Every Pitch Says AI. Here's What I Ask Instead

You sit in a swarm of pitch decks, and every one of them says AI, each claiming its version somehow wields magic. And yet you’re exhausted by all of it. The claims pile up into noise, and none of them tell you anything you can act on. You pause and ask yourself: this is madness, how am I supposed to tell which one uses AI better, which one has the sharper technology? You are asking the wrong question.

Here’s the thing nobody puts on a slide: an AI claim used to cost something to make. It doesn’t anymore. Any team can wire a model call into a product over a weekend, so “AI-powered” has stopped telling me anything about whether there’s a real business underneath it. A free claim should raise my scrutiny.

The claim that stopped meaning anything

I do spend real money on the AI stack: the interconnect fabric and the providers actually training frontier models. That’s a different bet with different economics.

The problem I’m describing lives one layer up: the application sitting on top of someone else’s model, selling a workflow, the “powered by GPT” or “powered by Claude” company. That’s where the free-claim problem lives.

I’ve also written about what founders get wrong in the room when they pitch me. This is the other side of that exchange: the four questions running in my head while they talk, and none of them are about whether the AI works.

The question that moves my price: who sets your COGS

The first thing I want to know is who controls the cost of serving each additional customer. In a SaaS company, marginal cost approaches zero. One more login costs almost nothing. In an AI application, marginal cost is a metered API call to a vendor who sets the price, changes it without asking you, and will keep moving it as their own economics shift.

Replit is the clean example. The Information reported that as Replit’s revenue scaled roughly 70x, gross margins swung between 36 percent and negative 14 percent, back and forth. That’s what happens when your COGS is somebody else’s pricing decision plus your own usage mix.

The standard founder answer is that inference gets cheaper every year, so today’s margin is a snapshot, not a trend. There’s real data behind that. Epoch AI has tracked the price of reaching a fixed capability level, GPT-3-level MMLU performance, falling from $60 per million tokens in November 2021 to $0.07 by October 2024. I don’t dispute that number. But Epoch’s own data shows the rate of that decline swinging from 9x to 900x a year depending on which benchmark you pick, and Replit’s margins went negative during a stretch when the industry-wide curve was supposedly working in every founder’s favor.

The average curve going down doesn’t mean your curve goes down. It means someone else’s roadmap and someone else’s pricing quarter decide where your curve points next.

The question about next Tuesday: what happens when the model gets better

Second, I ask what happens to the product the day OpenAI or Anthropic changes the model underneath it, because they will, on their calendar, not yours.

Model deprecation is routine. OpenAI gives roughly six months’ notice before retiring a GA model, and as little as two weeks for a preview model. Anthropic’s published policy is a 60-day minimum, and in practice the notices have landed every couple of months across its model families. A company built on one model’s specific behavior, its tone, its refusal patterns, its quirks, is re-qualifying its own product against someone else’s release schedule, repeatedly, forever.

There’s a sharper version of this risk than deprecation. In November 2023, OpenAI shipped GPTs and the Assistants API at its first DevDay, and TechCrunch named the mechanism at the time: a wide class of prompt-middleware businesses had just watched the platform absorb their entire feature into a checkbox. That’s a contemporaneous read of a mechanism, not a confirmed body count. But the mechanism hasn’t gone away. The platform doesn’t need to compete with you. It just needs to ship the thing you charge for as a default setting.

This is the question I was actually asking when I looked at Harvey, the legal AI company that raised at an $11 billion valuation in March 2026. Harvey doesn’t own a frontier model. It sells a workflow built on top of one. So the platform-risk question and the “what happens next Tuesday” question collapse into the same question: what happens to Harvey’s product the day a frontier model gets meaningfully better at legal reasoning on its own?

The question I ask last: what do you actually own

This is where Harvey is the honest example, not the cautionary one.

The bet underneath its valuation isn’t that GPT’s successors won’t get better at legal reasoning. It’s that Harvey’s workflow integration inside law firms, its data from real client matters, its compliance posture, and the switching cost of ripping it back out will keep outrunning the capability gap even as that gap closes.

I genuinely don’t know if that bet wins. Neither does Harvey, and to its credit nobody there pretends otherwise in the rooms I’ve heard about. The honest answer is that it’s unresolved, and it stays unresolved until either the model closes the gap faster than the moat compounds, or it doesn’t.

What I’m actually listening for is whether a founder can name the moat without pointing at the model: proprietary data nobody else has access to, workflow depth that took years to build into a customer’s process, switching costs measured in retraining time and compliance sign-off rather than a login, regulatory posture a general-purpose model can’t casually replicate. If the answer to “what do you own” is a description of how well you’ve integrated someone else’s model, that’s not defensibility. That’s a plugin with a valuation.

What I ask instead

None of this requires reading the deck twice. It requires asking four questions the deck doesn’t answer.

1. Who sets your cost of goods sold, and what happens to your margin if usage triples overnight?
2. What does your product stop doing well the day the underlying model changes?
3. If your model vendor deprecates the exact behavior your product depends on, what’s your notice period, and what do you have to rebuild?
4. What do you own that the model vendor doesn’t, and would you still have a company if a frontier model closed the gap tomorrow?

What the investor asks	What it’s actually testing
Who sets your COGS?	Margin control: whether unit economics are the founder’s to set, or a vendor’s
What happens when the model gets better?	Platform risk: whether the product survives being absorbed into the next frontier release
What happens when the vendor changes the deal?	Vendor dependency: exposure to deprecation and repricing on someone else’s timeline
What do you actually own?	Defensibility: what’s left once you subtract the prompts and the API

Fig 1. What each AI-pitch question is actually testing, drawn from my own experience as an angel investor and pitch coach. This is my own framework, not a survey result or published research.

The room is still full of decks that say AI. What’s changed is what I listen for: whether the business survives the day the technology stops being novel enough to matter on its own.

Next time a deck says AI, ask one of these four questions instead: which one can the founder actually answer, without changing the subject back to the model? The claim was never the question. It just used to feel like one.