How VCs Actually Evaluate AI Startups (It's Not What You Think)

Most founders think they know how VCs evaluate AI startups. Pedigree matters. Using the latest model matters. A big TAM matters. A demo that looks impressive matters. Check those boxes and you’re in.

Mercedes Bent, who spent six years at Lightspeed building the muscle to spot exceptional founders, has a different answer: almost every company will check those boxes. Which is exactly why they don’t matter.

The real evaluation happens in layers, and the first layer is brutal.

The Founder Filter: Separating 0.01% from the Top 10%

Mercedes has a seven-point framework for founder evaluation. The characteristics: urgently dissatisfied, resource magnet, learning animal, deep sector expertise, influential communicator, relentlessly resourceful, and decisive strategic execution machine.

“A lot of people are those things. But the skill set of an investor is being able to separate what the top 0.01% look like in each of those categories versus the top 10% or the top 25%,” she explains.

This is the part that can’t be gamed. You either have it or you don’t, and Mercedes spent six years at Lightspeed building pattern recognition for what 0.01% looks like. It’s not confidence. It’s not charisma. It’s the ability to operate in extreme uncertainty while moving fast and learning obsessively.

“Being generally impressive is not enough. They need to be exceptionally great.”

The Product Layer: Does It Feel Magical?

After founder, the second filter is product. And Mercedes has a specific test: does this product feel magical on first use?

Not: does the code work? Does the metric improve? Does the demo look impressive?

But: when I use this for the first time, does it pull me emotionally—either through utility or through entertainment? Is there a clear value prop? And does it have a retention mechanism that keeps users coming back?

The third layer is market. Not TAM. Market conditions. “Market doesn’t always have to be super large to start with. It needs to be rapidly growing and changing,” Mercedes says.

She references Uber. When Uber started, the black-car market was tiny. By every TAM-based model, the investment wouldn’t have made sense. But a tailwind was moving: mobile phones made calling a car frictionless. Behavior was shifting. That’s a market worth betting on.

The Vision Layer: Can They Paint the Future?

The hardest layer to evaluate is vision. Can the founder see 10 years forward? Can they draw a path from today to there?

“They might not know exactly what the world’s going to look like in 15 years, but they can have a hypothesis,” Mercedes says. “They should be able to infect the listener on the other side of the table from them with their energy and passion and enthusiasm, that that vision is a possibility.”

This is where Mercedes practices: she runs a sci-fi club where she reads science fiction and philosophies to keep her brain trained for imagining futures. “I need to do futurist exercises to stay pliable in my mind,” she says.

When a founder pitches a vision, she’s not fact-checking. She’s checking if the hypothesis is credible—if it’s something that could actually happen given current trajectories.

What Actually Fails: The Researcher Who Doesn’t Talk to Customers

Mercedes has watched brilliant researchers fail because they fell in love with the solution instead of the problem. She spent a decade in VR when VR was going to be the future. It still isn’t, not because the technology didn’t work but because no consumer wants it.

“They need to be a researcher who understands the customer voice, who seeks it out, who understands product value propositions, how to do hypothesis-driven experimentation around the ideal customer profile,” she says. “Not just around the technology.”

This is the common failure mode in AI: founders build something technically elegant that solves zero customer problems. They’re enamored with the capability, not the use case.

FAQ

What’s the single most important thing you look for in an AI startup founder?

Urgency. Do they have a deep, personal, almost obsessive need to solve this problem? Founders who are urgently dissatisfied with the status quo move fast and stay focused when the problem gets hard. Generic ambition fails in chaos.

Do I need to have worked at a FAANG or top startup to get funded?

No, but it helps with pattern recognition. What matters more is: have you built something before? Have you faced constraints and solved problems under pressure? You don’t need a brand-name pedigree, but you need evidence you can execute.

How much does using GPT-4 vs. Gemini vs. Claude actually matter in your evaluation?

Almost not at all. If your entire competitive advantage is “we use the latest model,” you’re building on sand. The model will be commoditized in six months. What matters is: what are you doing with the model that couldn’t be done before? How are you building defensibility on top of the commodity?

How should I think about TAM in my pitch?

Don’t lead with TAM. Lead with: what behavior is changing right now? What tailwind is making this newly possible? If you can justify why your market is growing fast and changing, a smaller TAM becomes relevant. Uber’s TAM looked tiny but the market was expanding in real time.

What’s the difference between a good product and a magical product?

A good product works. A magical product makes you feel something on first use—either because it solves a frustration you’ve lived with, or because it reveals a possibility you didn’t know existed. If you have to explain why it’s useful, it’s not magical yet.

Should I pitch my business plan or a prototype?

Pitch a prototype. Business plans are theater. Mercedes spends one minute reviewing pitch decks and moves on if there’s no clear intuition that it matches what she’s looking for. If you have working software that shows the insight, lead with that.

How do VCs separate hype from real product-market fit?

Retention data. If early users are coming back week after week, you have something. If you’re buying every user and they ghost, you have hype. Mercedes waits to see cohort data that shows retention working before writing a check on consumer products.

What’s the most common mistake founders make when pitching to VCs?

They pitch the vision without the path. They say “we’re going to change the world” without explaining the concrete steps from today to year one to year five. Vision without a credible narrative of how you get there reads as naiveté, not conviction.

Is it true that warm intros are the only way in?

Mostly true. A cold email to a VC is a signal that you’re either not good at networking or not willing to put in the effort to find an intro. VCs interpret it as: “You couldn’t even network your way to an introduction—how are you going to network your way to customers?”

How should founders think about competing with OpenAI or Google or other giants?

They’re research labs first. They’re not going to prioritize consumer use cases or vertical depth. They’ll release general-purpose tools and move on to the next frontier. Build a vertical, a use case, or a business model they’ll ignore. That’s your moat.

When do you actually fund someone? After the first meeting or after diligence?

Mercedes makes gut intuition calls in the first meeting based on founder quality, but she confirms it with data. She wants to see: (1) founder exceptional in multiple dimensions, (2) product that shows early users returning, (3) market that’s provably growing, (4) vision that’s credible. If all four align, it’s a go.