Autonomy levels and the strictest-wins mask

2 min read

A single on-off switch for AI autonomy is not useful in a clinical setting. Different practices have different risk appetites, the same practice feels differently about different actions, and trust is earned per capability rather than all at once. The answer is a ladder, with several parties able to pull it down and nobody able to push it up past the ceiling.

#The ladder

Level	What the agent may do
L0	Nothing. Observation only.
L1	Read. Find work, look things up, answer questions. No writes anywhere.
L2	Prepare and ask. Assemble drafts, raise tasks for humans, write factual status notes where the practice has enabled it.
L3	Act on non-clinical things within bounds, with consequential actions still gated.
L4	The maximum. Still cannot release a payer transmission or a clinical record write.

L4 is not autonomy

The top of the ladder does not remove the human gate. If it did, the ladder would be a way of eventually turning the safety property off, and a safety property with a configuration path to disable it is a default rather than a guarantee. L4 means the agent does everything it can do without a human, which still excludes sending.

#Strictest wins

Several things can constrain what an agent may do at a given moment: the practice's configured level, a ceiling set above it, the specific agent's own limit, a temporary quarantine, and a fleet-wide switch. They are evaluated together and the most restrictive applies.

This direction is deliberate. Anyone can pull autonomy down; nobody can push it above the ceiling. An operator responding to a problem at three in the morning should not need to reason about whether their restriction will be overridden by a setting somewhere else.

#Unavailable, not punished

When a capability is outside the current autonomy level, it is masked: the agent is not offered it. It does not attempt it and get refused.

This matters more than it sounds. An agent that repeatedly attempts a forbidden action and is rejected learns nothing useful, produces noise in the audit trail, and in a system with retries can spend a budget discovering the same wall. Making the capability absent is cleaner than making it forbidden.

#Earning a level

Movement up the ladder should be evidence-based rather than a matter of comfort. The relevant evidence is the eval scorecard for the capabilities the new level would unlock, on the slices that matter to that practice, with the number of cases stated. See Evaluating a clinical agent.

A practice asking to move from L2 to L3 is asking a specific question: has this system demonstrated it can be trusted with these particular actions, for patients like mine. That question has an answer, and it should be produced rather than asserted.