One question tells you more about an AI vendor than any demo, any case study, any reference call they arrange for you. Almost nobody thinks to ask it.
Not "does it work?" Not "who else uses it?" Just this: do you run your own company on this, or do you only sell it?
It goes unasked because it sounds impertinent, and because the standard diligence covers everything around it. Buyers check whether the thing functions. They check who else has bought it. They call the references they are handed. Every one of those questions is about the product, and every one of them can be answered honestly by someone who has never once had to live with what they are describing.
The instinct you already have, quietly suspended
Everywhere else in life, this is not a subtle point. You would not trust a kitchen to a chef who won't eat what comes out of it. The instinct is old and it is sound: the person worth listening to is the one who has paid for their opinion by living with the consequences of being wrong.
So why does that instinct switch off around AI? Because AI arrives as a specialist subject, and specialist subjects invite deference to fluency. Someone who can explain clearly how the technology works sounds a great deal like someone who knows what it is like to run a business on it. Those are different kinds of knowledge, and the distance between them is precisely where the money goes. Fluency about a thing is not experience of it. It never has been, in any other domain, and nothing about this one makes it the exception.
Selling it and operating on it are two different crafts
Only one of those crafts is the thing you actually need, and it is not the one that shows up in the room.
Selling it means knowing the promise: the clean story, the demo that runs, the number that makes a board lean in. That is real work and some people are excellent at it. Operating on it means knowing the reality. The exceptions that break it. The morning it did the wrong thing and someone had to answer for it. The unglamorous work of making it survive an ordinary week, in an ordinary month, against ordinary staff turnover.
Notice what sits in the operator's knowledge that cannot sit in anyone's deck. Where the thing does not pay, and they stopped. What they switched off after four months and never turned back on. The process they refused to automate because the failure mode was one they could not accept, whatever the savings looked like on paper. What it costs to keep running in month nine, when the person who understood it has left. None of that is proprietary and none of it is secret; it is simply unavailable to anyone who has not lived it.
That second kind of knowledge is the one you are trying to buy. There is no studying for it. You get it by living through it, or you don't have it.
The tell is the tense
Once you listen for it, the difference announces itself in grammar.
The one who only sells speaks in the future and the conditional: imagine, could, up to, someday. The one who runs on it speaks in the concrete past: we tried this, it broke here, we do it differently now. It is almost impossible to fake in either direction, because the past tense requires events and the events either happened to you or they didn't.
So ask either of them for the failures. The operator has a list and gives it without flinching, because the failures are not an embarrassment to be managed; they are how the knowledge was acquired in the first place, and they know exactly what each one taught them. The theorist has a brochure.
There is a second tell inside the first, and it is worth waiting for. When a seller who does not operate is pressed on failures, the failures they produce are always someone else's, or they are general enough to be safe: adoption is hard, change management matters, data quality is a challenge. An operator's failures have dates on them. They have a system named, a decision that went badly, a number attached, and a thing they do differently now as a direct result.
What a case study actually is
Hold a case study up to the light and ask what it is evidence of.
It is proof that someone else, somewhere, survived the thing. It has been curated, selected for its outcome, and stripped of the conditions you would most want to inspect. It tells you the technology can work, for some company, under circumstances you are not permitted to examine. That is not nothing. It is also not what you are being asked to bet on.
The operator has no need to point at another company. They are the company. They didn't study the proof; they are the proof.
Apply it to everyone in the room
The test is only worth having if it has no exceptions, so give it none. The platform vendor. The consultancy. The big familiar name whose logo does half the arguing. The confident newcomer.
The question is not "are they clever," because they are. It is not "is the technology impressive," because it is. The question underneath all of it is whether they have put their own operation, their own money, and their own risk on the very thing they are asking you to bet yours on.
The one objection worth taking seriously
There is a fair answer to the question, and you should know what it sounds like, because the good ones will have it ready.
Sometimes the seller's business genuinely does not resemble yours. A firm of thirty people does not run the payables of a company of thirty thousand, and it would be dishonest to pretend the two are the same problem. An operator can say so precisely, and the precision is the whole tell: the part of their own operation that does run on it, the exact point where their situation stops resembling yours, the piece they cannot speak to from experience, and who can. That answer names its own limits, and every one of those limits is something you can go and check.
So listen for what the objection is doing. In the operator's mouth it draws a line around a real body of experience and hands you the map of where it ends. In the theorist's mouth the identical sentence is a shield, and the difference in the business becomes the reason no concrete thing that ever actually happened has to be produced at all. One answer narrows a claim. The other excuses the absence of one.
The price of a theory
Someone selling a way of working they have not adopted is selling a theory and calling it experience.
The price of that theory is that you become the experiment. Not a metaphorical one. The experiment runs at your expense, on your timeline, in front of your customers, and it is your people who spend the year finding out where the promise stops matching the world.
And here is the part that should sharpen the question rather than merely justify it. The seller learns from that experiment either way. Everything your failure teaches them, they will carry into the next room, and they will sell better next quarter because of what it cost you this one. You will have paid for their operating knowledge and received none of it. The one asset the engagement reliably produces is exactly the thing you came in to buy, and it accrues to them.
That is the whole reason the question is worth the awkwardness of asking it. Someone must go first, live with the consequences, and turn the promise into knowledge. The only question on the table is whether that has already happened, or whether you are about to fund it.
The question to take into the room
So before you buy the transformation, make the seller answer for their own. Not the customer they will name. Not the demo they have polished. Their own house.
Do you run on this yourself? Where has it burned you, and what did you change?
And if the honest answer is that they will sell it gladly but would never run their business the way they tell you to run yours,
why is their advice worth any more than a recipe from a chef who won't eat their own cooking?