Is the AI real? What I ask before an investor backs it

Almost every pitch deck now has AI in it. That isn’t a problem in itself. The trouble is that the word covers everything from a company that trained its own model on data nobody else has, to a company that sends text to someone else’s model and shows you the answer. Both can be good businesses. They carry very different risks, and the deck rarely tells you which one you’re looking at.

Regulators have noticed. In 2024 the US Securities and Exchange Commission settled charges against two investment advisers for claiming to use AI in ways they didn’t, and it has a name for the practice: AI washing. Most startups pitching for seed money are nowhere near that. But the gap between the pitch and what is actually running is often wide. It’s cheap to check before the money goes in and expensive afterwards.

Here is what I ask. None of it needs an ML specialist, which is part of the point.

Whose model is it?

First, what’s actually running. Is it a model they trained, an open model they adapted, or a call to a provider such as OpenAI or Anthropic? I want to see it in the code, not on a slide.

Using someone else’s model isn’t a weakness. For most startups it’s the right call. But it changes the question. If the model isn’t theirs, the value has to be somewhere around it: their data, their workflow, their route to customers, the way they’ve fitted it to one job. I want to know what that is, and whether a competitor with the same API key could build it in a month.

What happens when it’s wrong?

Every AI system is wrong some of the time. What matters is what happens then. A feature that drafts an email for a person to check and a feature that sends it are different products, even if they share every line of code but the last.

Then I ask how they know when it’s wrong. A good team keeps a set of real examples with known good answers, scores the system against them, and runs that check before every change. A team without one is guessing, and it guesses again, in front of customers, every time the provider updates the model.

If the honest answer to “how do you know it works?” is “we try it and see”, what you’re looking at is a demo.

Is there a person behind it?

Some AI products quietly rely on people. In January 2025 the SEC found that most drive-thru orders taken by one company’s voice AI needed a human to step in. More often it’s staff who check, correct or redo the output before the customer sees it. That’s a perfectly reasonable way to start. It’s not a reasonable thing to leave out of a pitch, because it changes the economics.

I ask who touches the output, how often, and how much of it gets corrected. That work is a cost, and it belongs in the margins.

What does each use cost?

With most software, serving one more customer costs almost nothing. With AI, every request has a real cost from the provider, and the bill grows with use. That’s why AI-heavy companies often run at lower gross margins than traditional software.

I look at the cost per task, what the customer pays per task, and what happens to both if usage doubles. If a feature loses money each time it’s used, more usage means bigger losses. I don’t audit the financial model. I give whoever does a real cost per task to test it against.

What if the provider changes the terms?

A company built on someone else’s model depends on that provider’s prices, rate limits and policies, and on the model staying available at all. Providers retire older models on a published schedule. I ask how hard it would be to switch, and whether anyone has tried.

A team that has run its product on two providers, even briefly, has usually built the checks described above. A team that has never tried is usually more locked in than it thinks.

Can they use the data?

If the edge is data, I want to know where it came from, and whether the system treats it the way customer contracts say it should. Whether those terms hold up legally is a question for your lawyers.

What a good answer sounds like

None of these questions has one right answer. A seed company calling an API, with people checking every result and margins that only work at scale, can still be a good investment.

What worries me is a thin product described as a deep one, by founders who believe it.