We built a chat receptionist knowledge base in one afternoon

Illustrative scenario — one afternoon. Fifteen questions from the owner, one approved answer file, twenty hostile test questions — and a chat receptionist that knows exactly what it's allowed to say, and what to do when it doesn't.

Most of what makes a chat receptionist good isn’t the AI — it’s the answer file. Here’s the exact process we use to build one, and why it takes an afternoon instead of a month.

Watch it work in practice: a working sample that answers from approved business info and hands off to a human instead of guessing.

The problem: the same questions, every day

Your website gets visitors at 11pm. They have the same five questions your front desk answers all day — hours, pricing, booking, cancellations, parking — and there’s nobody there to answer. So they leave. Or they fill in your contact form, and you reply the next morning, and they’ve already moved on.

Meanwhile your team spends half the day retyping answers they’ve typed a hundred times, and every new hire learns the answers by getting a few of them wrong first. The questions never stop. The answers never change. The whole thing is a treadmill.

A chat receptionist gets off the treadmill: a visitor asks, the answer arrives instantly, the booking happens on the spot. Here’s the exact process we use to build one — and why it takes an afternoon instead of a month.

The step-by-step solution

Step 1: The 30-minute call

We ask the owner for their fifteen to twenty most-asked questions — the ones their staff answers on autopilot. Hours, pricing, booking, cancellations, parking, the weird one only regulars ask. Nobody knows this list better than the person who lives it.

Step 2: One approved answer file

We write up the answers — about two thousand words — and the owner approves every line. This file is the entire universe the receptionist is allowed to speak from. Anything outside it gets one response: “I don’t know — let me take a message.” That fallback is the most important sentence in the whole system.

Step 3: The router

A short system prompt tells the AI its job: match the visitor’s question to the approved file, answer only from it, and fall back the moment it’s unsure. This runs on cheap inference — the same class of models we measured at a fraction of a cent per decision — because routing is a matching task, not a reasoning marathon.

Step 4: Twenty hostile questions

Before anything ships, we write twenty questions designed to break it: trick phrasings, two questions in one message, questions with no approved answer, questions that almost match an approved answer but don’t. Every gap gets fixed — usually by sharpening the answer file, not by buying a bigger model.

The honest cost math: our measured inference cost per test decision is a fraction of a cent, so forty-odd test calls land well under a penny of AI spend (estimate). The afternoon goes to writing, approving, and testing — human work, reviewed by a human. That’s what our Basic chat package ($150) actually buys: not the AI, which is nearly free, but the answer file and the twenty hostile questions.

And the speed comes from reuse, not from skipping steps. The same approved file later feeds the voice receptionist and the follow-up automation — one source of truth, tested once, deployed three ways. That’s the subject of a later post in this series.