AI chatbots for customer service and lead capture: a practical guide

An AI chatbot for a business is a program that answers customers' messages in ordinary language, on the website, Messenger, Instagram or WhatsApp, using a language model and information you have approved. Done well, it answers routine questions at any hour, collects a potential customer's details and passes everything else to a person; done badly, it invents answers in your name.

This guide covers what to give a chatbot, what to keep from it, the channel rules that limit it, and how to test one without betting the business on it.

What a chatbot should handle, and what it should not

Good candidates are the questions you answer every day in the same words: opening hours, delivery terms, how booking works, what to bring, whether an order has shipped. The second job is capture: when someone asks about a service, the bot collects a name, a phone number and what they need, and hands that to whoever calls back.

Keep it away from complaints, refunds, prices that depend on the case, negotiation, and any answer where a mistake costs the customer money, health or rights. Those go to a person, every time.

Its wrong answers are your wrong answers

In February 2024 a Canadian tribunal decided a case every business with a chatbot should know. Air Canada's website chatbot gave a customer wrong information about the airline's bereavement fares, and the airline in effect argued that the chatbot was responsible for its own actions. The Civil Resolution Tribunal disagreed. In the tribunal's words, it should have been obvious to the airline that it is responsible for all the information on its website. The tribunal found that the airline had not taken reasonable care to make sure its chatbot was accurate, and ordered it to compensate the customer. One case in one country is not a rule everywhere, but the reasoning is hard to argue with. To the customer, the bot is you.

Made-up answers are a known property

Language models can write fluent text that is wrong. NIST, the US standards institute, names this risk confabulation in its generative AI risk profile: confidently stated but false content, known colloquially as hallucinations, by which users may be misled. Nobody can promise it away; it can only be contained. Three things contain it: the bot answers only from a knowledge base you wrote and approved; it is instructed to say "I don't know" and offer a person when the answer is not there; and someone reads a sample of real conversations every week.

Not every customer wants to talk to a bot

In a Gartner survey of 5,728 customers, 64% said they would prefer that companies did not use AI in customer service. That is not a reason to skip the bot. It is a reason to make it honest: say clearly that it is an AI assistant, and make the way to a person visible from the first message. Gartner's own advice in the same release is that a chatbot must tell the customer it will connect them to an agent when the AI cannot provide a solution. In the European Union it is also a legal duty: from August 2, 2026, Article 50 of the EU AI Act requires providers of AI systems that interact directly with people to design them so that people are informed they are interacting with an AI system, unless that is obvious. If your bot serves customers there, ask whoever built it how it meets that requirement.

The channel sets the rules

Messenger and WhatsApp limit when a business may write. On Messenger and Instagram, Meta's policy gives a business a standard messaging window to respond: for most user actions it lasts 24 hours, and when a person sends their first message after clicking a Click-to-Messenger or Click-to-Direct ad, it may last up to 7 days from that message. On WhatsApp, a 24-hour customer service window opens when the customer writes; once it closes, the business can send only pre-approved template messages. And the WhatsApp Business Messaging Policy lets a business contact people only after they have given opt-in permission. In practice: a chatbot can answer a lead tonight, but once the window has closed it can no longer follow up with a free-form message. Capture the details while the window is open, and let a person follow up.

Lead capture: ask less

Ask only for what the person calling back needs: a name, a phone number, what they are looking for and when. Send it at once to the place where inquiries live, with the channel it came from, and notify whoever is on duty. A bot that collects twelve fields and drops them into an inbox nobody checks is worse than no bot. How to wire that path is in our article on marketing automation for a small business.

Write it the way your customers write

Plan for messages that do not read like a brochure: mixed languages, Georgian typed in Latin letters, voice notes and photos, three questions in one line. Collect a hundred real messages before writing a single answer, and test the bot on those, not on tidy questions you invented. If customers write in two or three languages, the approved answers have to exist in each of them: a model translating your policy on the fly is one more place for a mistake.

A first pilot in four weeks

1. Week one: read. Sort a few hundred recent messages by question and pick the five that come up most and have a fixed answer.
2. Write the answers. Approved text for each, in every language you serve, plus the cases where the bot must hand over.
3. Week two: build and test it internally. Your team tries to break it: misspellings, mixed languages, angry customers, questions just outside its scope.
4. Weeks three and four: one channel, limited hours. For example, the website chat outside working hours, with every conversation reviewed.
5. Decide. Keep it, widen it or switch it off, based on what the conversations show.

When not to use a chatbot

When you get a handful of messages a week: a person answers faster and better. When the answers change every day and nobody will keep the knowledge base current. When nobody is available to take the conversations the bot hands over, because then the hand-off leads nowhere. And when the subject is sensitive: health, money, legal rights, a complaint.

How to tell it works

Four measures, read monthly: the share of conversations the bot finished without a person and without a complaint; the inquiries it captured that a person then followed up; the time from hand-off to a human reply; and the wrong answers found in the weekly review, which should fall. A bot that resolves a lot but gets things wrong is not saving time. It is moving the cost to later.

Where to start

How chatbots fit among other uses of AI in marketing is in what AI marketing is. How we design, build and tune assistants is on the AI chatbots and assistants service page. If you want us to read your current inquiries and tell you whether a bot is worth it, request a free audit.