How to read AI claims in medicine

What a 90% match really means

A 90% diagnostic match is not quite what it sounds like.

The AMIE study from Google and Beth Israel Deaconess, published in The Lancet last week, deserves a proper read. Ninety-eight patients at a single Boston primary care clinic chatted with the AI remotely before an urgent care visit.

The 90% means the final diagnosis appeared somewhere in AMIE's top seven suggestions. Its first choice matched the final diagnosis in 56% of cases, according to the preprint. Blinded internists found no significant difference from the doctors in differential quality, or in how appropriate and safe the management plans were. They rated the doctors' plans better on practicality and cost-effectiveness. A doctor watched every conversation live, and none needed stopping.

None of that makes it a weak study. It's prospective and used real patients, which is still rare in this field.

The finding I'd underline is the workflow. The AI took the history before the patient walked in. Where the doctor had reviewed the transcript beforehand, they said it helped them prepare in three out of four surveyed visits.

Most diagnoses come from the history. It's one of the first things medicine teaches you, and nearly two decades of practice haven't changed my mind. A good history waiting for me changes my first five minutes. It doesn't replace the rest of the consultation.

For Gulf and Nordic health systems and investors hearing "AI doctor" pitches: are you being sold the diagnosis or the history? One needs outcome trials. The other needs workflow and safety data from your own clinics.

What happens when it is wrong?

At HIMSS 2026 last week, Oracle, Google, Microsoft, Amazon, and Epic all unveiled AI agents designed to work autonomously in clinical workflows.

The technology is genuinely impressive. I've seen some of what's coming.

The clinicians in the room were asking the question that matters. The vendors on the main stage weren't answering it: what happens when it's wrong?

The validation frameworks have not kept pace with deployment timelines. Organisations are deploying first and validating later, relying on vendor benchmarks rather than independent clinical evaluation. The EU AI Act's high-risk classification for healthcare AI takes full effect in August 2026. The gap between those two timelines is where the liability lives.

I'll say it plainly. A probabilistic system that occasionally hallucinates a drug interaction isn't producing a bad recommendation. It's a patient safety event. And in healthcare, patient safety events have consequences that no product roadmap accounts for.

The Stanford-Harvard ARISE network put this clearly in their State of Clinical AI report in January: real-world performance after deployment looks very different from controlled benchmark results. That finding came from reviewing the most significant clinical AI studies of 2025. It wasn't a theoretical concern. It was the evidence.

Technology readiness and clinical readiness are not the same thing. They never have been.

Health system leaders procuring AI right now are the ones who need to hold this line. The vendors won't.

Front door and engine room

The Great Bifurcation: Front Door vs Engine Room

This week marked a clear shift in how AI is entering medicine. The experimentation phase is ending. A more defined structure is emerging.

On 7 January, OpenAI launched ChatGPT Health, aimed at patients and consumers.

On 11 January, Anthropic advanced Claude for life sciences, focused on researchers and clinicians.

As a physician with a strategic lens, I don’t see routine product launches. I see a decisive split in the healthcare value chain.

The Front Door

OpenAI is competing for agency and triage. The goal isn’t to replace clinicians but to meet patients where they already are, seeking answers, context, and reassurance before they ever reach a clinic.

The real challenge is no longer raw accuracy. It’s the anxiety gap: the distance between what a patient reads and what a clinician later validates. Widen that gap and trust erodes. Narrow it and access improves.

The Engine Room

Anthropic is playing a different game. Speed and precision inside the system. Parsing complex clinical data, accelerating research, reducing friction between discovery and delivery.

This isn’t conversational medicine. It’s infrastructure, the tools that quietly reshape how trials are designed, how evidence is synthesised, and how clinicians make decisions under pressure.

The Strategic Question

Medical knowledge is becoming commoditised. The scarce asset is no longer information. It’s judgement and execution.

For leaders in boardrooms and on hospital floors, the question isn’t whether to adopt AI. It’s where you position yourself in this emerging ecosystem.

Are you empowering the patient? Or optimising the provider?

These aren’t mutually exclusive, but they demand different architectures, different metrics, and different risk tolerances.

Trying to do both with the same tool usually delivers neither well.

Technology is the lever.

Clinical judgement remains the fulcrum. Build faster systems by all means, but make sure they’re pointing in the right direction.

What it means

For Gulf and Nordic buyers the practical rule is the same in each case. Ask what was measured, in which patients and in which workflow, and check that independent clinical evaluation exists before deployment rather than after it.