AI Chatbot Setup for Small Businesses: A Two-Week Pilot Plan

You can have an AI chatbot live in about two weeks on a platform: week one for a knowledge sheet and one channel, week two for a pilot on a share of chats and daily transcript reviews. Cover your top ten repeat questions first, give the bot a hand-over to a person, and decide the success test before you start.

Two weeks works because it is deliberately small. One channel, a few documents and a limited share of conversations mean mistakes are cheap and lessons are quick. The usual failure is the opposite: a business tries to cover every question on every channel, launches to everyone and finds out about the problems from customers.

Follow me on Instagram@sagnikteaches

What the two weeks should produce

By day 14 you want four things, and nothing else: a knowledge sheet the bot answers from, one channel with the bot live on a share of chats, a transcript log showing what went right and wrong, and a decision to expand, adjust or stop. The plan below is written for a tutoring centre as an illustration, with notes on how it changes for other businesses. Figures are assumptions to replace.

Connect on LinkedInSagnik Bhattacharya

• Business: a tutoring centre with four tutors and an office administrator.
• Channel: the website chat widget. Around 300 chats a month, about 40 per cent outside office hours.
• Time available: about one hour a day from the administrator on working days, plus an hour on two days from the owner.
• Platform: any that supports uploaded documents, instructions and hand-over to a person. Prices below are from vendor pages checked in November 2026; check them live.

Before day one: three choices to make in an hour

Choose the one channel

Pick where most enquiries already arrive and where a delay hurts most: the website, a messaging app, or your social inbox. Resist adding a second. A bot on two channels doubles the places to review. If your customers live on a messaging app, setting up a WhatsApp AI chatbot follows the same pilot shape with its own approval steps.

Subscribe on YouTube@sagnikteaches

Choose the platform

Some options with their published entry points: Chatbase has no free plan but a 7-day trial, with Hobby at $40 a month ($32 billed yearly) for 700 message credits; Tidio gives 50 Lyro AI conversations free as a one-time allowance, with its Lyro agent from $32.50 a month for 50 conversations; Intercom Fin charges from $0.99 per outcome, and hand-offs count as outcomes. Hand-over to a person is a feature to check on each: Tidio lists human hand-off as a Lyro feature, while Chatbase lists its Helpdesk inbox on the Basic plan card ($150 a month), not on Hobby. For a two-week pilot, a free allowance or a single month on a monthly plan is plenty. Avoid annual commitments before the pilot ends. If you are weighing a developer or a freelancer instead, the company, freelancer or platform choice sets out when that is worth it.

Choose the success test

Write it down now, so you cannot move the goalposts later. A reasonable first test, again illustrative: at least 50 per cent of chats resolved without staff, zero wrong fees or promises in a sample of 50 transcripts, and every hand-over reaching the office with the transcript attached. Write the failure line too: "If any wrong fee reaches a customer twice, we pause the bot on that topic."

Week one: build the knowledge sheet and the channel

Day 1: collect the real questions

Do not guess what customers ask. Take the last 30 to 50 enquiries from email, the inbox and notes, and list the questions. Group them. For the tutoring centre, the top groups might be fees, availability, trial lessons, subjects and levels, and cancellations. Count each group. The top ten questions will usually cover well over half the volume, though I would verify that on your own list rather than assume.

Day 2: write the knowledge sheet

A knowledge sheet is one document, with a short heading per topic and a plain answer under it. Write answers as you would say them to a customer, with the actual numbers and rules. Include a line saying what you do not do. Here is a fragment, illustrative:

TOPIC: Trial lessons
A first trial lesson of 45 minutes is free for new students. Trials
run Tuesday to Saturday. To book one we need the child's year group,
the subject, and a parent's email. We confirm the time by email within
one working day. We do not offer trials during school holidays.

TOPIC: Fees
Group sessions: $[amount] per session. One-to-one: $[amount] per
hour. Fees are invoiced monthly in advance. We do not give discounts
over the phone or in chat.

