What enterprise buyers actually ask in AI deals

My day job is enterprise AI go to market. I sit in rooms where large companies decide whether to put this technology in front of their customers and their employees, and I have watched hundreds of those conversations play out. Everything in this essay is pattern level. No company names, no deal names, no numbers that belong to anyone else. The patterns are the interesting part anyway.

If you learned about AI from the internet, you would guess that buyers ask about model quality. Benchmarks, context windows, which lab is ahead this quarter. They almost never do. In two years I can count the benchmark conversations on one hand, and every one of them was started by a vendor, not a buyer.

The demo is not the deal

Every AI deal starts with a moment of genuine wonder. The demo lands, the room goes quiet, someone says they did not know it could do that. Two years ago that moment was most of the sale. Today it is the entry fee. The wonder wears off in about a week, and what is left is a committee of people whose jobs depend on this thing not embarrassing them.

That committee is who you are actually selling to, and their questions have almost nothing to do with intelligence. They have to do with trust, ownership, and what happens on the bad day.

The five questions that keep showing up

Across industries, across company sizes, the same five questions surface in nearly every serious evaluation. Not always in these words, but always in this shape.

• Where does our data go, and does your model learn from it? This is question one in almost every room now. The buyer has usually been burned by a headline, not an incident. They want the boring answer, stated plainly, in writing.
• What happens when it is wrong, and who finds out first? Enterprises do not fear a model that is wrong. They fear a model that is wrong confidently, in front of a customer, with no human in the loop and no audit trail.
• Who in our organization owns this once it is live? The deals that stall are almost never blocked by technology. They stall because nobody inside the building wants to be the name attached to the thing. When a buyer has an answer to this question, the deal moves.
• What does my team stop doing? This is the ROI question in its honest form. Nobody buys hours saved in the abstract. They buy a specific task that a specific team no longer does, and they get suspicious when a vendor cannot name one.
• What does this cost when it works? Not the pilot price. The price at scale, when ten thousand employees use it daily. Usage-based pricing reads as a growth story to vendors and as an uncapped liability to CFOs, and that gap kills more renewals than quality does.

What changed in two years

In 2024 the questions were vocabulary questions. What is a token, what is a hallucination, is our data training your model. Budgets were curiosity budgets, carved out of innovation lines, spent on pilots that were really theater for the board. A pilot could succeed technically and still go nowhere, because it was never wired to a decision.

Now the buyer is literate. Procurement has AI clauses. Security teams have AI questionnaires that run to dozens of pages. Legal wants to know about indemnity for model output. The conversation moved from what is this to how does this fit inside the machinery we already run, and that is a much harder conversation for a vendor and a much better one for a buyer.

The other shift is quieter. Two years ago the champion in an AI deal was usually the most technical person in the room. Today it is just as often someone from operations or finance who has used these tools personally and has a concrete, slightly boring idea about where they fit. The boring ideas are winning. The moonshot use cases make the press release, and the document processing use case makes the renewal.

What this means if you are betting your career on it

I write a lot about careers on the other side of this site, so here is the bridge. If you sell software, you are already in the AI business whether your product is AI or not, because your buyer's questions are converging on the five above. The sellers who are pulling ahead are not the ones who can recite model names. They are the ones who can sit with a nervous committee and translate capability into what changes Monday morning.

That translation skill is rarer than technical skill right now, and it is what hiring managers are actually screening for when a job description says AI fluency. Not prompt tricks. The ability to hold the wonder and the fear in the same conversation and move a room from one to the other.

This is the first essay of Baldwin Builds. I am writing it because the distance between how AI is discussed in public and how it is bought, adopted, and lived with inside real companies is enormous, and I happen to stand in that gap all day. More soon.