What Medical AI Is Actually Doing Today

A radiologist reviewing a scan is comparing it, in their head, against every similar case they have personally seen across a career — a few tens of thousands at most. A model trained on scans can be shown millions of labelled examples before it ever looks at a real patient's image. That gap in raw exposure, not some deeper form of understanding, is where most of medical AI's advantage actually comes from.

It is not reasoning about your case the way a doctor does. It is recognising that a pattern in front of it statistically resembles patterns it has seen before, at a scale that makes rare patterns easier to catch than they would be for any one person.

That gap has a limit. Name it early, because it explains a failure mode the rest of this part of the track keeps returning to. A model recognises a case because it resembles cases in its training data — it has no mechanism for reasoning about a case that resembles nothing it has seen. Feed it patients who look, in the data sense, like the patients it trained on, and the pattern-matching holds. Feed it a patient the training data barely represented, and the same model that looked brilliant on paper gets quietly worse, without ever announcing that it has.

Medical imaging is the clearest success story. Several imaging-assist tools already hold regulatory clearance for narrow, specific jobs — flagging a suspected stroke on a CT scan fast enough to change how quickly a patient reaches a specialist, or measuring the exact volume of a lesion on a follow-up scan so a radiologist is checking a number instead of redrawing an outline by hand. They run in real hospitals, a radiologist reviews every result before it reaches a chart, and they catch patterns a tired reviewer on their fortieth scan of the day might miss.

Triage and worklist prioritisation use a patient's existing record to flag who is likely to deteriorate soon. A well-known example reorders a hospital ward's sepsis watch list hours before a nurse working strictly in order of who was admitted first would otherwise have noticed. The nursing staff does not get bigger. The list they work from gets ranked instead of arbitrary.

Drug discovery uses models to narrow an impossibly large space of candidate molecules down to a shortlist for testing in a lab — the subject of the next chapter.

None of the above is where most of medical AI's actual deployed hours go. The highest-volume use, by a wide margin, is duller than any of it: turning a conversation into a chart, a chart into a bill, and a full clinic schedule into one with fewer empty slots.

1. 1

   The visit is recorded

   An ambient listening tool picks up the conversation in the room, with the patient's consent.

2. 2

   A draft note is generated

   The model turns the conversation into a structured clinical note in the format the electronic record expects.

3. 3

   A clinician edits and signs it

   Nothing reaches the permanent record without a clinician reading it first and making it theirs.

4. 4

   Codes are suggested

   A separate model maps the finished note to the billing codes it implies, for a human coder to check rather than submit blind.

A quieter, separate category does the same kind of pattern-matching on a hospital's calendar rather than a patient's chart — predicting which appointments are likely to go unfilled, or which slot actually fits a given procedure's typical length rather than the generic thirty minutes the booking system defaults to. Adopted faster than anything diagnostic, for an unglamorous reason: getting a documentation draft or a schedule slightly wrong costs an edit, not a delayed diagnosis, so hospitals took the risk sooner.

Fully autonomous diagnosis — a model that receives your symptoms and returns a confident, unsupervised verdict with no clinician in the loop — is barely deployed today. The rare exceptions are tightly scoped screening tasks: the best-known is an FDA-authorised system that checks retinal photos for diabetic eye disease without a specialist reading the image, and even its output is a decision about whether to refer the patient to an eye specialist, not a treatment plan. The next chapter covers why the boundary sits there.

Predicting rare, highly individual outcomes from limited data, the kind of headline that promises a model can tell you your exact personal risk of a one-in-a-million condition, usually overstates what a model trained on population-level patterns can actually say about one specific person.