parking, accessibility and how to find the door.
• Your policies in plain words: cancellations, late arrivals, refunds, returns.
• Your ten most common questions with the answer you'd give on a good day.
• A short list of things the AI must never do: offer discounts, promise dates, diagnose, or argue.
• Two or three real replies you're proud of, so it can copy your tone.

Draft mode first: AI suggests, a person sends

My strong recommendation is to start in draft mode. The AI reads each message and writes a suggested reply, and someone on your team reads it, fixes anything off and taps send. You lose a few seconds per message compared with full automation, and you gain something more useful: you see every mistake before a customer does.

Draft mode also shows you exactly what's missing from your answer sheet. If the drafts keep fumbling questions about gift vouchers, that's because you never wrote the voucher rules down. Add them, and the next draft is right. After a few weeks of drafts that need no edits for a particular kind of question, you can decide whether to let that one kind answer on its own, if your tool allows it.

A worked example: a physio clinic's first month

Here's a worked example. The business and every number in it are made up, so swap in your own. Say a two-room physio clinic gets about 160 messages a month across WhatsApp, email and its website form. The receptionist sorts a month of them and finds that roughly 100 are routine: hours, prices, parking, whether a referral is needed, how to book. About 40 are booking changes. The last 20 are the hard ones: complaints, insurance questions and people describing symptoms.

In this made-up month, the clinic writes its answer sheet and turns on AI drafts for the 100 routine messages only. If checking and sending a good draft takes 30 seconds instead of three minutes of typing, each routine message saves two and a half minutes: 100 × 2.5 = 250 minutes, a little over four hours a month, or about 50 hours a year. At a made-up NZ$35 an hour for reception time, that's roughly NZ$1,750 a year of time moved from typing to the phone and the front desk. The 40 booking changes keep going to the receptionist, who can now get to them faster. The 20 hard messages never touch the AI except to be flagged as urgent.

At the end of the month, the clinic counts how many drafts needed changes and why. Maybe twelve did, and eight of those were about one new service it forgot to add. Adding that one service to the answer sheet fixes 8 of the 12 next month, so only 4 drafts in 100 need editing. That's the whole loop: write, draft, check, fix the sheet.

Build the handover before you need it

The moment that decides whether customers like your AI is the handover. Gartner's advice in the same release is that a chatbot should make it clear it can hand off to a person and that the handoff should be smooth, picking up where the bot left off. Set that up on day one.

1. 1Name the triggersComplaints, refunds, anything medical or legal, the words “speak to someone”, and any question the AI can't answer from the sheet.
2. 2Decide who catches itOne named person per time of day. A handover to “the team” is a handover to nobody.
3. 3Pass the contextThe person taking over should see the whole thread, plus a one-line summary of what the customer wants.
4. 4Tell the customerSay plainly that a person will reply, and roughly when: “Sam will get back to you by 3 pm.”
5. 5Test it yourselfPretend to be an annoyed customer and see how long it takes to reach a human. If it's more than one message, fix it.

Tell customers it's AI, and stay accountable

Be upfront. Salesforce's research found 72% of customers say it's important to know if they're communicating with an AI agent, and in my view hiding it is the fastest way to lose trust when someone notices. A short line in the first message is enough: “I'm the clinic's AI assistant. I can answer common questions, and I'll pass you to the team for anything else.” If you serve customers in the European Union, law firm summaries of the EU AI Act say its rule that people must be told when they're interacting with an AI system applies from 2 August 2026.

Also remember that what the AI says is what your business says. In Moffatt v Air Canada, decided by British Columbia's Civil Resolution Tribunal in February 2024, the airline argued its chatbot was responsible for its own words. The tribunal rejected that, noting the chatbot was still just part of Air Canada's website, and ordered the airline to pay the difference between the fare the customer paid and the bereavement fare the chatbot had wrongly promised. It's a small-claims decision in Canada, not law in Australia or New Zealand, but the lesson travels: you own your bot's answers.

Keep customer details safe

Customer messages are full of personal information: names, phone numbers, health details, addresses. Australia's privacy regulator, the OAIC, published guidance in October 2024 on using commercially available AI products. Its best-practice advice is not to enter personal information, especially sensitive information, into publicly available generative AI tools, and to do due diligence on any product before you use it. New Zealand's Privacy Commissioner has guidance explaining how the Information Privacy Principles in the Privacy Act 2020 apply to AI tools, and recommends a privacy impact assessment before you start.

In practice, for a small business, that means a few simple habits. Don't paste customer chats into a free chatbot to get a reply written. Use a tool built for customer messages, read where it stores data and whether it trains on your conversations, and update your privacy policy to say you use AI to help answer messages.

