How to use an AI chatbot for customer service in a small business: what to hand over, how to set it up, and your first seven days
You run a small team, the same questions arrive all day on WhatsApp and your website, and you want an AI chatbot to take the repetitive ones without upsetting customers. This guide covers the decision (what the bot should and should not handle), the setup, and a day-by-day plan for the first week, which is where most launches succeed or stall.
Quick answer
To use an AI chatbot for customer service, give it the repetitive questions (prices, hours, availability, bookings, order basics) and keep complaints, refunds and anything needing judgment with a person. Load it with your own content, tell it to hand over when unsure, test it with 20 real questions, then launch on one channel and review every conversation daily for the first week.
Which customer service jobs should an AI chatbot take?
The ones that are frequent, answerable from written information, and low risk if a person checks later. Keep anything emotional, money-related or one-off with a human.
| Conversation | Who handles it | Why |
|---|---|---|
| Prices, opening hours, location, parking | AI | Same answer every time, written once |
| "Do you have a slot on Saturday?" | AI | Checkable against your calendar and bookable in the chat |
| Service details, what is included, preparation instructions | AI | Answerable from your brochure or website |
| New enquiry: what they need, budget, timing | AI, then you | The bot collects details; you close the sale |
| Rescheduling an existing booking | AI, with your rules | Fine if your policy is written down |
| Complaints and angry customers | Human | Needs empathy and authority to fix things |
| Refunds, discounts, exceptions | Human | Money and judgment; never let a bot promise these |
| Medical, legal or financial specifics | Human | Liability; the bot should share general information only |
The handover rule matters as much as the split. The WhatsApp Business Messaging Policy says businesses using automation must have "prompt, clear, and direct escalation paths" to a person. Write yours down before you start; our guide to escalation rules has a template.
Website, WhatsApp or both?
Start where most of your questions already arrive. For many small businesses in Asia, Latin America, Africa and the Middle East that is WhatsApp; for others it is the website. Launch on one, get it right, then add the second.
- Website chat catches visitors who are browsing and would otherwise leave. They are often anonymous, so the bot's job is to answer and capture a way to continue.
- WhatsApp catches people who already have a question and expect a fast reply. Speed matters: the Harvard Business Review lead study found firms that responded within an hour were about seven times more likely to qualify the lead than those that waited longer.
If you are unsure, our piece on whether you need both walks through the trade-offs.
The afternoon setup checklist
Most of the work is writing, not configuring. Set aside two to three hours.
- List your top 20 questions. Scroll back through a month of messages and write down what people ask most, in their words.
- Gather your knowledge. Website pages, current price list, service descriptions, policies (cancellation, refunds, delivery). Delete anything outdated. Our guide on training a chatbot on your own data covers preparation.
- Write the instructions. Who the bot serves, what it should do (answer, book, collect details), the tone, and the rule: if the answer is not in the knowledge, say so and hand over.
- Write the handover triggers. Complaints, refund requests, a customer asking for a person, anything medical or legal, and any question the bot cannot answer.
- Connect the channel. Add the website code or connect your WhatsApp number. Use a dedicated business number if you can, not your personal one.
- Test with your 20 questions. Type them the way customers do, including typos and voice notes, and fix the content behind each wrong or vague answer.
- Tell your team. Who watches for handovers, how fast they reply, and how to take over a chat.
The week-one plan, day by day
The first week is a supervised trial, not a launch-and-leave. Here is what to do each day, and roughly how long it takes.
| Day | What to do | Time |
|---|---|---|
| Day 1 | Go live on one channel. Read every conversation by the evening. Note every wrong, vague or missing answer. | 45 minutes |
| Day 2 | Fix the content behind yesterday's notes. Add Q&A pairs for questions you did not expect. Check every handover was picked up. | 30 minutes |
| Day 3 | Read conversations again. Look at tone: too long, too formal, too many questions at once? Adjust the instructions. | 30 minutes |
| Day 4 | Test the edge cases yourself: an angry message, a refund request, a question in another language, a photo. Confirm each hands over or answers properly. | 20 minutes |
| Day 5 | Review bookings or leads the bot created. Are the details complete? Did any slot clash with your real diary? | 20 minutes |
| Day 6 | Ask two regular customers or a friend to message it with real questions and tell you honestly how it felt. | 15 minutes |
| Day 7 | Count the numbers below, decide whether to add the second channel, and switch from daily to twice-weekly reviews. | 30 minutes |
Expect day 1 and day 2 to surface the most gaps; that is normal and a sign the review is working. For what weeks two to four tend to look like, see the first 30 days with an AI agent.