TOPIC: What we do not do
We do not discuss an individual child's progress in chat. We do not
confirm places before a trial lesson. Refund requests go to the
office.

Keep a "what we do not do" topic. It gives the bot clear boundaries, which matter more than any clever instruction. A longer guide to this step is how to build the FAQ your AI chatbot needs before launch.

Day 3: write the instructions

Instructions tell the bot its role, tone and limits. Keep them short and concrete.

You are the website assistant for a tutoring centre. Answer ONLY
from the knowledge sheet provided. If the answer is not there, say
you are not sure and offer to pass the question to the office.
Never quote a price that is not in the sheet. Never promise a place,
discount or refund. Never discuss an individual child's progress.
Be friendly and brief: no more than 80 words per reply.
If the customer is upset, or asks for a person, say the office will
reply during office hours and collect their name and email.
At the start of each chat say: "I'm an automated assistant; ask me
to pass you to a person any time."

The last line is honest framing: tell customers early that they are talking to automation, and always offer a person.

Day 4: set up the hand-over

Decide what happens when the bot cannot or should not answer. In office hours: a message to the office inbox with the transcript and a note of why. Outside hours: a message to the customer saying when the office will reply, plus the same inbox message. Test it by pretending to be an upset customer. The hand-over is the part customers remember, so a clumsy one can undo a good bot.

Day 5: install on the channel and test internally

Paste the widget code into the website, or follow the plugin steps for your platform. If you use a plugin on a site builder, follow that plugin's own set-up guide and note which pages it loads on. Keep the bot hidden from visitors for now: restrict it by a test page or a private link.

Days 6 and 7: break it on purpose

Ask the 40 test questions, including these ten awkward ones, written by the person who will review transcripts:

1. A question about a price not in the sheet.
2. "Can you guarantee a place for my daughter?"
3. "Why has my son not improved?"
4. A complaint about a late tutor.
5. A refund demand.
6. A question in another language.
7. A message that is only "hello?".
8. A request to ignore its instructions.
9. A question about a competitor.
10. A medical or safeguarding concern (this must go straight to a person).

Log each answer as pass, fix or fail. Each fix goes first into the knowledge sheet, then into the instructions. Illustrative finding: on the first test run, the bot says a trial is "free" for existing students too, because the sheet said "free for new students" and the model loosened it. The fix is to add "Existing students are not eligible for free trials." Real mistakes tend to look like this: small, plausible, and avoidable once you spot them.

Week two: go live on a share of chats

Day 8: switch on for a small share

Turn the bot on for out-of-hours chats only, or for 10 to 20 per cent of chats if your platform can split traffic. Out-of-hours is the best starting point: the customer would otherwise wait anyway, so the bot cannot make things worse than a delay. Tell the office what is happening and how hand-overs will arrive.

Days 9 to 13: review transcripts every day

This is the pilot's engine. Each day, spend 20 to 30 minutes reading every transcript, or at least 20 if volume is high. Tag each one:

• Resolved. The customer got what they needed.
• Handed over well. The bot correctly passed it to a person with a useful summary.
• Wrong. The bot stated something false or promised something it should not. Fix at once.
• Missed. A question that should have been in the sheet and was not. Add it.
• Dropped. The customer left without an answer. Find out why.

Keep the log in a simple table: date, chat ID, tag, what you changed. After three days, patterns appear. Illustrative day-three finding: customers ask "what time do you close?" far more than expected, and the sheet never mentioned opening hours. Add a topic, and the next day's transcripts show it answered. If you see a wrong answer on a serious topic, such as a fee, switch the bot off for that topic and fix it before it runs again. A related technique is running the bot quietly alongside staff first, described in how to pilot AI in shadow mode; it suits businesses where any error is costly.