Measure it every week

Pick a fixed half hour each week to read a sample of conversations. You're looking for three things: answers that were wrong, answers that were right but cold, and customers who had to ask twice to reach a person. Each one tells you what to change.

Track two numbers alongside that read-through: how long customers wait for a first useful reply, and how many conversations needed a person. The first should drop quickly. The second should drop a little and then settle, because some conversations should always go to a human. If it keeps climbing, your answer sheet has a gap.

Copy this: an AI answer sheet to give your chatbot

Here is the answer sheet I'd start from, written so you can paste it into whatever tool you use: a chatbot's knowledge or instructions box, a help desk's AI settings, or the top of a document the AI reads. Fill in the square brackets, delete what doesn't apply, and put the date at the top so you know how fresh it is.

1. 11. Who we areWe are [business name], a [type of business] in [town or area]. We offer [main services, one line each]. We don't offer [things people ask about that you don't do]. Most of our customers are [who they are, e.g. local families, office managers].
2. 22. Prices[Service]: [price], which includes [what's included]. [Service]: [price]. Extras: [call-out fee / weekend rate / deposit, with amounts]. If someone asks about a price that isn't listed here, reply: “That depends on the details, so [name] will confirm a price by [time].” Never estimate, round or discount a price.
3. 33. Hours and bookingsWe're open [days and hours]. Public holidays: [hours, or closed]. Book at [booking page link] or by [method]. We can usually fit new customers in within [timeframe]. Don't confirm, move or cancel a booking yourself; say “[name] will sort that for you by [time]” and pass it to the team.
4. 44. Finding us[Where you are, or the area you cover]. Parking: [where, and whether it's free]. Access: [steps, ramps, lift, which door]. If you visit customers instead: we cover [suburbs or radius] and charge [travel fee, if any] outside it.
5. 55. PoliciesCancellations: [notice needed and any fee]. Late arrivals: [what happens]. Refunds and returns: [the rule]. Explain these exactly as written. If someone asks for an exception, don't agree or refuse; pass it to [name].
6. 66. How to behaveStart every new conversation with: “Hi, I'm [business name]'s AI assistant. I can answer common questions, and I'll pass you to the team for anything else.” Hand over to a person straight away for complaints, refunds, anything medical, legal or financial, or when someone asks for a person, and say: “I'm passing this to [name], who'll reply by [time].” Keep replies short and friendly. Here are two replies in our style: [paste two real replies you're proud of].

Keep one master copy of this sheet and change it there first, then in the tool. When a draft gets something wrong, the fix almost always belongs in the sheet: a price that changed, a service you forgot, a policy that was only in your head. In my view, a dated, one-page sheet that's right beats a long one that's half out of date.

If you only do three things this week

• Scroll back through a month of messages on every channel and tally the questions; circle the ten that come up most.
• Fill in the answer sheet above for those ten questions, date it, and give it to whatever tool you use or plan to trial.
• Pretend to be an annoyed customer on each channel and count how many messages it takes to reach a person. If it's more than one, fix the handover first.

Where Kahu fits

If you'd like help without handing over the keys, here's what our tools do. WhatsApp Inbox is live on every plan and works the way this guide recommends starting: Kahu drafts each reply from your prices and hours, and a person on your team checks it and taps send. Kahu Assist adds website chat that answers questions from your site's own pages and can hold a booking slot, but never changes a booking. Work Kahu drafts waits in your approvals queue until you approve it, and automations run only when you switch them on. Your website, ads and CRM are included from Growth (US$179 a month, or US$149 billed yearly). Starter (US$69) covers social posting, Creative Studio and chat. Every plan starts with a free 7-day trial, no card needed.

Where Kahu fitsWhatsApp InboxKahu reads each WhatsApp message, drafts the answer from your prices and hours, and a person on your team taps send.Your existing WhatsApp Business numberReplies drafted from your prices and hoursA person presses sendSee WhatsApp Inbox →Nikau Reevesvia WhatsApp BusinessHey, anything free Thursday after 5? And roughly how much for a cut and beard trim?17:04Draft · from your price list and hoursKia ora Nikau, yes, 5:30 Thursday is free. Cut and beard trim is $55, about 45 min. Want me to hold it for you?EditSend

I'm Muskaan Sharma, User Researcher at KahuLogic, a New Zealand company with a team across Australia, New Zealand and India. My honest view is that AI customer service works best when it's boring: it answers the questions you've already answered a hundred times, and it gets out of the way the moment a customer needs a person.

Frequently asked questions