IPM Take
This is where the AI governance debate becomes real. The technology did not merely rank images. It made autonomous discharge decisions inside a live cancer pathway. The capacity gain is impressive. So is the responsibility. Systems deploying autonomous AI need continuous surveillance, false-negative detection, clear accountability and a route back into care when the machine is wrong.
Executive Summary
Real-world findings presented at the 2026 EADV Congress evaluated autonomous AI within urgent suspected skin-cancer pathways across two U.K. hospitals. The system managed 8,391 patients, representing 94% of urgent referralsin the participating services. After validation, it autonomously discharged 31% of patients at one site and 25% at the other without clinician review. Researchers estimated that the pathway saved 2,851 clinician hours, equivalent to more than 8,500 additional face-to-face appointments. Biopsy rates fell from 43% in conventional face-to-face care to 27% in the autonomous pathway. Post-market surveillance identified six false-negative discharges—five basal cell carcinomas and one melanoma in situ—with no reported adverse outcomes during available follow-up.
Why it matters
- Patients / advocates: Autonomous discharge changes who—or what—makes a cancer-pathway decision.
- Clinicians: AI may release specialist capacity, but clinicians remain responsible for the system around it.
- Regulators: Static premarket validation is not enough for autonomous systems operating in real care.
- Hospitals / providers: Deployment requires surveillance, escalation protocols, audit and clear ownership of failure.
AI in oncology has spent years promising to “support clinicians.”
This system is doing something more consequential.
It is sending some patients home without a dermatologist reviewing the case.
That is why the NHS data deserve attention.
Across more than 8,000 urgent suspected skin-cancer referrals, autonomous AI removed a substantial share of apparently benign lesions from the specialist pathway. Researchers estimate the model released enough clinician time for more than 8,500 additional appointments over 16 months.
In a workforce-constrained system, that is enormous.
But the most important part of the study may be the six errors.
Post-market surveillance detected five basal cell carcinomas and one melanoma in situ among patients the system had discharged. No adverse outcomes were reported in available follow-up. That does not make the system unsafe. It makes the governance question visible.
Humans miss cancers too.
The difference is that autonomous AI can scale a decision across thousands of patients very quickly.
So safe deployment needs something stronger than an accuracy figure frozen at approval. It needs continuous monitoring. False-negative review. Version control. Performance by subgroup. Patient safety-netting. Clear responsibility when an algorithm, rather than a dermatologist, ends the referral.
This is what real AI regulation looks like.
Not debating whether AI will enter medicine.
Deciding what happens now that it already has.
