Clinical versus administrative: two risk modes

AI in healthcare: an overview · Lesson 2 / 20

Why one yardstick does not fit both

Administrative and clinical applications differ not in model sophistication but in what happens after an error. In an administrative task a wrong output creates work: somebody rewrites the text, calls the patient back, fixes the schedule by hand. In a clinical task a wrong output can travel further down the chain and affect a person, and nobody will notice unless verification is deliberately built into the process.

Four features that separate the modes

Reversibility. Can the consequence be undone in minutes at negligible cost? A scheduling error is reversible; a decision about patient management rarely is.
Observability. Is it visible that the system was wrong, without a special check? A missed callback is noticed by the patient; a quiet error in prioritisation is noticed by nobody.
Time to detection. Does the problem surface immediately, or months later when the trail has gone cold?
Who pays. The organisation, in money and time, or the patient, in health.

How to use this in practice

Before any discussion of a tool, run the task through these four features and write the answers down. If even one answer points towards irreversibility or invisibility, the task belongs to the clinical mode, with a different approval path, a different depth of documentation and a mandatory answer to the question of who re-checks the output.

Risk mode card:
Task: ____
Is the consequence reversible? yes / no
Is the error visible without a special check? yes / no
Time to detection: minutes / days / months
Who bears the consequence: organisation / patient
Mode: administrative / clinical

Insight. The boundary runs between tasks, not between departments. The same patient chat assistant is administrative while it reports opening hours and becomes clinical the moment it starts advising whether to come in urgently.

Common mistake. Classifying a whole product instead of individual scenarios. The vendor sells a front desk assistant, and inside it there is a complaint triage scenario that carries an entirely different order of risk.

Pro tip. Record the risk mode in the same document that records the purchasing decision. Six months later nobody will remember on what grounds the tool was considered safe.

Cheat sheet

Risk mode follows from consequences, not from model complexity.
Four features: reversibility, observability, time to detection, who pays.
Classify scenarios, not whole products.
Write the mode down next to the purchasing decision.

1. What moves a task into the clinical risk mode?

2. Why is it wrong to classify a product as a whole?

3. What does the observability feature describe?

Task — checked by AI

Pick one AI tool that is used or under discussion in your organisation. Write out two or three concrete scenarios of its use and fill in a risk mode card for each: reversibility, observability, time to detection, who bears the consequence. State which scenario turned out to be the riskiest and why.

← Back

🔒 Answer the question correctly to move on to the next lesson.