Every regulated firm we talk to opens with a version of the same question. How much autonomy is appropriate? It sounds like the responsible thing to ask and it is nearly impossible to answer, because autonomy is not a quantity. It is a property of specific actions, and the actions differ from each other far more than they resemble one scale.
The question that can be answered is: if this goes wrong, what does putting it right require? Sorting actions by that produces three groups, and the groups are not evenly sized.
- Reversible by the agent. A draft withdrawn, a ticket reopened, an allocation changed before close. Automate freely at depth 3.
- Reversible by a person, at a cost. An email already sent, a journal already posted, a case already routed. Automate with a ceiling and a legible trail.
- Not reversible at all. A settled payment. A submitted regulatory filing. A record in a clinical system another clinician has since acted on. Depth 4, and mostly withheld.
The third group is sector-specific, and that is the whole difficulty
Nobody can write a general rule for the third group, because irreversibility is a property of the rail, not the intent.
Payments. A push payment on a real-time rail is gone the instant it settles; you can request a return, which is a new payment somebody else has to agree to make. Curiously the case that needs the most care is the one where reversal is technically possible — inside a single institution both accounts are ours and a reversal is a database write. That it is possible is precisely why refusing it has to be a decision rather than a limitation. We wrote that one up separately.
Clinical. A record does not become irreversible when it is written; it becomes irreversible when another clinician reads it and acts. The window is invisible and it closes without any event the system sees. This is why our healthcare floor is not a depth cap but a categorical refusal: no depth ever signs a discharge summary, and no certificate raises that.
Insurance. A coverage decision communicated to a claimant is, in most jurisdictions, difficult to withdraw and in some effectively binding. The agent may assemble the file, cite the clauses and recommend — and a person communicates.
Telecom. Disconnection is reversible in the billing system and not reversible for the person who missed an emergency call. That distinction does not exist in the schema and has to be carried by the guardrail.
Reversibility is not a technical property you can read off the API. It is the answer to “who has to do what, and can they still do it” — which is why it belongs in the guardrail rather than the integration.
What took us three sectors to state properlyThe floor beneath the ceiling
Customers set a ceiling: the maximum depth they will permit, regardless of certification. Sectors have a floor underneath it: actions no ceiling can raise, because the refusal is not ours to trade away. The floor is uncomfortable commercially — it means telling a customer that a capability they are willing to pay for is not available at any depth — and it is the single most useful thing we have to say to a regulator.
What this buys the compliance function
Sorted this way, the autonomy conversation stops being a philosophical one and becomes a list. A list can be reviewed line by line, disagreed with specifically, and re-derived when a rail changes. “Appropriate autonomy” cannot be reviewed at all, which is why it survives in policy documents for years without anyone noticing it constrains nothing.