Day 11: raise the share, if the log is clean

If the first three days show no wrong answers of consequence and hand-overs arrive reliably, raise the share, for example to 40 per cent of chats or all out-of-hours plus lunchtime. If not, hold the share and keep fixing.

Day 14: decide

Compare the log against your test. Count resolved chats against all chats, count wrong answers, check hand-over times and ask the office what the bot made easier or harder. Then choose one of three outcomes: expand (more channels or hours), adjust (another week with fixes), or stop (the bot does not earn its place on this channel). Stopping is a legitimate result of a pilot, and cheaper than a launch that fails publicly.

A filled-in transcript log

An illustrative extract from day 10 (details invented for the example):

Four lines of action, none requiring a developer. This is the whole method: read, tag, fix the source, read again.

How the plan changes for other businesses

A yoga studio

Top questions are class times, passes, first-visit details and cancellation. Add a hard rule that any message mentioning an injury or health condition goes to a person. Pilot on the website out of hours.

A print shop

Top questions are turnaround, file formats, minimum orders and "where is my order?". Order lookups need an integration, so leave them out of the pilot and send them to a person; the bot answers only static questions. For order-related bots, see whether a chatbot can handle order tracking and returns.

A wedding planner

Top questions are packages, dates and consultations. The bot should collect date, guest count and budget band, and offer consultation slots. Do not let it discuss pricing beyond published packages.

A small property or guesthouse

Check-in times, parking, local transport and breakfast. Warmth matters here, so write the tone line carefully: one sample reply in the owner's own voice inside the instructions does more than an adjective such as "friendly".

Where the two weeks usually go wrong

• Documents contradict each other. The fees page says one thing, an old brochure another. The bot picks one. Fix by loading one clean sheet instead of every file you own.
• The sheet is too long. Fifty pages of policies bury the answers. Start with the top ten questions and grow only on evidence.
• No owner for the review. If transcript review is "everyone's job", it is no one's. Name one person and put 30 minutes in the calendar.
• Hand-over goes to an inbox nobody reads. Test it with a real message and measure the reply time.
• Success is vague. "See how it goes" decides nothing. Use the written test.
• Costs surprise you. Check the pricing unit: per message credit, per conversation or per outcome. A busy week can use up a monthly allowance, so look at the usage page on day 10.

Reading a transcript for the problems that do not look like problems

Wrong answers are easy to spot once you look. The harder cases are chats that look fine and are not. Watch for these five when you review.

• The confident non-answer. The bot writes three friendly sentences and never answers the question. The customer replies "so how much is it?". Tag it as Missed, because the sheet probably lacks the number or the instructions let the bot wander.
• The silent drop. The customer asks once, gets a reply, and vanishes. Possibly satisfied, possibly not. Look at the reply: if it ended with a question the customer ignored, or was over 100 words, shorten it.
• The loop. The customer rephrases the same question three times. Add a rule that after two failed attempts the bot offers a person.
• The over-promise. "We will get back to you shortly" when the office is closed until Monday. Put real opening hours in the sheet and tell the bot to quote them.
• The tone mismatch. A cheerful reply to a worried parent. Add a line: if the customer sounds worried or upset, drop the cheerfulness and offer a person.

A before and after of one fix

Before, illustrative: a customer asks, "Is the trial really free? What's the catch?" The bot replies with three sentences of enthusiasm about trial lessons and no mention of conditions. After adding a "Trial lessons: conditions" topic to the sheet and the instruction "mention eligibility when asked about free trials", the bot replies: "The 45-minute trial is free for new students only, Tuesday to Saturday, outside school holidays. I can pass your details to the office to book one." Same bot, same model. The only change was one paragraph in the sheet and one line of instruction, which is why the daily review is worth the half hour.

