Every vendor who walks into your office this year will say the word "artificial intelligence" inside the first two minutes. Most of them are renting the same three tools you could sign up for this afternoon on a credit card.
I am not against any of it. We build these systems. That is exactly why I will tell you where they fall down, because the people selling them to you will not.
When I disagree with something at work, I do it in three stages. It has served me well for twenty years and it works just as well on a vendor as it does on a founder.
Stage one. The question
The question is always the same one. Have you thought this all the way through?
Nine times out of ten that is where it ends. Not because it is clever, but because saying the logic out loud is usually enough for a person to find their own hole in it. Ask a vendor what their AI actually does inside your workflow, and then be quiet. Genuinely quiet. Count to five.
If what comes back is a gesture at "our proprietary AI" with no named capability, no process and no example, you have not got a technology partner. You have got a subscription with a markup on it.
Where these tools genuinely fall down
Four things, and any vendor worth your money will list them before you do.
It is built to be average. A language model produces the most statistically likely next words. That is the precise opposite of sounding like you. Left alone it drifts to the same confident, frictionless, forgettable middle as every other agent in your market who bought the same tool the same week. Sounding like yourself is the one thing it cannot do on its own, and in this business sounding like yourself is most of the job.
It invents things, beautifully. It does not know the square footage, the taxes, the HOA terms or the school district, and it will produce all four in flawless English with total confidence. In a category with fair housing law sitting over it, an invented detail is not an oops. It is exposure, with your licence number printed underneath it.
It has no taste and no judgement. It optimises for what you asked, not for what the situation needs. It will never tell you the price is wrong, the photos are the problem, or that this campaign is quietly eating the last one. That remains stubbornly human work and I have yet to see anything change it.
The demo is not the job. Anyone can make this look miraculous for four minutes with a hand-picked example. Production is brand rules, edge cases, revisions, deadlines and a hundred assets a month at a consistent standard. The distance between those two things is where most of these claims quietly die.
Stage two. The paper
If the question does not settle it, I go away and come back with one page. Their position stated fairly, my position, and the three questions I cannot answer. I fight in writing because writing strips the charm out of the room, and charm is where a good salesperson lives.
So here is the page. Six things that should make you slow down.
Nobody can tell you who checks the work. If there is no name attached to who owns accuracy and compliance before something goes live, then nobody owns it.
They cannot name a single failure mode. This is the one I care about most. People who actually run these systems in production have scars and will show you them unprompted. A shop that tells you the tools "just work" has either never run them at scale or is not being straight with you, and I am not sure which is worse.
Everything is instant, unlimited and cheap. That is the price of raw output. Somebody still has to make it correct, on brand and legal. If that labour is not in the price it is not in the product.
Their whole portfolio looks the same. Scroll it. If ten clients share one look, the AI capability is a default setting, not a practice.
They go vague on the data questions. Who owns the output. Whether your material trains anybody's model. Where your client list lives. Hesitation is itself an answer.
They want the annual contract before the pilot. Anyone confident in the work takes thirty days with defined deliverables, because the work closes the sale for them. Insisting on the year first tells you where their confidence actually sits.
Stage three. The flat sentence
If it matters and they still will not move, I say the thing plainly so nobody can claim later that they were not told.
Here it is. Ask what it cannot do.
That is the single most revealing question you can put in that meeting, and it costs you nothing. A serious firm answers instantly and specifically, because they walk into those walls every week and they remember the bruises. A pretender improvises, and you will hear them improvising.
Then ask for a live run against your brief. Not a case study and not a sizzle reel. Your actual listing, your actual brand, this week, in the room. Real capability survives contact with your work. A rented demo does not.
Any road up
None of this is anti-technology. We build the stuff. The point is that "AI-powered" is not a capability, it is a sentence, and sentences are free.
Ask what it cannot do. Ask who checks it. Ask for thirty days. If the answers are good you have found somebody worth working with, and if they are not you have saved yourself a year.
Ta, bab. Put the kettle on.