How to evaluate AI startups in 2026: the signals that actually matter now

The standard due diligence playbook was already straining under the pace of AI in 2024. By now it's genuinely broken. Frameworks built around team, market size, and technology differentiation don't map cleanly onto a landscape where the technology is commoditising in real time and the moat lives somewhere else entirely. Here's what we look at instead.

Start with the data question, not the model question

The first thing to establish is not what model the company uses, but what data they have that nobody else can access. Proprietary data is the only durable technical advantage in the application layer right now. Everything else, the prompts, the UX, the orchestration, can be replicated.

The data question has two parts. First: do they have it? Second: does it actually improve the product? A lot of companies claim data advantages that turn out to be historical datasets with no feedback loop. What matters is whether the product gets meaningfully better as more customers use it, generating signals that compound over time.

The signals worth paying attention to

1. Retention shape, not just retention rate

A 70% 90-day retention rate means different things depending on the shape of the curve. If retention drops sharply in the first two weeks and then flattens, that's a product that people commit to once they get past the initial friction. If it slopes steadily downward, that's a product people are slowly giving up on. Ask to see the full curve, not the headline number.

2. Who is churning

Churn analysis by segment tells you more than aggregate churn. If a company is losing small customers but retaining enterprise accounts, that may be intentional portfolio shaping. If they're losing power users, that's a product signal worth interrogating. Ask them to walk you through the last ten churned accounts and why they left.

3. The model dependency test

Ask a direct question: if your primary model provider increased prices by 40% tomorrow, what would you do? The answer tells you how embedded the model is in their unit economics and how much pricing power they actually have. Companies that have already stress-tested this scenario are meaningfully more sophisticated than those that haven't thought about it.

4. Evaluation infrastructure

Ask how they know when a model update makes their product better or worse. Teams with serious evaluation infrastructure, even informal ones, have a fundamentally different relationship with their product quality. Teams without it are flying blind and don't know it. This is one of the most reliable signals of technical maturity in an AI company right now.

The competitive question has changed

Two years ago, the competitive question was mostly about other startups. Now it's primarily about the foundation model providers themselves. OpenAI, Anthropic, and Google are all moving up the stack, shipping features that were previously the territory of application-layer companies.

The question to ask isn't "who are your competitors?" It's "which of your features could a foundation model provider ship in the next twelve months, and what happens to your product if they do?" Companies that have a clear answer to this, and a credible plan for why they'd still be valuable, are worth the conversation. Companies that haven't thought about it are not.

What good looks like in 2026

The AI companies worth backing right now share a few traits: they have a specific customer with a specific workflow problem that recurs frequently; they have some form of data advantage that compounds; and they've thought seriously about what happens when the models underneath them improve. The ones that check all three boxes are rare. The ones that check two are worth a closer look.

The bar for "AI company" is now so low it's nearly meaningless. The bar for "AI company that will matter in three years" is higher than it's ever been.