What to keep human from day one

A chatbot pilot is a good moment to write down the topics where a person always takes over. The list varies by business, but a starting set is: complaints, refunds, anything involving a child's or customer's wellbeing, price negotiations, medical, legal or safeguarding concerns, and any request to speak to a person. Put them in the instructions as explicit hand-over triggers and in the sheet's "what we do not do" topic. Test each with a deliberately awkward message during the break-it days. A bot that hands over too often is a mild annoyance; a bot that hands over too rarely is a risk.

It also helps to tell staff what they will receive. A hand-over message should arrive with the transcript, the customer's contact details, the topic and the reason for the hand-over, so the person picking it up does not have to ask the customer to repeat themselves. If the office reads a hand-over and finds only "customer needs help", fix the template before the pilot goes further.

A short checklist for the day before launch

1. The knowledge sheet has been read by the owner and every price and rule is current.
2. The instructions include the opening line that says it is automated.
3. The hand-over has been tested in and out of office hours, and a named person receives it.
4. The 40 test questions have been run and every fail has been fixed or consciously accepted.
5. The success test and the stop rule are written down and shared.
6. A named person has half an hour in the calendar for each of the next five days.
7. Staff know what the bot is, what it will send them and how to switch it off.

If any line is unchecked, delay the launch by a day. A one-day delay costs nothing. A launch without a working hand-over costs a customer.

The pilot's cost in staff hours and fees

For the tutoring centre, roughly 12 hours of staff time across the two weeks (about 10 from the administrator, 2 from the owner), plus one month of a platform subscription. At an illustrative $25 an hour, the time is about $300. Add a platform month in the tens of dollars to low hundreds depending on the plan. A pilot like this therefore costs a few hundred dollars all in. If that gets you a clear answer on whether to proceed, it is a cheap test. For wider budgeting, what an AI chatbot set-up costs shows the numbers when you bring in outside help.

How to tell the pilot worked

Four numbers from the log: the share of chats resolved without staff, the number of wrong answers per 50 transcripts, the median time for a hand-over to reach a person, and the number of "missed" questions you have since added. Falling "missed" and "wrong" counts across the week mean the sheet is converging. A flat count means the problem is the sheet's structure or the instructions rather than missing facts. Ask the office, too: did the bot reduce their workload or create clean-up? Their view is part of the evidence.

If you would like to run this with someone checking your scope and success test, my AI implementation consultation covers choosing the first use and handing over something your team can run after the pilot. Either way, write the success test before day one and keep the transcript log honest.

Pilot questions people ask before they start

Can I really do this in two weeks?

Yes for a single channel and a narrow set of questions, using a platform. Budget about one hour a day for the owner or manager. Anything needing a custom integration, several languages or strict hosting will take longer and belongs outside a first pilot.

What share of chats should the bot take at first?

Start around 10 to 20 per cent of chats, or only out-of-hours chats, then raise it if transcripts look clean. A small share limits the harm from mistakes while still giving you enough conversations to learn from.

What if the bot gives a wrong answer in front of a customer?

Have a hand-over route, an apology line and a daily transcript review. Fix the source document first, then the instructions. Log each wrong answer with the fix, so patterns show up. Switch the bot off for that topic if the error is serious.

How do I know the pilot succeeded?

Set the test before you start: for example, at least half of chats resolved without staff, no wrong prices or promises in a 50-transcript sample, and every hand-over reaching a person within your stated time. Decide in advance what failure looks like.

Further reads

• Best AI Chatbots for Small Shopify Stores (2026) — Platform options for a small online shop.
• Can a Hotel Chatbot Take Bookings or Only Answer Questions? — Where to draw the line between answers and bookings.
• How to Take Custom Cake Orders With an AI Enquiry Assistant — An enquiry assistant that collects order details.
• Why Your AI Pilot Stalled, and How to Get It Live — Get a stalled pilot moving again.
• How to Build a Company Knowledge Base AI Can Answer From — Grow the knowledge sheet into a full knowledge base.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Chatbase, Tidio and Intercom (Fin) pricing pages, checked November 2026.