How we build an AI chatbot — and what it taught us
From brief to launch. Which stack we choose, why the system prompt eats the most time, and the mistakes we won't repeat.
The start: a good brief is half the work
Most companies that want a chatbot answer the same questions every day — delivery, returns, opening hours, prices. The goal is always the same: let the bot handle the common questions on its own, and let a human deal only with what really needs them.
The first thing we do isn't code. It's a map of questions: what customers ask most often, where the company already has answers and where it doesn't. Without this prep, every chatbot is just an expensive toy.
The stack we choose
- A modern web framework + a database as the knowledge base — a source of truth the company can edit itself
- An advanced language model to generate answers — we pick whichever is the most accurate for factual answers at the time, not by brand
- A streamed response, so the user sees the text appear gradually instead of ten seconds of silence
- Contact capture on escalation — when a question exceeds the knowledge base, the bot politely asks for an email
What takes the most time
It's not wiring up the API. It's writing the system prompt — the rules the bot answers by. It typically goes through several iterations before it stops being overconfident and learns to say "I don't know this, let me connect you with support." A chatbot that makes things up does more harm than good.
Mistakes we won't repeat
- We don't build it without an admin panel. If the company can't edit the answers itself, it has to call us for every change. The knowledge base has to be editable without a deployment.
- Rate limiting from day one. One link on a forum and suddenly you have hundreds of messages in a few minutes. A limit on messages per minute per IP is a must.
The takeaway
A good chatbot isn't about the model. It's about how precisely it knows what the company should answer — and when it should honestly admit it doesn't know. The technology is available to everyone today; preparation and honesty make the difference.