What happens if the AI can't answer?

The failure everyone fears is a confident wrong answer. A quoted price that is out of date, a policy that was never yours, a promise nobody can keep.

The defense is a hard boundary rather than a clever model. The assistant answers from what you gave it and treats anything outside that as a handover, not a guess.

This is why what you write at setup matters so much. A thin brief produces a lot of handovers; a thorough one produces a lot of completed calls.

First, it tries to recover in conversation. If the caller phrased something oddly, it asks once more in plainer words, which resolves a surprising share of near-misses.

Second, it checks whether this is a transfer case. You set those rules: a specific topic, a caller who asks for a person, or anything the assistant has no answer for.

Third, it takes a message. Name, number, what they wanted, in their own words rather than a summary, so whoever calls back has something to work with.

Fourth, it tells someone. A text to the on-call phone for anything you have marked urgent, which is how after-hours emergencies stop waiting until morning.

Say you own a flooring showroom. Your brief covers hours, the brands you carry, how measuring visits work and a rough price range per square foot. On Wednesday a caller asks, "Can you put vinyl plank over a heated floor?" Nothing you wrote covers radiant heat.

A badly built assistant would produce a plausible answer, and plausible is the danger here, because the right answer depends on the product and the system temperature. What should happen instead looks like this in the transcript: the assistant says it doesn't have that detail and won't guess, offers to have an installer call back, and takes the caller's name, number and the question in her own words. The entry is marked as handed over, with her sentence quoted.

You read it that evening, ring her back with the real answer, and add two sentences about radiant heat to the brief. Next month the same question gets answered on the call.

Don't take any vendor's word for this behavior, ours included. On a demo, ask for a price you never supplied. Bring up a service you don't offer. Then try something that needs professional judgment, like whether a symptom sounds serious or whether a contract clause is enforceable. You want to hear a clean refusal and a handover each time, and you want to hear it without a long apologetic preamble.

Then look at the human end of the chain, because that's where fallbacks quietly fail. A rule that transfers urgent calls at 2am is only as good as the person whose phone rings. Is it on loud? Does it break through do-not-disturb? Who covers when that person is on vacation?

Ask how you'll find the handed-over calls the next morning. If they're mixed in with sixty routine transcripts and nothing flags them, some will sit unread for a day, and a callback that arrives a day late has lost most of its value.

You choose what happens at each step, per topic. A dental practice usually transfers pain calls and takes messages about billing, because one is urgent and one is not.

The transcript shows the caller's own words at the point of handover, which is the useful artifact. A summary would smooth away exactly the phrasing you need to see.

Handovers cluster. After a week you will find three or four questions causing most of them, and each is a paragraph of setup away from being handled.