What should an AI voice agent do when it doesn't know the answer?
Say so, capture the question, and route it — quickly, without a second guess. The dangerous failure is not silence, it is confident invention: a made-up price, an imagined warranty term, an arrival time nobody can keep. The reliable defense is structural. Define a closed list of things the agent is permitted to state as fact, and make everything else a capture-and-escalate path by design rather than by instruction.
Instructions are not guardrails
Telling a model not to quote prices reduces how often it quotes prices. It does not prevent it. Under pressure from a persistent caller, a model that has been asked the same question three different ways will often produce an answer because producing answers is what it does.
Guardrails that hold are outside the conversation. The agent can only read from a defined knowledge source; anything not in that source has no text to retrieve and the flow routes to capture. That is an architecture decision, and it is the difference between a system that usually behaves and one that reliably does.
Define the commitment set explicitly
Write down every category of statement the agent may make on your behalf. A typical list is short.
- Hours, service area, and what services you offer. Stable facts, safe to state.
- Appointment windows returned by the live schedule. Never invented, always read.
- Published flat fees where you actually publish them. If it is not published, it is not sayable.
- What happens next. Who will call, and roughly when — backed by a task that exists.
- Everything outside that list is a capture: record the question verbatim, tell the caller a person will answer it, create the follow-up.
Do not rely on the model's own confidence
Self-reported confidence is unreliable, and it is least reliable exactly where it matters — plausible-sounding domain questions. Use structural signals instead: whether a retrieval returned anything, whether required fields resolved, whether the caller has restated the same question after an answer.
Those are observable and testable. They also give you a log you can review, which self-assessed confidence does not. The same principle governs any automated action in an operational AI system: gate on evidence, not on the model's mood.
There is one more useful signal: the caller correcting the agent. If someone says "no, that is not what I meant" the flow should treat that as a hard stop rather than another attempt, because the second attempt is rarely better and the caller is already spending patience.
Make the unknowns visible
Every capture-and-escalate event should land somewhere a human reads. Cluster them weekly and you get a free content roadmap: the questions callers actually ask, ranked by frequency, with the exact wording they use. Some become knowledge base entries the agent can then answer. Some belong on your website, which is the same input that drives an answer hub. The rest stay human, permanently, and that is a legitimate outcome.
Topics: fallback · hallucination · guardrails · escalation
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.