An AI due-diligence checklist for boards and investors

Right now a company can put “AI” in a deck and add a zero to the ask. Some of it is real. Most of it is a thin layer over someone else’s model. Here’s how to tell them apart before you write the cheque.

You don’t need to be technical to do this well. You need to ask a handful of blunt questions and watch how confidently they get answered. Vagueness is the tell.

Is it a real edge, or a wrapper?

Plenty of “AI companies” are a nice interface on top of a model anyone can rent. That’s not always bad, but it’s not a moat. Ask what they have that a competitor couldn’t rebuild in a weekend with the same off-the-shelf model. If the answer is the model, they don’t own it. If the answer is their data, their distribution or their workflow, keep listening.

Who owns the model and the data?

If the whole thing depends on one external provider, they’re a tenant, not an owner. What happens to their margins and their product when that provider raises prices, changes the terms, or ships the same feature themselves. Ask it plainly.

Do the unit economics survive scale?

AI has a real cost per use. Some businesses look great at ten customers and fall apart at ten thousand because every query costs money. Ask for the gross margin per unit today and at 10x volume. If they haven’t modelled it, that’s your answer.

A demo tells you it can work once. Unit economics tell you whether it can work as a business.

Is the metric real, or a demo?

“95% accurate” on what test, chosen by whom. Ask to see it work on messy, real, adversarial inputs, not the happy path. Ask what the accuracy needs to be for the product to be useful, and whether they’re actually there.

Are people using it, or just trialling it?

Pilots are easy to sign and easy to abandon. Ask for retention and real usage, not logos on a slide. A wall of pilot customers who all quietly churned is a red flag dressed as traction.

The rest of the list

Who on the team has actually shipped this kind of thing before.
Where does customer data go, and does that create a regulatory or trust problem later.
What’s the plan when the underlying models get better and cheaper, which they will. Does that help them or gut them.

None of this requires you to understand transformers. It requires you to be unimpressed by the word “AI” and interested in the boring parts underneath it. That’s where the real answer lives.