Every small business owner asks the same question before they’ll trust an AI with real calls: “What happens when it doesn’t know the answer?”
It’s the right question. It’s just usually imagined backwards.
The failure people picture
Most people picture the AI grinding to a halt, repeating itself, or — worse — making something up to sound helpful. A caller asks about a price that changed last month, or whether a specific appointment slot is open, and the system just guesses. That’s a real failure mode. It’s also exactly why we didn’t treat the fallback as an afterthought.
Escalation is a design decision, not a safety net
The instinct is to make the AI smarter so it never has to say “I don’t know.” That’s the wrong target. No system — AI or human — should be answering questions it isn’t sure about, and the goal isn’t zero uncertainty. It’s handling uncertainty honestly.
Every call gets classified by what kind of answer it needs. General questions — hours, location, what services you offer, how booking works — get answered directly, every time, because the system is actually certain about them. Anything specific to an individual case — exact pricing for a nonstandard job, whether a particular slot is still open, anything that depends on information the system doesn’t have — routes to a human during business hours, or a structured callback request after hours.
The confidence problem
The failure that actually costs you the customer isn’t “the AI got something wrong.” It’s “the AI got something wrong confidently.” A caller forgives “let me get you someone who knows for sure” instantly. They don’t forgive being told something false and finding out later.
That’s why the system is built to recognize the edge of its own knowledge rather than talk past it. If a question falls outside what it can answer with certainty, it says so, and hands off — it doesn’t fill the gap with something plausible-sounding.
What this looks like on a real call
A caller asks if you can fit them in Thursday. The AI checks the calendar and answers directly — that’s certain, so it answers. A caller asks if you can match a competitor’s quote. That’s not something the AI should be guessing at, so it takes the details and gets a human to call back, instead of inventing a number.
The difference isn’t intelligence. It’s knowing which question it’s actually being asked.
Why this matters more than the demo
A slick demo answers everything smoothly because demos ask easy questions. Real callers don’t. The system that matters is the one that behaves well on the question it can’t answer — because that’s the call that decides whether you keep the customer or lose them to a competitor who picked up the phone.
Oleksii Kocherev
CEO, Optima Voice
AI Voice Agents for Business, Education & Government
Originally published on Medium.