What happens when your AI doesn't know the answer
Agentric Team ·
The worst thing an AI receptionist can do is confidently make something up. The entire safety model is built around a simpler rule: know what you know, and handle the rest honestly. Here's how that works.
The guardrail hierarchy
1. Knowledge base first. The assistant answers only from what you've written — services, prices, FAQs, notes. If the answer's in there, it uses it. If it isn't…
2. Honest uncertainty. The trained response is "I want to make sure you get the right answer — let me take your details and the team will confirm." Never a guess, never a hallucinated price.
3. Capture and escalate. Whatever the caller needed becomes a structured enquiry for you — name, contact, question — flagged appropriately.
4. Human path. For anything the caller needs a person for, the escalation rules kick in — handoff or callback.
Why this design beats "smarter AI"
You could build an assistant that tries to answer everything — and get an assistant that's wrong confidently, promising prices you don't charge and policies you don't have. The honest-uncertainty model trades a small amount of "I don't know" for zero hallucinated commitments. For a business whose name is on the line, that's obviously correct.
How the boundary sharpens over time
Every "I don't know" is a signal: the gap is visible in your transcripts and enquiry queue. Add the answer to the knowledge base, and that question never stumps it again. The knowledge base is designed to grow — the first month of transcripts usually surfaces the 15 questions you forgot to write down.
What this looks like in practice
Caller: "Do you do asbestos removal?" Assistant: "That's not something we cover — let me take your details and Sam can confirm or suggest someone. Can I grab your number?"
Honest, helpful, captured. That's the correct response — not a guess.