What You Should Refuse to Automate

The pitch in every clinic AI deck right now is the same: automate everything, remove the busywork, watch the hours come back. It makes a great slide. It also quietly skips the only question that matters, which is where automation should stop. Because there is a line, and the practices that get burned are the ones that automated straight past it before anyone had drawn it.

Knowing where that line goes is not caution, it is the actual skill. Anyone can point a model at a task. The judgment is in telling apart the work a machine should own, the work it should assist, and the work it must never be allowed to decide on its own.

The three axes that decide it

You do not need a framework with twelve boxes. You need three questions, and the honest answer to the worst of them.

First, reversibility: if this goes wrong, how easily is it undone? A mis-drafted recall message costs a phone call. A symptom wrongly categorised as routine is a patient who went home, and you may not learn what you missed for months. Second, judgment: is this a fixed rule, or does it need weighing, context, and values? Third, stakes: what is actually on the line if the output is wrong, a little time, or somebody’s health?

The rule for combining them is deliberately blunt. The worst single axis wins. A task can be routine on two axes and still belong to a person because of the third, and pretending otherwise is exactly how the expensive mistakes happen.

Grade a task and watch the line move

Here is the boundary as a tool rather than a lecture. Score a task, or load one of the examples, and watch where it lands.

Notice what happens with “decide a symptom is routine.” Every axis is at its worst: hard to undo, heavy judgment, somebody’s health at stake. That is the textbook keep-human decision, and it is the one a scribe vendor will eventually offer to help with. Now notice “draft a referral letter”: low stakes, easily undone, a clinician reads it before it goes. The verdict slides straight to copilot. Same technology, opposite answer, because the axes changed, not the model.

The line everyone gets wrong

The most expensive misreading of this whole idea is thinking “keep it human” means “keep AI out.” It does not.

A decision that stays with a clinician can still be wrapped in AI that does the heavy lifting around it: pulling the prior results, summarising a decade of chart, drafting a first note with its reasons, and flagging the thing a tired person at the end of a long list might miss. What stays human is the call itself, and the accountability that comes with it. This is the copilot pattern, and it is where most high-value work actually belongs: fast because the machine did the gathering, safe because a named person still signs it.

Automate the work. Keep the decision. The line between those two is the whole job.

So “keep-human” is not the slow option. It is how a practice goes fast without betting its reputation on a model’s worst day. The clinics that skip it, automate a clinical judgment, get burned, and then ban everything outright are the ones that actually move slowly, because one incident costs more time than a boundary ever would.

The fork

A task lands on your desk. Where does it belong?

Automate it

Reversible, no judgment, low stakes. A plain rule runs it. Adding AI just adds a new way to be wrong.

Copilot it

AI drafts or ranks, a person approves before it counts. Fast, and still accountable.

Keep it human

Hard to undo, heavy judgment, high stakes. AI assists, a person decides and owns it.

Where this rule misfires

Two honest caveats, because a rule you apply blindly becomes its own kind of mistake.

First, the axes tell you who owns the decision, not who is better at it. On a genuinely high-stakes call, the uncomfortable truth is sometimes that a measured model beats a tired human, and imaging is the standing example. Keep-human does not mean ignoring the better reader. It means the machine does the analysis and a named person ratifies and owns the outcome, with the authority to overrule. And whether AI should assist at all is a question you answer with evals on your own data, not with a hunch. The boundary decides accountability; measurement decides capability.

Second, this rule is a scalpel, not a shield. Its failure mode is timidity: labelling everything “keep-human” to avoid the work of automating, or standing up an approval step that is really a clinician clicking yes on twenty notes at nine at night without reading them. A rubber-stamp is not human judgment, it is latency with a signature, and it is worse than no AI because it launders the error. If you draw the boundary so conservatively that nothing moves, you have not been careful, you have just described the status quo with extra steps.

What to do Monday

Take your three most tempting automation ideas, the ones somebody keeps saying should just run themselves. Grade each on the three axes above, honestly, using the worst answer not the average. You will usually find one that is genuinely safe to automate, one that wants a copilot, and one where you were about to hand a machine a decision that belongs to a clinician.

Write the boundary down for each workflow before you build, not after the incident. It is a fifteen-minute exercise that settles a fight you would otherwise have loudly, later, in front of a patient or a regulator.

---

Not sure where the line falls in your practice? Book a free consult and we will grade your top automation candidates with you, and map which ones to automate, which to copilot, and which to keep firmly in clinical hands. See more on PHIPA and Guardrails.