What to check at the end of week one
Four numbers tell you whether the chatbot is helping. You can count them by hand from a week of conversations.
Answered without you
The share of conversations the bot handled from start to finish. Do not chase a high number in week one; a correct handover is a success too.
Wrong answers
Count answers that were factually wrong. The target is zero, and each one points to a fix in your content or instructions.
Handover pick-up time
How long customers waited after the bot handed over. If people waited hours, the problem is the team process, not the bot.
Bookings or leads captured
Appointments booked or enquiries with full details collected, especially outside your opening hours.
After the first month, track the same four monthly and add one more: how many enquiries came in outside your opening hours and still got an answer.
What usually goes wrong, and the fix
In the accounts we see, week-one problems come from the content and the process far more often than from the AI itself.
- The bot says "I'm not sure" a lot. The knowledge is missing those answers. Add them; it is the system working as designed. More in why your agent says it is not sure.
- An answer is wrong. Usually two documents disagree, such as an old PDF and a new website price. Delete the old one.
- Handovers sit unanswered. Nobody owned them. Name one person per shift and set notifications on their phone.
- Replies feel robotic. The instructions are too formal or the bot asks three questions at once. Ask for one question per message and shorter replies.
- Customers ask for a human straight away. Check the greeting says what the bot can do, and make sure the person who takes over replies quickly.
Running this plan on ReplyKit
ReplyKit is an AI agent built for this exact job in a small business: it answers WhatsApp and website chat from your own content, qualifies leads, and books appointments into real open slots with double-booking protection and a reminder about 24 hours before. If it does not know an answer, it says so and notifies you, and you can take over any chat in one click and hand it back just as quickly. Unanswered questions are listed so you can teach the answer once, which is most of the week-one work above.
WhatsApp connects through Quick Connect, which pairs your number by QR code in about two minutes. It is unofficial: it is not the WhatsApp Business API, Meta does not sanction it, and a number can be restricted. Before connecting you see one plain notice, and pressing Connect accepts section 3 of our terms. We recommend a dedicated number, and new numbers start with safety limits of 30 AI replies an hour and 200 a day, with a warm-up. The website assistant is one line of code. Plans start at $29 a month; see pricing, or start a 7-day free trial with no credit card and run the week-one plan on it.
When ReplyKit is not the right fit: if you need to send broadcasts or marketing campaigns, work in WhatsApp groups, cover Instagram or Messenger, or require the official API with a verified business presence, choose an official-API platform. If you need tickets, SLAs and integrations with a help desk, a help desk tool fits better.
Frequently asked questions
Can an AI chatbot replace customer service staff?
For a small business, it usually replaces the repetitive part of the job, not the people. Staff stop retyping prices and hours and spend their time on complaints, sales and anything that needs judgment, which a chatbot should hand over anyway.
How long does it take to set up an AI chatbot for customer service?
The technical setup can take minutes; the writing takes an afternoon. Plan two to three hours to gather your content, write instructions and test, then about 30 minutes a day in the first week to review conversations.
What should a customer service chatbot not do?
It should not handle complaints, refunds, discounts, exceptions to your policies, or medical, legal and financial specifics. It should also never guess: if the answer is not in your content, it should say so and hand over to a person.
How do I stop the chatbot giving wrong answers?
Use a tool that answers only from your own content and admits when it does not know. Then remove outdated documents, fix contradictions between your website and PDFs, and review every conversation daily in the first week.
Do customers like talking to AI chatbots?
Some customers prefer a person, which is why disclosure and a fast route to a human matter. Customers tend to accept a bot that answers quickly and correctly and hands over smoothly; what they dislike is being stuck in a loop.
How much does an AI customer service chatbot cost?
For a small business, roughly $29 to $250 a month depending on the tool, channels and seats. ReplyKit is $29, $49 or $99 a month with AI replies included and no Meta fees.
Run your first week with an AI agent, free
Answer WhatsApp and website questions from your own content, with you one click away. 7-day free trial, 300 AI replies, no credit card.
Sources and further reading
- WhatsApp Business Messaging Policy · automation must have prompt, clear and direct escalation paths to a person
- Harvard Business Review, The Short Life of Online Sales Leads (Oldroyd, McElheran, Elkington, 2011) · firms that responded within an hour were about seven times more likely to qualify the lead
- ReplyKit, Will WhatsApp ban my number for automation? · the risk of unofficial QR pairing and the limits we enforce
ReplyKit is an independent product and is not affiliated with or endorsed by Meta or WhatsApp. Third-party prices and policies were checked in October 2026 and may change. This article is general guidance, not legal advice.