I’ve reviewed somewhere north of 400 AI startups over the past three years. Most of them weren’t good investments. But the ones that were good shared certain characteristics that, in retrospect, seem obvious.

Here’s the framework I’ve developed for evaluating AI startups. It’s not perfect, but it’s helped me avoid some expensive mistakes and identify some genuine opportunities.

Question 1: Where’s the Moat?

This is the most important question, and the one most AI startups can’t answer well.

Building with OpenAI’s API is easy. Training a basic model on your data is easy. That means if your entire value proposition is “we applied AI to X,” you don’t have a moat. Someone else can do exactly what you did, possibly cheaper.

The moats I look for:

Proprietary data is the strongest moat. If you have access to data that competitors can’t get, you can build better models that competitors can’t replicate. This is why companies like Recursion (massive biological datasets) and Tempus (clinical data from healthcare systems) are interesting.

Workflow integration is underrated as a moat. If your product becomes embedded in how customers work - their daily processes, their data flows, their decision-making - switching costs are high even if the underlying AI isn’t differentiated.

Network effects can work but are rare in AI. If your product gets better as more people use it, you have something. But most AI products don’t actually have this dynamic - they just have scale effects, which are different and weaker.

Speed of iteration matters in a rapidly evolving field. The team that ships fast, learns fast, and adapts to new model capabilities will outpace slower competitors even if they started behind.

Red flag: If the answer to “what’s your moat?” involves generic claims about “proprietary algorithms” or “unique AI capabilities” without specifics, that’s not a moat.

Question 2: What’s the Wedge?

How do they acquire customers? This matters more than most investors realize.

AI startups often have brilliant technology and terrible go-to-market. The product is cool, but the sales motion doesn’t work, the pricing doesn’t make sense, or the target market isn’t actually reachable.

Strong wedge strategies I’ve seen:

Replace a specific, expensive workflow. If you can show a CFO that your product replaces $500K of outsourced work with a $50K SaaS subscription, the value proposition is clear. The ROI conversation is straightforward.

Land with a free tool, expand with paid features. Some AI startups grow through viral free products that demonstrate value, then monetize through advanced features or enterprise versions. This can work but requires truly viral mechanics.

Partner with system integrators. B2B AI often gets sold through consultancies and SIs who are trusted advisors to enterprises. If you can make yourself part of their toolkit, you inherit their distribution.

Weak wedge strategies:

We’ll sell to IT. IT departments are overwhelmed and skeptical. Unless you’re solving a problem IT specifically owns, going through IT is slow and painful.

We’ll run pilots with innovation labs. Innovation labs don’t have budget authority and don’t control deployment. Pilots that don’t convert to production deployments are just free consulting.

Word of mouth. This isn’t a strategy, it’s a hope.

Question 3: Does the Unit Economics Work?

AI is expensive. Training models costs money. Running inference costs money. And those costs need to fit within a business model that works.

The questions I ask:

What’s the gross margin? Some AI products have thin margins because inference costs are high relative to what they can charge. This limits their ability to invest in growth and makes them vulnerable to price competition.

How does margin change at scale? Model efficiency improvements and volume discounts from cloud providers should improve margins over time. If the startup hasn’t thought through this trajectory, they’re not thinking about the economics seriously.

What’s the payback period on customer acquisition? If it takes two years to recoup the cost of acquiring a customer, the business needs a lot of capital to grow. That’s not fatal, but it’s a factor.

Red flag: Startups that handwave about economics with “we’ll figure it out at scale” or “the market is so big it doesn’t matter.” It always matters.

Question 4: Foundation Model Dependency Risk

This is specific to the current AI landscape: how dependent is the startup on foundation models they don’t control?

If your product is a thin wrapper around GPT-4, you have a problem. OpenAI could ship a feature that makes you redundant. Or raise prices. Or change their terms of service. You’ve built your house on someone else’s land.

Lower-risk positions:

Model-agnostic architecture. The product works with multiple underlying models and can switch as the landscape evolves.

Proprietary fine-tuning or training. The base model might be commodity, but the startup has trained specialized capabilities on top that aren’t easily replicated.

Value in the workflow, not the model. The model is one component of a larger product that includes data integration, user interface, business logic. The AI is important but not the entire value.

Higher-risk positions:

Competing with foundation model providers directly. If OpenAI, Google, or Anthropic could ship your product as a feature, you’re in trouble.

Complete dependence on a single provider. No alternative if that provider changes direction.

Question 5: Team Reality Check

This sounds generic, but the specific things I look for in AI teams:

Do they understand the problem domain, not just the AI? The best AI startups have deep expertise in the problem they’re solving, not just the technology they’re applying.

Have they shipped AI products before? There’s a huge gap between building AI in research contexts and building AI products that work reliably in production. Prior experience matters.

Are they realistic about timelines? Teams that promise autonomous AI solutions in six months either don’t understand the technology or are being deliberately optimistic. Neither is good.

Can they talk about limitations? If I ask about cases where their product doesn’t work well, and they don’t have good answers, they either haven’t deployed widely or aren’t being honest.

Putting It Together

When I evaluate an AI startup now, I’m looking for:

1. A defensible position that won’t evaporate as the technology commoditizes
2. A clear path to acquiring customers at reasonable cost
3. Economics that work at the unit level
4. Reasonable positioning relative to foundation model providers
5. A team that understands both the technology and the problem

Most AI startups fail one or more of these criteria. That doesn’t mean they won’t succeed - some will get lucky, or the market will evolve in unexpected ways. But for systematic investment decisions, these filters help separate the opportunities worth pursuing from the ones that are just riding the hype.

If you’re evaluating AI investments and want to compare notes on specific opportunities, my inbox is open. The framework keeps evolving as the market does.