Any vendor can play you a smooth recording. What you're buying is what the agent does on the calls that go badly, because those are the ones that carry your name.
Here are five calls to make yourself, live, before you let anything near a real lead. All five together take about twenty minutes and none of them require you to know anything technical.
Let the agent start its opening, then talk over it about four seconds in. Ask something completely off-script. What year was your company founded, or who's the owner.
What you're watching for is whether it stops. An agent that plows through its opening while you're talking is not listening, it's playing. Real people interrupt constantly, and a seller who feels talked over hangs up.
Then watch whether it recovers. After it handles your interruption, does it get back to what it was doing, or has it lost the thread entirely?
Answer with an edge. Who is this. How did you get my number. I've had four of these calls today.
A good agent does two things: it answers the actual question about where the number came from, and it drops its energy to match yours instead of staying relentlessly upbeat. An agent that stays cheerful while somebody is irritated reads as a machine that doesn't care, and it is.
Then say you're not interested and see whether it tries a rebuttal. One graceful attempt is a sales call. Three is harassment with your logo on it.
Say take me off your list. Then, on a second call, try the versions people actually use. Stop calling me. Don't contact me again. Lose my number.
All of them should stop the call immediately, without a question, without a retention attempt, and without asking you to confirm. If any version gets a different result than the others, the agent is pattern-matching one exact phrase and it will miss the real ones.
Ask the vendor afterward to show you the record. The opt-out should exist somewhere you can see, and it should survive the next time that number gets uploaded.
Give it something contradictory. Say you want to sell fast, then say you're not in a hurry. Say you own it outright, then mention the mortgage.
What you're testing is whether it notices. A good agent circles back and asks about the contradiction, gently. A weak one records both answers and hands you a lead sheet that contradicts itself, which you won't discover until you're standing in the kitchen with the seller.
This is the test most people skip and it's the one that costs the most later. Bad data from a qualifying call doesn't announce itself. It shows up as a wasted appointment three days out.
Ask something genuinely outside its knowledge. What's the current tax assessment on my parcel. Will you cover my mother's storage unit. Do you buy in Alaska.
The only acceptable answer is some version of I don't know, let me get you to someone who does. What you are checking for is invention. An agent that makes up a number, a policy or a commitment has created an obligation with your name on it, and you'll be the one explaining it.
Ask the vendor directly: what does your agent do when it doesn't know? If they say it never happens, that's the answer to a different question and you should press.
On every one of the five calls, count how long before it tells you it's an AI. Unprompted, inside the first ten seconds, every time. If it only discloses when asked, that's not disclosure.
After the five calls, go look at what the agent actually wrote. Not the transcript, the structured fields. Is the motivation captured or just the words? Is the timeline a date or a phrase? Is the appointment on a real calendar?
A perfect-sounding call that produces a useless record has not saved you anything. The output is the product, and the conversation is just how it gets there.
Not on the vendor's demo line. On a number that behaves like a real seller's, from where you actually are, at a time when nothing has been warmed up for you.
And record all five. When you get to the second vendor, and the third, you'll want to compare the same five calls side by side rather than compare your memory of a good feeling.
Run at least one of the five from a place with background noise. A car with the window down, a kitchen with a TV on. That's where sellers call from, and a system tuned in a quiet room behaves differently when it isn't in one.
If any real share of your market speaks Spanish, test in Spanish. Not a translated demo, a live call. Ask whether it handles a caller who switches between the two mid-sentence, because plenty of people do and an agent that panics at the switch will lose them.
And test somebody who talks slowly, and somebody who talks over themselves and restarts. Both are extremely common on a call about selling a house, and both break agents that were tuned on clean speech.
Make a five-row grid before you dial: one row per call, and columns for did it stop, did it disclose inside ten seconds, did it notice, did it invent anything, and what landed in the record.
Fill it in as you go. By the third vendor you will not be able to separate call two from vendor one and call two from vendor three, and you'll end up choosing on which demo felt smoothest. The whole point of the grid is to make you choose on the failures instead.
A vendor who fails one row honestly and tells you they're working on it is often a better bet than one who passes all five on a call that was clearly rehearsed. Ask when each behavior was last changed. Recent changes mean somebody's still working on it.
Because these five cover the four ways an AI phone call actually fails: it doesn't listen, it doesn't stop, it doesn't notice, and it makes things up. Everything else is polish.
Twenty minutes now, against a system that's going to talk to every seller who calls you. It's the cheapest diligence available.
Every follow-up process looks fine when the team is full and the volume is normal. Here is how to find the point where yours quietly stops, before a bad week finds it for you.
Consent, disclosure, opt-out and calling hours. The four questions that separate a vendor who built this carefully from one who did not, and what a good answer sounds like.