AI in Healthcare: Documentation, Triage, Diagnosis
5 min read · updated August 3, 2026
Sort healthcare applications by capability and you get a list. Sort them by how close the output sits to a decision about a patient and you get a plan, because that distance is what determines who has to sign, what evidence they need, and how much oversight the thing attracts. Nothing here is medical or legal advice, and no claim is made about what any regulator currently requires.
The ladder, bottom rung first
| Rung | Description |
|---|---|
| admin | Prior-authorisation letters, appointment logistics, coding support, denial appeals. The output is a document sent to another organisation. Nobody is treating a patient on the strength of it, and errors surface as rework. |
| documentation | Drafting the encounter note from the conversation. The clinician reads, edits and attests. The output enters the medical record, so an error persists — but a human sits between the model and the record by construction. |
| triage | Answering a patient directly, or advising on urgency. No clinician is in the loop at the moment of use, which changes the problem entirely: the safety property is about the worst case, not the average one. |
| decision-support | Suggesting a diagnosis, flagging a finding, proposing a treatment. Software intended to inform a clinical decision sits in a different world of evidence, validation and post-market obligation — one that varies by jurisdiction and by exactly what the software claims to do. |
Most organisations should start at the top of that table and stay there for a while, and the reason is not timidity. It is that the bottom two rungs require a body of clinical evidence and a regulatory strategy as the first deliverable, and an engineering team that has not yet learned what clinical text does to a model is not ready to produce either.
Why clinical text breaks summarisers
Clinical language has three properties that general-purpose summarisation evaluation does not test for, and each has produced exactly the failure you would predict.
- Negation carries the meaning. A large share of a note is a record of what was not found. “No evidence of pneumothorax”, “denies chest pain”, “ruled out”. A summariser optimising for brevity drops the negation and inverts the finding. Any evaluation set for clinical summarisation needs negated findings deliberately oversampled, because they are the failure and they are not rare.
- Laterality and dose are not details. Left versus right, and a decimal point in a dose, are the two transcription errors with a direct path to harm. These deserve dedicated checks — extracted doses compared against the medication’s plausible range, laterality echoed back for confirmation — rather than being folded into a general quality score.
- Abbreviations collide by department. “MS” is mitral stenosis in cardiology, multiple sclerosis in neurology and morphine sulfate on a drug chart. “PT” is physical therapy, prothrombin time or the patient. The disambiguation is local, sometimes to a single unit, which means a model that performs well on one department’s notes has not been shown to perform on another’s. Evaluate per specialty.
Ambient documentation, and the number that decides it
Note drafting from an ambient recording of the encounter is the application with the clearest value proposition, because documentation time is the thing clinicians most want back. It is also the one where the success metric is most often chosen wrongly.
The metric is not similarity to a reference note. It is the edit burden before attestation — how much the clinician changes, and how long they spend, compared with writing it themselves. A draft that is ninety per cent right but requires reading every line to find the ten per cent can easily cost more time than dictation, and clinicians discover this in week two and stop using it. Alongside edit burden, one error class needs its own count with a target of zero: assertions of things that did not happen. A generated line recording a normal examination of a system that was never examined is a patient-safety problem and a billing problem simultaneously, and it is exactly the kind of plausible filler a language model produces when the transcript is thin. The general treatment is in grounding and measuring hallucination; the healthcare-specific part is that the acceptable rate is not a tuning parameter.
Triage: the asymmetry runs one way
A symptom checker or nurse-line assistant has an error cost that is radically asymmetric. Telling someone with an urgent presentation to wait until morning is unlike any other error in this cluster. Everything about the design follows from that:
- Red-flag detection runs as a separate deterministic layer over the transcript, not as part of the generative answer, so that a model failure cannot suppress it.
- The system is permitted to escalate but never to stand someone down. “This may need urgent attention” is an output; “this is not serious” is not one.
- The evaluation set is built from adjudicated cases and reports sensitivity on the highest-acuity band, at whatever specificity that costs. A version that improves overall agreement while losing one point of high-acuity sensitivity is a worse system.
The data does not move easily, on purpose
Clinical data arrives as HL7 v2 messages with pipe-delimited segments, or FHIR resources, or a CDA document, or — very often — a PDF that was faxed. Health information is handled under access, disclosure and retention duties that differ by country and by care setting, and the practical effect on architecture is consistent everywhere: identified data does not casually leave the environment it lives in, and every processor of it needs a contract before the pilot, not after.
What that means in practice is that the deployment question comes before the model question. De-identification before egress, a processing environment with the right contractual and residency properties, or self-hosted inference are the three shapes, and they have very different costs — see health data and language models, confidential inference and self-hosting for a team. De-identification is itself imperfect on free text: a note that mentions a rare condition, an occupation and a town can identify a person with every explicit identifier stripped, which is why de-identification is a risk reduction rather than a release from the obligations.