Training your AI receptionist: the knowledge base explained
Agentric Team ·
An AI receptionist isn't psychic — it's a fast learner that reads exactly one textbook: the knowledge you give it. Feed it well and it answers like your best employee on their best day. Feed it nothing and it can only take messages.
The three layers of knowledge
1. Structured facts (the setup wizard collects these):
- Services: name, duration, price, description — what it quotes and books.
- Hours: when you work — what it can offer for bookings.
- Contact details: phone, email, address, website.
2. FAQs (question → answer pairs): The things customers always ask: "Are you Gas Safe registered?", "Do you charge for call-outs?", "Do you cover Wetherby?", "What's your cancellation policy?" Each FAQ is matched when a caller asks anything similar.
3. Free-text notes (the catch-all): Anything else in plain English: "For gas smells advise calling 0800 111 999 first", "we don't do oil boilers", "parking is on the street — mention permits", "offer 10% off boiler services in summer". The assistant reads these verbatim and applies them sensibly.
How the model uses it
Everything you write goes into the assistant's context on every call and chat. It's not pattern-matching keywords — it understands the knowledge and applies it to whatever the customer actually asks. "Do you lot do combi swaps?" → uses your boiler installation service and FAQ. "Can I park?" → uses your parking note.
What good knowledge looks like
- Write like you talk. "Call-out is £95 for the first hour, then £60/hour" beats a formal paragraph.
- Cover the top 10 questions. Check your texts and emails — the questions that repeat are your FAQ list.
- State boundaries explicitly. What you don't do matters as much as what you do — it stops the assistant improvising.
- Keep prices honest. It quotes verbatim; "from £40" is fine, mystery is not.
The maintenance habit
Ten minutes weekly at first: skim transcripts, find the question it stumbled on, add the answer. The knowledge base compounds — within a month it handles questions you never explicitly taught, by combining what it knows. That's when it stops feeling like config and starts feeling like staff.