My friend Meera is a radiologist, twenty years in, very good at her job — good enough that for a long time she eyed every AI tool her hospital tried to hand her with open suspicion.
"It's not that I don't think they work," she told me. "It's that most of them are trying to replace me, when what I actually need is a second pair of eyes on the last scan of a long shift."
That line has stuck with me ever since, because it's the sharpest description I've heard of where AI belongs in a hospital. Not as a replacement for the clinician. As the colleague who's still sharp at 11pm when you're running on fumes.
Mostly, it's plumbing
Most of the useful AI in healthcare today is unglamorous. It's not making diagnoses. It's flagging a spot on an X-ray that sits in an awkward place behind the heart shadow. It's noticing that a patient's overnight vitals are drifting in a way that, on their own, would be easy to miss. It's scheduling appointments in an order that means fewer people wait three weeks to be told they're fine.
I visited a community clinic in rural Georgia last spring that had one of these systems installed. A nurse walked me through how they used it to screen for diabetic retinopathy — a preventable cause of blindness that usually gets caught too late in underserved communities, because there aren't enough specialist ophthalmologists to go round.
The system takes a photograph of the back of the eye, runs it through a trained model, and flags anyone who needs to see a specialist. In six months, it had picked up over 200 cases of early-stage disease that would otherwise have been found only after vision loss started.
There's no drama in that story. Nobody's being disrupted. But two hundred people are going to keep their sight.
Where I'd be careful
The things that worry me here are the same things that worry the clinicians I trust, so I'll name them plainly.
Bias in the training data. If a diagnostic model is trained mostly on white, male, American patients, it will work worse for everyone else — invisibly, which is the dangerous part. This isn't hypothetical. It has happened, repeatedly, and the fix is slow, careful attention to what data we collect and from whom.
Automation without accountability. The worst version of AI in healthcare is the one where a clinician signs off on a machine's recommendation without really engaging with it, because the caseload is punishing and the tool is convenient. Every good deployment I've seen treats the AI as a suggestion, never a verdict.
Privacy, properly done. Medical data is some of the most sensitive information about a person that exists. If you're looking at AI in a clinical setting, the privacy conversation isn't a checkbox — it's the starting point.
Three rules we've earned the hard way
When we help hospitals and clinics think about AI, the same advice keeps proving itself:
- Start with administration, not diagnosis. Appointment scheduling, prior authorisation, documentation. These aren't glamorous, but they're where the fastest, least-risky wins live, and the team will thank you for them.
- Pilot in parallel, not in place of. Keep the current process running while the AI runs alongside it. Compare notes for a few months. Then decide.
- Let clinicians shape the tool. The hospitals that love their AI are the ones where a senior nurse spent three afternoons with the product team, arguing about the alert UI. That argument is the whole secret.
Back to Meera
She did, in the end, start using one of the tools. It catches a couple of things a year she reckons she'd have missed — not because she's bad at her job, but because nobody is perfect at 11pm after twelve hours of scans. And she still catches everything it does, every time. That's the whole arrangement. Neither one replacing the other; both watching the other's back.
That's the version of AI in healthcare I'd argue for. Not the demo that gets the press. The one where clinicians are better rested, fewer things slip through, and nobody loses their sight because the scheduling was bad. If you're a clinician wondering where AI might genuinely help, or a health-tech team trying to ship something that earns a place in a hospital, come and talk to us before you build it. That conversation is cheaper than the one you'll have after.
