Before you sign, I'll tell you what's actually in the code, what it costs to run, and who on the team is holding it up. And if you buy it, I'll stay to build the team that makes the AI pay off, then hand it to them and leave.
The point isn't to kill deals. Most companies with AI in the deck have something real in there. The question is how much, what it costs to run, and what it'll take to make it hold up. Sometimes the answer is "a good team with a thin wrapper and a real plan," and that's a fine thing to buy at the right price.
I talk to the engineers, not only the founders. I've built this stuff, so I know which engineers are load-bearing and how much of the demo actually works.
Is there real engineering between the user and the model? Wrappers aren't automatically bad. Wrappers priced like a moat are.
Red flagThe "proprietary AI" is one API call and a prompt a competitor could copy over a weekend.
Which models, trained by whom, under what license, and what happens if the vendor changes price or terms next quarter.
Red flag"Our model" is a vendor's model with a fine-tune the company can't move.
Where the data came from, and whether customer contracts allow what the product is already doing.
Red flagThe data advantage is customer data the contracts don't clearly cover.
What one request, user and customer cost to serve, and how that moves at 10× volume, next to your margin model.
Red flagInference cost grows faster than revenue, and nobody has run the numbers.
Do they measure quality on production data, or only on the demo? How do they know when an update made things worse?
Red flagNo evals, or an accuracy number nobody can reproduce.
How locked in they are to one provider, and who actually built the AI, can change it, and is staying.
Red flagThe whole system lives in one engineer's head, and their options vest at close.
How the system is really put together, where the AI sits, and what's custom versus bought.
Code quality, test coverage, how they ship, and whether the model layer can change without breaking things.
Data lineage and rights, quality on production traffic, and who notices when the model is wrong.
Conversations with the engineers who built it: who's load-bearing, who's leaving, and whether the culture survives an acquisition.
Inference cost per customer against your margin model, the memo, the risk register, and a walkthrough with the deal team.
Was the AI in the code, or just in the valuation?
Calling something AI used to be marketing. Now it's a representation. The SEC has a dedicated unit looking at it, and the FTC has brought cases. Nate raised $42M on AI-powered checkout while humans did the work. For a fund, the question someone will ask later is: what did you check, and who did you ask?
Turn the risk register into a plan. Measure delivery, quality, AI spend and how the people are doing. Fix the serious findings first.
One team running the AI workflow with review gates and cost controls, one real improvement shipped, and key people retained on purpose.
A before/after scorecard for the board and the operating partner, and an in-house team that runs it.
The device version adds RF, factory and certification. I tested iPhones to FCC specs and designed the factory testing so only compliant iPhones and iPads shipped, built the wireless test lab for what became Google Nest WiFi, and defined the FCC specs for software-defined radio.
Machines build. People direct. Customers decide.
Happy team, happy customers. Let's have some fun and make some money.
If you are growing a team, changing how engineering works with AI, or navigating a difficult technical transition, I would love to hear what you are building. Based in Denver; working with teams across the U.S.