"But what if it makes something up?" You will hear this on nearly every voice AI sales call. It's not really a question about hallucination — it's a question about who's liable when the agent is wrong in front of my customer. Answer the surface question and you lose. Answer the real one and you close.
Here's the reframe, the guardrails to name, and the things that quietly kill the deal if you say them.
First, don't get defensive
The instinct is to argue the model is accurate. Don't. Arguing accuracy concedes the frame that the agent is a free-floating brain that might say anything. Instead, reframe it as a scoped system with guardrails — which is the truth.
The reframe (say this)
"Totally fair — and it's the first thing we lock down. The agent isn't answering from general knowledge. It only speaks from the information you give it: your services, your hours, your FAQs. When a caller asks something outside that, it doesn't guess — it does one of three things: confirms, defers, or hands off to a human. Let me show you exactly what that looks like."
Now you've moved from "trust the AI" to "here are the rails." Rails are concrete and demoable. Trust is not.
The three guardrails to name
1. Scoped knowledge. The agent answers from a defined knowledge base, not the open internet. Out-of-scope questions trigger a fallback, not an invention. (On Vapi/Bland/Retell this is your system prompt + knowledge config — say so plainly; technical buyers respect specifics.)
2. Confirmation on anything that matters. Names, numbers, appointment times, addresses — the agent reads them back. "I've got you for Thursday at 2, at 14 Oak Street — is that right?" A confirmed booking can't be a hallucinated one.
3. Hand-off on uncertainty or stakes. If the caller is upset, asks for something out of scope, or says the word "emergency," the agent stops trying to be clever and routes to a human with context. The escalation layer is the safety net, and naming it defuses the fear directly.
Then demo the failure, on purpose
The strongest move: ask them to try to break it. "Ask it something we didn't program — go ahead." When the agent says "That's a great question — let me have someone from the team follow up on that," the objection evaporates. You've shown the rail working live. A demo that survives a hostile question sells harder than any reassurance.
What NOT to say
- "It's 99% accurate." A made-up stat invites a made-up counter, and the 1% becomes the entire conversation.
- "AI never makes mistakes anymore." Everyone has seen one make a mistake. You just torched your credibility.
- "Don't worry about it." Dismissing the concern confirms they should worry about it.
- Deep technical jargon about temperature, tokens, RAG. The buyer doesn't care about your stack; they care about who answers for it. Translate everything into caller experience and liability.
The one-line version
When you only have a sentence: "It only speaks from what you give it, it reads back anything important, and it hands off to a human the moment it's unsure — so the risk isn't 'what might it say,' it's 'a missed call,' and we're fixing that."
The hallucination objection is a trust objection wearing a technical costume. Name the rails, demo one failing gracefully, and keep the conversation in the caller's world — not the model's.