Healthcare-adjacent work: clerical or clinical

"Healthcare-adjacent" is most of a health system: the practice manager, the medical secretary, the physiotherapist writing reports, the study coordinator, the app developer whose product is "wellness" until a regulator disagrees. This page is for that ring of work — close enough to touch patient data and clinical consequences, far enough that no one handed you a governance process.

The reference document is WHO's guidance on large multi-modal models, and its most useful property is that it sorts uses instead of blessing or banning the technology.

The line the guidance draws

The clerical side is where WHO sees the immediate, defensible win. On drafting responses to patient messages, the guidance describes the purpose as to "decrease the burn-out of health-care workers, who field thousands of messages daily, and to enable them to focus on their clinical duties" — it even has a name for it, "keyboard liberation". Letters, summaries, scheduling logic, report drafting: this is ordinary review-everything office use, and the adjacent worker is exactly who does it.

The clinical side has a bright line with a WHO sentence under it. Governments should ensure that "LMMs or applications for supporting clinical decisions that are not yet approved for use not be used on an experimental basis outside an authorized clinical trial setting". A chat window is not an approved clinical decision-support tool. Drafting the letter about a decision is clerical; anything that shapes the decision is clinical, and it enters through your institution's front door or not at all.

The test for which side you are on is consequence-shaped: if the output being wrong could change what happens to a patient's body, it is clinical, whatever the task looked like when it started.

Patient data is the second wall

The clerical win assumes the data is handled lawfully, and health data is the strictest category in every regime this site covers. Before a patient's name, condition or message goes into any tool, the what-data-may-go-in questions apply at maximum strength — legal basis, processor status, retention and training settings for Claude, ChatGPT or Gemini — and "I removed the name" is weaker than it feels: a rare condition plus an age plus a town is an identity. Describe patterns, use genuinely synthetic examples, or use a tool your organisation has actually contracted for health data. A personal account is never that tool.

The professional risks WHO names

The guidance is specific about what goes wrong in the professional, not just in the tool. It warns that "LMMs are likely to encourage automation bias in experts and health-care professionals", with the failure spelled out: "In automation bias, a clinician may overlook errors that should have been spotted by a human" — the same mechanism as the general case, at clinical stakes. And it extends the deskilling problem further than most documents dare, warning that delegating moral judgements could leave "physicians become unable to make difficult judgements or decisions", and that accountability gets murky once professionals "have outsourced certain responsibilities to AI".

For the adjacent worker the translation is: the review habit is the job. The model drafts; the human who signs is the one the duty attaches to.

The patient who arrives pre-informed

Chatbots have replaced search as the waiting-room second opinion, and WHO's warning about it is carefully double-edged: "Even if the information is correct, individuals without medical training who use such information for self-diagnosis could misinterpret or misuse it" — on top of "the propensity of chatbots to produce incorrect or wholly false responses", invented references included. Front-desk and triage roles now field both kinds. The useful stance is the one this site applies to its own domain: not "the tool is wrong", but "let's check that against a source that carries accountability".

What goes wrong

The patient in the personal chat. One pasted message thread: health data processed with no legal basis, retention unknown, on an account the organisation cannot see or purge.

Clerical scope creeping clinical. The tool that drafted appointment letters starts prioritising the appointment list. Nobody approved a triage system; one is running.

The wellness app in denial. Software that interprets symptoms or guides treatment is a medical device in most jurisdictions regardless of its marketing category, and "powered by AI" changes nothing about that.

The summarised guideline. A model's summary of a clinical guideline is a paraphrase without accountability; the guideline itself is the document that was governed.

Trust transferred instead of earned. The confident tone that misleads patients works on staff too.

How to check it worked

Take one week of your own AI-touched tasks and sort each into clerical or clinical using the consequence test. For everything clerical, confirm the data path: whose account, what retention, is the organisation's contract under it. Everything that lands clinical — or will not sort cleanly — goes to whoever owns clinical governance, as a question rather than a confession. The sorting itself is the safety mechanism, and a week's sample is enough to show whether it is happening.