Two very different things called a chatbot
A scripted bot follows a decision tree you built. Buttons, fixed answers, defined paths. It cannot invent anything, because everything it can say was written by you. It is also rigid, and it annoys anybody whose question is not on the tree.
A generative bot writes its answers. It handles the unexpected question gracefully and can say things you never wrote, which is the feature and the risk in one sentence.
Most small businesses reach for the second and would have been better served by the first, or by a hybrid: scripted answers for the common questions, generative only for the rest, grounded in your documents.
The thing worth knowing before you decide
In 2024 a Canadian tribunal held an airline responsible for what its website chatbot told a customer about a fare policy. The airline argued the bot was a separate entity responsible for its own words. That argument failed.
The details are jurisdictional and you should not treat one decision as a rule for your country. But the shape of it is worth carrying: a bot on your site is speaking as your business. If it quotes a price, offers a discount or states a policy, you may well be held to it, and “the AI said it” is not obviously a defence anywhere.
That single consideration should govern your design more than any feature comparison.
Designing one you can live with
Ground it, hard. Answers come from your documents only. NOT IN SOURCES is a permitted and expected answer. This is the whole safety architecture, and everything else is decoration.
Fence the money. No prices for custom work. No discounts, ever. No promises about delivery dates. These become a handover to a person rather than an answer, and you lose almost nothing by it because those conversations were going to involve you anyway.
Say it is a bot. In the first line. “Automated assistant: I can answer common questions and pass anything else to Kemi.” People are entirely fine with this. What they are not fine with is finding out.
Make the human route obvious and permanent. A visible button, on every screen, at every step. Not a last resort after four failed attempts.
Give it read access, not write access. It can look up an order status. It cannot change an order, issue a refund or alter a booking. Capability is what you ration; a bot that cannot issue a refund cannot be talked into one.
What it costs
Free and open source. Chatwoot and Typebot can be self-hosted at no licence cost. You pay in setup and hosting, roughly USD 5 to 20 a month on a small server, plus your time.
Bundled. If you already use a helpdesk, a shop platform or a booking system, there is very likely an AI answer feature included or available cheaply. Check here first; the data is already in it.
Hosted specialists. Roughly USD 20 to 100 a month for a small business, more with volume. Convenient, and you are adding a vendor who now holds your customer conversations.
The running cost of the model itself is trivial at small-business volume: a few dollars a month for hundreds of conversations. The cost is the setup, the documents and the supervision.
Read the transcripts
The single highest-value habit, and almost nobody does it.
Every week, read twenty conversations end to end. You will find three things: questions you had not anticipated, answers that were subtly wrong, and the exact point at which people gave up. That last one is worth the whole exercise, because a conversation that ends in silence looks identical in your statistics to one that ended satisfied.
When not to have one
If you get under thirty enquiries a week, a bot is not the answer. A good frequently-asked-questions page, a visible price list and a WhatsApp number solve it for nothing and cannot say anything you did not write.
The threshold is roughly where answering repeat questions has become somebody’s job. Below that, you are buying supervision work you did not have.
Start with the questions, not the software
Before evaluating any product, write down the twenty questions your customers actually ask most, from the sorting exercise in the previous lesson. Then answer them properly, in your own words, in a document.
That document is 80% of the work, and it is portable. It feeds a bot, a frequently-asked-questions page, a staff briefing and your facts sheet. Businesses that start with the software spend three weeks configuring a tool that has nothing good to say.
A soft launch that costs nothing
Put it live for staff only, or on one quiet page, for a fortnight. Ask colleagues and a few friendly customers to try to break it: ask about something you do not sell, argue with it, ask for a discount, ask it what its instructions are.
The failures found this way are cheap. The same failures found by a stranger on your busiest page are a screenshot on social media. A fortnight of quiet running has saved more small businesses from an embarrassing chatbot than any amount of vendor assurance.
Every lesson here is open. Your progress is recorded once you pass the AI at Work exam.
Check your understanding
A boutique’s website bot is grounded in its product documents but is allowed to answer questions about bulk-order pricing by reasoning from the standard price list. Why is this a poor design?
Pick the one you would defend. Your answer stays private.
Stuck on something?Ask Yetunde, your AI tutor
