Starbase / 01 PlainAgent

PlainAgent / How to build an AI chatbot

Guide · Chatbots

How to build an AI chatbot

A chatbot is the easiest kind of AI agent to start with: someone asks, it answers. The hard part isn't the AI. It's giving it the right knowledge and knowing when it should stop and hand over to a person.

1. Decide what the chatbot is for

Pick one: answering customer questions, booking calls, helping staff find internal information, or guiding visitors to the right page. One chatbot, one purpose. A support bot that also tries to sell usually does both badly.

2. Collect the real questions

Pull 30 to 50 real questions from your inbox, tickets, chat logs or sales calls. These become your test set and show you which documents you need.

3. Prepare the knowledge

A chatbot can only be as good as the content behind it. Check that your docs actually answer the real questions. Remove outdated pages: old prices in a forgotten PDF will end up in answers. Short, clearly written facts work better than long marketing pages.

4. Pick a builder

  • Inside an AI app: Custom GPTs (ChatGPT) or Claude Projects, for internal use or sharing a link.
  • On your website: chatbot platforms such as Voiceflow, Botpress, Intercom Fin or Tidio, which give you a widget and handoff.
  • Inside Microsoft 365: Copilot Studio.
  • Custom: an LLM API with retrieval (RAG) if you write code.

Features and prices change often; check each vendor's current pricing page.

5. Write the instructions

Tell it who it serves, to answer only from the knowledge, how long answers should be, and exactly when to hand off. Start from the support chatbot template or the FAQ chatbot template.

6. Set up the handoff

Decide where conversations go when the bot can't help: a ticket, an email, a live agent. Make "talk to a person" always work.

7. Test, launch small, read transcripts

Run every real question. Launch on one page, read all conversations for a week, fix the docs where answers were wrong, then widen.

Common mistakes

  • No handoff, so customers get stuck in loops.
  • Letting it answer from general knowledge, so it invents policies.
  • Launching everywhere at once without reading transcripts.

Ready-made support workflows for n8n are on RecipeAgents. See also chatbot examples.

Sponsor slot · availableFeature your product here
Newsletter

One agent template a week

A short Friday email: one agent spec, the system prompt, and what to test before you switch it on. Free, and the unsubscribe link is in every email.