The current batch of AI health tools is chasing the wrong metric. Walk through any funding announcement from the last six months and the pitch is almost always some variant of "we added GPT-4 to a clinical workflow." That framing treats the model as the product, which puts the feature list in the driver's seat and quietly shoves the patient into the back.
Teams that get traction in the next five years will do this in reverse. They will pick a behavior that actually moves a health outcome, and only then figure out which features (AI or not) support it.
A few behaviors that belong at the top of any health-product roadmap right now:
- Adherence without the guilt spiral. Patients know when they missed a dose. They do not need an app reminding them they are letting themselves down. The design question is how to reduce the silent shame that kills compliance in week three.
- Asking for help before the crisis. Most expensive health outcomes start as small, ignorable problems the patient did not want to bother anyone with. Tools that shorten the distance between "something feels off" and "I told someone who can help" are worth more than any triage chatbot.
- Retaining information between appointments. Clinicians know patients forget 80% of what they hear. AI that writes a plain-language recap a patient can send their spouse is worth more than another symptom checker.
None of these require a new foundation model. They require a team that watched real patients struggle and built around what they saw.
The feature-first teams will ship more polished demos this year. The behavior-first teams will ship products patients come back to next year. Those are not the same bet.