Every few weeks someone asks me to build them an AI chatbot. Some of them need one. Some need a better FAQ page. A few need something else entirely, an agent that does the work behind the question rather than answering it. This guide is the conversation I have with them, written down.
The question behind the question
"Do I need a chatbot?" usually means one of three things:
- We keep answering the same questions. Opening hours, pricing, what is included, how returns work, whether you serve a certain area.
- People ask things only our systems can answer. Where is my order, is Tuesday free, what did you quote me last month.
- People want something done, not explained. Book it, quote it, change it, cancel it.
The right answer is different for each, and a tool that is right for the first is the wrong purchase for the third.
Three situations, three answers
| Your situation | What you need | Build or buy |
|---|---|---|
| The answers live in documents you already have | A chatbot that answers from those documents and hands off to a person | Buy. Any decent tool does this |
| The answers live in your systems (orders, calendar, CRM) | A bot with tools that can look things up on a customer's behalf | Build, or buy a tool that integrates properly with your specific systems |
| The customer wants an action taken | An agent: tools that can book, quote, update, with approval where money or commitments are involved | Build |
The first row covers most small businesses, and it is the cheapest to get right. The third row is where the interesting work is, and it is also where a bad build does real damage, because the bot can now do things.
When you do not need one
- A good FAQ page would do. If the same five questions cover most enquiries, answer them on the page where people are looking. Search engines and AI assistants will read that page too.
- People want a human. In some businesses the first conversation is the sale. A bot in front of it costs more than it saves.
- The volume is small. A chatbot pays for itself on repetition. A handful of enquiries a week is a job for a person and a template.
- Your information is a mess. A bot trained on out-of-date pages will confidently give out-of-date answers. Fix the pages first.
Set up an off-the-shelf chatbot properly
If you are in the first row, you can do this yourself in an afternoon. The tool matters less than the routine.
Collect the real questions
Go through the last month of emails, messages and calls and write down the questions as people actually phrased them, not as you would phrase them.
Write the answers you would send
Short, plain, current. If two people on your team would answer differently, decide now. The bot will inherit whatever contradiction you leave.
Give it your documents and your site
Upload the help pages, the price list, the policies. Remove anything stale first.
Turn on the handoff
Anything the bot is unsure about goes to a person, with the transcript attached, and the customer is told that is happening. A bot that guesses is worse than no bot.
Read the transcripts every week
The questions it could not answer are your next FAQ entries. The answers people argued with are the ones to rewrite. Twenty minutes a week keeps it useful.
When a custom build pays off
Build when the bot needs to do something in your software. Looking up an order, checking a calendar, creating a quote from your price rules, updating a record in your CRM. At that point the bot needs tools, and tools need permissions, limits and logs. That is an agent, and it is engineering rather than configuration.
A custom build is also the answer when the questions can only be answered from your own systems and no rented tool integrates with them properly. Integrations that "support" your CRM often mean a nightly export, which is not the same as looking up a live record.
What good looks like
A chatbot that works has three properties. It answers from material you control and keep current. It hands off to a person cleanly, with the conversation attached. And someone reads its transcripts. Everything else, including which model it runs on, is secondary.
An agent that works has those three and two more: every action it can take is a small, permissioned tool on your systems, and anything involving money or a commitment waits for a person. I cover the engineering side in how to build an AI agent, and I build them under AI agents. If you are in the first row of the table, you do not need me. Set up the tool, follow the five steps, and read the transcripts.