AI & Automation

What I learned building a 24/7 AI chatbot

By Florent · Sep 25, 2026 · 8 min read

Most "AI chatbot for your business" posts are written by people who installed a widget and called it a day. This post is the opposite: I designed and shipped a real chatbot for a small business website — the one answering questions on my studio's own site — and I want to walk through the architecture that actually survived contact with real visitors.

The mistake everyone makes: AI-first

The default architecture people imagine is: visitor types → AI answers. Simple, and wrong for business use. A raw language model asked about your prices, your hours or your service area will cheerfully make something up. For a restaurant, a hallucinated "yes we're open Mondays" is a real-world failure with a real customer standing at a locked door.

The architecture that worked: three tiers

What I shipped instead is a three-tier flow, and I'd build it the same way again:

Details that mattered more than the model

Short replies, no markdown. Visitors read chat on phones; asterisks and bullet walls feel like documentation. The system prompt enforces plain, short sentences.

Never repeat yourself. The bot offers the WhatsApp handoff once per conversation, not after every reply. Repeating contact details reads as desperation.

An editable brain. The knowledge base and prompt are editable from an admin panel without touching code — because every week of real conversations teaches you what people actually ask.

Fail gracefully. If the AI API is down or slow, the visitor gets the human handoff links, not a spinner or an error. The worst outcome is a silent failure nobody notices.

What it's genuinely good for

After running it live, the honest scorecard: the chatbot earns its keep on after-hours triage and repetitive questions. The questions that arrive at odd hours used to be morning backlog; now most are answered on the spot or arrive pre-qualified. It does not replace a human for anything revenue-critical — and it shouldn't try.

What I'd do differently

Two things. I'd seed the knowledge base from real support messages earlier instead of writing it cold — the first week of live traffic rewrote half my assumptions about what people ask. And I'd put the API key behind a server proxy from day one rather than moving it later; it's a five-minute job at the start and an annoying one after launch.

Should you add one?

If your site gets real questions — bookings, quotes, "do you deliver to my area" — a constrained, KB-first chatbot is one of the highest-leverage upgrades available to a small business right now, and it's now a service I offer through Nexa Dev Studio. If your site gets three visitors a day, fix traffic first. The chatbot is a multiplier, not an engine.

Related: the other half of "answers instantly" is the site itself loading fast — The 3-Second Rule.