— What we hear
"We have a working model and nobody is using it."
The most common failure mode in clinical AI. The model is fine. The interaction shape is wrong. Clinicians don't adopt tools that feel like they're being graded by software.
"We're locked between FDA pathway requirements and the EU AI Act and we don't know which to optimise for."
Both. Both pathways. The good news is the overlap is real — the bad news is each requires distinct documentation, and most teams pick one and discover the other six months too late.
"The model degrades on our deployment data and we can't tell when until something bad happens."
Clinical data shifts — populations, equipment, protocols, coding. Models that don't have drift monitoring with clinical context will silently fail in ways the post-market surveillance frameworks aren't designed to catch in time.
"Our pilot site loves it. Site #2 hates it. We can't tell what's different."
Clinical workflow is local. The same AI feature lands in two hospitals very differently depending on EHR setup, attending rotation patterns, and unwritten clinical culture. Generalisation is a research problem; rollout is a field problem.
— Pattern recognition
— What fails
"Defer to AI" interaction patterns. Senior clinicians don't defer — they verify. Designing around that is a UI decision, not a model decision.
Optimising only on model metrics. Sensitivity and specificity matter, but adoption rate is the metric that determines clinical impact. A 95% accurate model used 8% of the time is worth less than an 87% model used 89% of the time.
Single-site pilots without rollout in mind. The pilot site is over-engineered for success. The real test is site #4, where the EHR is older and the attending is sceptical.
Treating clinicians as users instead of co-designers. The reason most clinical AI fails adoption is that it's built without the clinical workflow at the centre. They know the workflow; the team rarely does.
— What works
Authority-preserving UI. AI orders the queue, suggests context, surfaces patterns — but the clinician decides. The framing keeps trust where the medical liability already is.
Structured disagreement loops. When a clinician overrides an AI suggestion, ask one question: why. Five preset options plus free text. This is your training feedback and your trust-building mechanism in one.
Clinical co-design from day one. Shadow sessions with attending physicians during Diagnose. Not focus groups — actual workflow observation. What surfaces is usually what fixes the engagement.
Drift monitoring with clinical context. Statistical drift is the wrong signal in healthcare. Population-level drift matters; equipment swap detection matters; coding pattern shift matters. Build the monitoring to clinical realities, not ML defaults.
— Engagement patterns
— 01
Triage, prioritisation, clinical alerts. Built with clinician shadowing, authority-preserving UI, and disagreement loops as feedback infrastructure.
— 02
Ambient scribing, note generation, summarisation. Faster than the EHR, accurate enough for the chart, structured enough for billing.
— 03
Charge capture, coding accuracy, denial prediction, appeals automation. Quiet, measurable, fast payback. Often the warm-up engagement that builds confidence for the clinical work.
— 04
FDA 510(k) / De Novo strategy, post-market surveillance design, EU AI Act compliance documentation. The boring work that determines whether you ship.
— Series B teleradiology (illustrative)
A Series B teleradiology platform had spent 18 months building a strong triage model. Adoption sat at 8%. We didn't change the model — we changed how it asked. Time-to-triage on priority scans dropped from 37 minutes to 4. The model accuracy was the same throughout.
— Composite case drawn from engagement patterns; anonymized for client confidentiality. Real metrics, real outcome arcs.
— If this resonated
If you're a HealthTech leader sitting with one of the problems on this page, applications are open for next quarter. We read every one.