TL;DR: Health AI is where "move fast" goes to die and where durable products get built. The PM skill is navigating three gates that don't exist elsewhere: clinical evidence, regulation, and workflow adoption. Get them right and you have a moat measured in years.
| Use case | Status (2026) | Why it works |
|---|---|---|
| Radiology/imaging triage & detection | The workhorse: ~76% of FDA-authorized AI devices are radiology | Clear ground truth, digital data, assists (not replaces) the read |
| Ambient clinical documentation ("AI scribes": Abridge, Nuance DAX, etc.) | The most visible current land-grab in health AI | Attacks clinician burnout; human reviews every note → survivable errors |
| Administrative ops: coding, prior auth, scheduling, claims | Quietly huge | High volume, pattern-heavy, non-clinical error tolerance |
| Patient communication: portal message drafts, triage chat | Growing, carefully scoped | Drafts reviewed by staff; escalation paths mandatory |
| Drug discovery & clinical trial matching | Deep-tech track | Different PM species (research PM) |
| Diagnosis-adjacent decision support | The hard, regulated frontier | This is SaMD territory: see below |
The cautionary anchor: IBM Watson for Oncology. >$4B spent on acquisitions, trained on small volumes of synthetic cases curated at one hospital, marketed as revolutionizing cancer care, internally documented recommending "unsafe and incorrect" treatments (STAT, 2018), sold off for ~$1B in 2022. Every health-AI failure mode in one story: marketing ahead of evidence, no real-world ground truth, workflow ignored. (Full case)
flowchart LR
A[Health AI product] --> B["🔬 GATE 1: Clinical evidence<br/>does it work, proven how?<br/>retrospective → prospective → RCT ladder"]
B --> C["📋 GATE 2: Regulation<br/>FDA/CE pathway if SaMD ·<br/>HIPAA/GDPR data compliance"]
C --> D["🏥 GATE 3: Workflow adoption<br/>fits the 15-second clinical moment ·<br/>EHR integration · who's liable · who pays"]
Health buyers ask "where's your study?" not "where's your demo." PM owns the evidence roadmap: retrospective validation → prospective pilot → (for strong claims) randomized trial + peer-reviewed publication. Evals here are clinical validation studies: same discipline as evals.md, higher ceremony, with site-level generalization as the recurring killer (models that work at Hospital A degrade at Hospital B, because scanners, populations, and workflows differ).
- SaMD (Software as a Medical Device): if your software diagnoses, treats, or drives clinical decisions, you're likely a medical device needing FDA authorization (mostly 510(k)/De Novo). FDA has authorized 1,000+ AI-enabled devices (cumulative ~1,451 through 2025).
- The line that matters: "flags studies for radiologist priority" (device, cleared) vs. "generic admin automation" (not a device). Feature wording and claims move you across it. Those are PM decisions, made with counsel.
- Adaptive models: FDA's Predetermined Change Control Plans (PCCP) (2025 guidance) let you pre-specify how a model may be updated post-clearance; otherwise every retrain is a new submission. Ask about this in any med-AI interview; it signals real domain knowledge.
- LLM caveat: most LLM chat tools currently sit outside cleared-device pathways by scoping claims to admin/documentation. The gray zone is the industry's active debate.
- HIPAA: PHI through an LLM requires a BAA. Consumer ChatGPT is not HIPAA-compliant; enterprise/cloud offerings (Azure OpenAI, Bedrock, Google Cloud, ChatGPT Enterprise) provide BAAs. BAA is necessary, not sufficient: minimum-necessary data, audit trails, retention rules still apply.
- EU: medical AI is high-risk under the AI Act (obligations timeline shifted to Dec 2027 by the 2026 omnibus) plus the existing MDR regime.
The clinician has ~15 seconds and 40 clicks of accumulated EHR resentment. Integration = Epic/Cerner (EHR) embedding, not a separate portal. Alert fatigue is the enemy: a 94%-precision alert still fires wrongly dozens of times/day at volume, so clinicians learn to ignore it (automation complacency), or it erodes trust in everything you ship. And map the buyer ≠ user ≠ beneficiary triangle: hospital CFO buys, clinician uses, patient benefits; your value story must work for all three (and often: who bills for it? Reimbursement codes can make or kill the category).
- Assist, don't decide: outputs framed as draft/flag/prioritization with clinician sign-off; changes liability, regulation, and adoption simultaneously.
- Show the evidence trail: link to the guideline/source/image region behind every suggestion (grounding = clinical trust).
- Fail visible, fail safe: low confidence → route to human, never to a guess; measure escalation correctness.
- Bias is a patient-safety issue: performance sliced by demographics is mandatory (pulse-oximeter history looms large); underrepresented groups in training data → measured, disclosed, mitigated gaps.
- Update discipline: versioned models, monitored drift (patient populations shift, coding practices change), PCCP-governed updates.
"Design an AI feature to reduce ER wait times." Strong path: decompose waits (triage → bed → doc → tests → dispo) → pick a non-diagnostic wedge first (predictive bed management, documentation acceleration) vs. diagnosis-adjacent (triage severity scoring = SaMD + evidence program; say so explicitly) → data reality (EHR access, HL7/FHIR, PHI under BAA) → clinician-in-loop design + alert-fatigue budget → metrics: door-to-provider time, LWBS (left-without-being-seen) rate, guardrails on triage-error/adverse events → rollout: shadow mode against current triage, single site, prospective validation before scale. Mentioning site generalization and who is liable when it's wrong separates domain-fluent candidates.
Epic/Cerner-Oracle (EHR gravity wells) · Abridge, Nuance DAX (ambient scribes) · Aidoc, Viz.ai (radiology AI) · Tempus (precision medicine) · FDA AI device list · PCCP · FHIR (data interop standard) · pulse-oximeter bias case · Watson Health post-mortem.
Regulatory status verified July 2026. This domain moves; re-verify before interviews. Related: responsible-ai · failures