Three headlines from the last six months tell the same story. Microsoft launched Copilot Health, a consumer-facing assistant that reads lab results and suggests next steps. Amazon opened up its Health AI tool beyond the One Medical membership wall, pushing it at the broader AWS customer base. Google started pulling continuous glucose monitor data into its Fitbit coaching loop, so your ring, your watch, and your CGM can now agree about whether you ate too much rice at lunch.
The capital behind these launches is enormous. Bessemer's State of Health AI 2026 report puts venture funding into health AI at a multi-year high, with more than half of enterprise health systems now running at least one generative AI pilot. The models are real. The distribution is real. The press coverage has been nonstop.
And almost none of it is pointed at the actual bottleneck.
The gap nobody wants to talk about
AI in health is still solving the wrong problem. The question the industry keeps answering is "can a model generate a reasonable clinical plan?" The answer has been yes for about two years. Any decent LLM can produce a plausible differential, a reasonable medication list, a well-structured diet plan. That was the easy part, and we're done with it.
The hard part sits downstream. It's whether a 42-year-old nurse working 12-hour shifts, raising two kids, and carrying a mortgage will actually follow through on what the AI suggested. For 90 days. Without a care team nagging her. Without her motivation collapsing the third week in when work gets busy.
MIT Technology Review ran a piece in March 2026 asking a version of this question directly: there are more AI health tools than ever, but how well do they actually work? Their reporting cut through most of the launch-day narrative. The tools get evaluated on accuracy and clinical reasoning. They rarely get evaluated on whether the patient on the other end did the thing.
What the data actually says
The benchmark for digital health behavior change is grim, and it hasn't moved much in five years. A systematic analysis of mental health apps published in the Journal of Medical Internet Research found that roughly 4% of users who download an app are still using it 15 days later. 3% at 30 days. These aren't obscure products either. These are the well-funded, well-designed apps that got real user-research budgets.
A product with 100,000 downloads and 3,000 monthly actives is an expensive marketing funnel attached to something that doesn't stick, which is a different thing from a health intervention even if the pitch deck calls it one.
Now watch what happens when AI gets added to that same pattern. The app generates a better diet plan. The recommendations are more personalized. The onboarding is smoother because a language model is guiding it. Retention barely budges, because the problem was never that the plan was bad. The problem was that the plan didn't survive contact with the patient's Tuesday.
The question that doesn't get asked
What happens in the space between "AI suggests an action" and "person performs that action consistently for 90 days"?
That's the design problem. It's almost entirely behavioral, and it's almost entirely ignored in the way these tools are built. A CGM-integrated coach that tells you to eat more fiber is useful exactly once, on the day you read the message. Whether you still eat more fiber in week six depends on things the AI doesn't know and usually isn't asked to account for: your kitchen, your schedule, your partner's preferences, the grocery store nearest your office, whether you have the energy to cook after work or not. Clinical files don't capture any of that.
The teams that are making real progress here tend to look different from the ones getting funded. They talk about friction before they talk about features. They build for what actually happens on a Wednesday night at 9pm when the patient is exhausted, not for what the patient said they'd do during an onboarding flow at 10am on a Saturday. They measure whether behavior changed at day 45, not whether the user rated the onboarding five stars.
Some specific patterns worth noticing:
A medication-adherence product designed around the family member who reminds the patient, not around the patient's phone. An AI diabetes coach that adjusts its recommendations when the user's schedule changes, without asking the user to update their profile. A sleep product that assumes the user will fail for a week and designs the recovery arc into the core flow, rather than treating failure as a support ticket.
None of these require better models. They require teams that watched real people struggle and built around what they saw.
Where the real work is
Bessemer's report makes one point in passing that probably deserves a whole separate conversation: the companies showing durable retention metrics in their portfolio aren't the ones with the fanciest models. They're the ones with the clearest picture of what their user's life actually looks like. The model is a component. The behavior design is the product.
This doesn't mean AI doesn't matter in health. It matters a lot. But most of the current wave is pointing the technology at the wrong seam. You can't solve a behavior problem by being better at generation. You have to solve it by being better at what happens in the 167 hours a week the patient isn't using your app.
The industry is about to learn this the expensive way. A lot of well-funded AI health products are going to post impressive DAU numbers for the first six months and then watch retention curves collapse the same way they always have. The conclusion will be that the model wasn't good enough, and the next funding round will chase a better one. That's the wrong conclusion.
More AI isn't the answer. Better behavior design using the AI we already have is.
For a deeper look at the behavioral frameworks underneath this, behaviorux.com/insights has more.