The Off-Script Test: Why AI Voice Agents Break When Callers Ask What Isn't in the Prompt

You build the prompt. You map the conversation flows. You test it with the dozen scenarios you planned for. Inventory questions, appointment booking, lease vs. finance, trade-in values. It handles all of them. You feel good about it.

Then a real caller gets on the line and the first thing they say is: "I just opened a recall notice about the transmission control module on my 2022 Tahoe, and I need to know if I can still trade it in while that's outstanding, and also is the recall repair going to take more than a day because I can't be without a car."

None of this was in your prompt design doc. The recall. The trade-in-with-open-recall question. The repair timeline. The "I can't be without a car" constraint that changes the entire interaction.

A good human BDC rep would say, "Let me check on that recall timeline for you, and yes, you can absolutely trade it in — we'll factor the open recall into the appraisal, it's not a dealbreaker." They'd pull the service advisor into the conversation. They'd handle it.

Most AI voice agents would say something like, "I can help you schedule a service appointment for that recall. Would you like me to transfer you to service?" Which answers a question the caller didn't ask, ignores the two they did, and broadcasts that this system can't handle conversations that go off the map.

The prompt boundary problem

Every AI voice agent has a knowledge boundary. Inside the boundary: the prompt, the training data, the integrated systems. Outside: everything else. The real world. The specific, strange, practical questions that callers bring to every dealership interaction.

The boundary is unavoidable. You can't prompt for every possible caller scenario. The question is what happens when a caller crosses it.

Most systems have three responses. Option one: deflection. "I'm sorry, I didn't quite catch that. Could you rephrase?" The caller rephrases. The AI still doesn't know. The caller gets frustrated and either hangs up or demands a human. Option two: hallucination. The AI generates a confident answer based on its language model's general knowledge, which may or may not be accurate for this specific dealership, this specific recall, this specific situation. Option three: escalation with no context. The call gets transferred to a human who has no idea what was already discussed, so the caller repeats everything and the "time-saving" AI interaction saved nothing.

None of these are acceptable. But they're what most AI voice agent platforms ship with.

The callers who find the edges first

The problem isn't evenly distributed across all callers. Most calls follow predictable patterns. Someone wants to know if a vehicle is in stock. Someone wants a lease payment. Someone wants to schedule service. These are the calls that demos look great on.

The edges get found by specific types of callers, and they're often the callers you most want to convert.

The well-researched buyer who has been reading forums and knows more about the vehicle's known issues than the salesperson does. They ask about specific TSBs and option codes. They're not trying to stump the system — they just know what they're talking about and expect the person (or AI) on the other end to keep up.

The caller with an unusual situation. A lease that's in a business name but the driver is an employee. A trade-in with a lien from an out-of-state credit union. A customer who bought three cars from your dealership over the years and wants to know if that history matters for pricing. These aren't edge cases to the callers — they're just their lives. But they're edge cases to the AI's training data.

The opportunistic caller who heard something from a friend and wants to verify. "My neighbor said you guys are doing 0% financing on leftover 2025s, is that true?" The answer might be no. It might be yes for certain trims. It might have been true last month but not anymore. The AI needs to know which answer is real and deliver it without making up a promotion that doesn't exist.

A human rep can navigate all three of these. They might not have every answer immediately, but they know how to say "let me check" without sounding like they've hit a wall. The AI doesn't have that muscle.

The hallucination trap

This is where off-script questions get dangerous, not just annoying.

When a language model doesn't know something, it doesn't default to "I don't know." It defaults to generating the most statistically probable response based on its training data. For a general-purpose model, that training data includes millions of web pages, forum posts, and articles about cars, dealerships, and automotive topics.

So when a caller asks about the transmission control module recall on a 2022 Tahoe, the AI might generate a response based on forum discussions about similar GM recalls. It might cite a timeline that was accurate for a different model year. It might describe a repair process that doesn't match your service department's actual protocol.

The answer sounds confident. It sounds specific. It's also wrong.

The caller takes that information as official — it came from the dealership's phone system — and makes plans around it. When they show up and reality doesn't match what the AI told them, the trust damage isn't with the AI. It's with your store. You're the one who deployed the system that gave them bad information.

Why context transfer matters more than the answer

There's a scenario that's actually worse than getting the answer wrong: getting the escalation right but dropping all the context.

The AI recognizes it can't handle the recall question. It transfers to a human. Good call, in theory. But the human picks up and says, "Service department, how can I help you?" They have no idea who the caller is, what vehicle they're asking about, or what was already discussed with the AI.

The caller sighs and starts over. "I was just explaining to the automated system that I got a recall notice about my 2022 Tahoe's transmission module..."

The AI interaction didn't save time. It cost time. The caller is now more frustrated than if they had reached a human directly, because they had to explain the situation twice.

Proper context transfer means the human agent sees a summary before they pick up: "Caller has a 2022 Tahoe, open recall on transmission control module, wants to know: 1) repair timeline, 2) whether they can trade it in with open recall, 3) they need a loaner if repair takes more than a day." That's three sentences that turn a restart into a continuation.

Most AI voice agent platforms don't do this. The escalation is a transfer, not a handoff. The difference is everything.

What the fix requires

Building an AI voice agent that handles off-script questions isn't about a bigger prompt or a better language model. It requires architectural decisions that most platforms haven't made.

The system needs to know what it doesn't know. This sounds obvious but it's technically non-trivial. Confidence scoring on generated responses, especially for factual claims about specific vehicles, recalls, promotions, and policies. When confidence drops below a threshold, the system needs to flag it before speaking — not after delivering a wrong answer confidently.

The system needs graceful degradation paths that sound human. Not "I don't understand" or "please rephrase" or "I'm sorry, I can't help with that." Something closer to: "That's a good question about the recall timing. Let me connect you with our service team — they'll have the exact answer for you, and I'll make sure they have all the details from our conversation." This preserves the caller's confidence in the system even as it admits its limits.

The system needs genuine context handoff. Not just transferring the call. Transferring the conversation — the identity, the history, the open questions, the customer's emotional state. When the human picks up, they should know the caller's name, vehicle, situation, and what's still unresolved.

And the system needs continuous learning. Every off-script question that gets escalated should feed back into the knowledge base. If callers keep asking about that transmission control module recall, the AI should learn to answer it. Not through manual prompt updates. Through automated pattern recognition that identifies frequent out-of-scope topics and expands the system's reliable knowledge boundary.

The test that actually matters

Forget the scripted demos. The test for any AI voice agent is simple: hand it a phone call from a real person with a real, specific, slightly unusual question about a real vehicle situation.

Does the AI recognize when it's out of its depth? Does it handle the gap without sounding broken? Does it hand off to a human with enough context that the conversation doesn't restart?

Most systems fail one of these three. Most buyers don't test for them because the demo only covers the happy path. But the happy path isn't what separates an AI voice agent people trust from one they hang up on. The edges are. And most AI voice agents still don't have edges — they have cliffs.