What a chatbot should actually do for a small business

We have built conversational assistants for a dental clinic, an independent garage and a wholesale roastery. The ones that earned their keep have almost nothing in common with what people picture when they hear the word chatbot. They are narrow, they are unglamorous, and a good share of their value is in knowing when to hand over to a human.

The job is triage, not conversation

A small business does not need a system that can discuss anything. It needs one that can handle the four or five questions that consume most of the front desk's day. For the clinic, that was which practitioner do I need. For the garage, roughly what will this cost. Both were answered dozens of times a day by a person who was in the middle of something else.

So the first deliverable on any assistant project we take on is not code. It is a document: sit with whoever answers the phone and write down the actual decision tree they use. That document is the product. Everything after it is plumbing.

Handing over is a success state

The clinic assistant escalates about one conversation in eight, and we count that as working correctly. A patient describing something the tree does not cover gets a human, with the transcript already attached so nobody has to repeat themselves.

An assistant that never escalates is not confident, it is guessing. The failure mode people actually remember is not it could not help me, it is it confidently told me something wrong. Those cost trust in a way that is very hard to win back, particularly in health and in trades.

Give it real data or do not ship it

The difference between a useful assistant and an expensive novelty is almost always whether it is connected to something true. Real calendar availability, not a promise that somebody will call. Real prices for the fixed-price work. Real stock levels.

The garage assistant can quote an oil change outright because it knows the actual price list. It cannot quote a suspension noise, so it books a diagnostic instead of inventing a number. That boundary is drawn deliberately, and drawing it is most of the work.

Three rules we hold to

• It never speculates. If the answer is not in the source data, it says so and offers a person.
• It works in both languages. In Montreal this is not a nice extra. A French-only visitor meeting an English-only assistant is a lost customer.
• It is measured on outcomes, not conversations. Bookings made, quotes issued, calls avoided. Message volume is vanity.

When the answer is no

We have talked clients out of this twice. If your enquiry volume is ten a week, an assistant is a solution looking for a problem, and a well-written FAQ page with an honest response-time promise will serve you better for a fraction of the cost. The threshold is roughly when repetitive questions start eating an hour a day. Below that, fix the website copy first.