Eighteen months ago, “AI-powered therapy” was a strength on a pitch deck — a signal of technical sophistication and scale potential. Today, in the rooms I’m in, it’s more often the line that gets the most follow-up questions.
What Changed Isn’t the Technology
The shift isn’t really about the models getting better or worse. It’s that investors are asking harder clinical-validation questions of AI mental health tools than they were a year ago — not because they’ve become skeptical of AI, but because the first wave of scrutiny (regulatory, journalistic, and clinical) has made clear what a well-built AI mental health product needs to be able to answer.
The Questions That Weren’t Being Asked Before
Is there a licensed clinician meaningfully in the loop, or is the AI operating without real clinical escalation paths? What’s the evidence base behind the model’s therapeutic claims, versus what’s inferred from general-purpose training data? What happens, specifically, when a user’s inputs suggest crisis-level risk — and has that pathway actually been tested, not just designed?
The startups that survive the harder questions are the ones that could already answer them before anyone asked.
The Startups That Are Surviving the Scrutiny
The pattern I’m seeing is straightforward: the startups that survive the harder questions are the ones that could already answer them before anyone asked — because clinical validation was built into the product from the start, not retrofitted once a diligence team or a regulator started asking. The ones caught flat-footed are almost always the ones treating clinical rigor as a later-stage problem, something to solve after product-market fit instead of alongside it.
Curious where you’re seeing this shift too — if you’re evaluating an AI mental health tool right now, I’m happy to talk through what the harder questions actually look like.