The Honest Map of Clinical AI

Five Places AI Shows Up in Care

It helps to separate the field into five zones, because the evidence, the risk, and the regulatory position differ enormously between them. Administrative and documentation tools sit furthest from the patient and carry the clearest current benefit case. Imaging and diagnostics is the most studied clinical zone and the most mature in regulatory terms. Clinical decision support sits closest to the diagnosis or treatment decision and is the hardest to evaluate honestly. Operational triage and prioritisation reorders work rather than changing clinical conclusions. Patient-facing tools reach people directly, without a clinician between the output and the person acting on it. Almost every argument about "AI in healthcare" collapses once you ask which zone is being discussed.

• Administration and documentation — coding, scheduling, prior authorisation, note generation
• Imaging and diagnostics — detection, quantification, and worklist prioritisation
• Clinical decision support — risk scores, deterioration prediction, alerting
• Operations and triage, plus patient-facing tools such as symptom guidance and messaging
• The FDA publishes a public list of AI-enabled medical devices it has authorised; radiology accounts for the large majority of entries

Distance From the Decision Predicts the Evidence

There is a rough pattern worth holding on to: the further a tool sits from an individual clinical decision, the easier it is to justify and the faster it moves. Documentation support can be evaluated on time, clinician burden, and note quality, with a clinician reviewing and signing every output. A deterioration prediction model, by contrast, has to prove not just that it predicts something, but that acting on the prediction improves outcomes — a much higher bar that many systems have never cleared. This is why the current landscape looks lopsided. It is not that clinical AI is impossible; it is that the evidentiary and regulatory burden rises steeply as you approach the patient, and much of the market has not paid it.

• Far from the decision: measurable on workflow outcomes, with a human signing every output
• Close to the decision: must show that acting on the output changes patient outcomes, not just that it predicts well
• Many widely marketed clinical tools have never been evaluated against that higher bar

Back-Office Does Not Mean Harmless

The convenient story is that administrative AI is low risk. It is lower risk, which is not the same thing. A generated clinical note becomes the record; it is what the next clinician reads, what a coder bills from, and what a court examines later. A coding tool that systematically shifts documentation patterns has financial and regulatory exposure. A scheduling or prior-authorisation system that behaves differently for different patient groups produces access inequity even though it never made a clinical claim. Risk in healthcare AI follows consequence, not category label. The right question is never "is this clinical?" but "what happens downstream if this output is wrong, and who would notice?"

• A generated note is the record — downstream clinicians, coders, and reviewers rely on it
• Administrative systems can create access inequity without ever making a clinical claim
• Ask what happens downstream when the output is wrong, and who is positioned to notice
• Lower risk still requires review, audit, and an owner — it does not mean unmonitored

Try It Yourself

The five-zone map is only useful once you can place a real tool on it. This takes about fifteen minutes, a pathway you already know well, and no patient data at all.

Tool and the zone it sits in:
Pathway step where its output appears:
Who sees the output first:
Who is accountable for acting on it:
What happens downstream if it is wrong:
Who would notice, and how: