Autonomy changes are risk decisions

Informal promotion

A workflow begins with human review. It performs well in demos and handles routine cases without an obvious incident. Over time, review becomes sporadic and then disappears. The system has gained authority, though nobody declared the change or recorded what justified it.

After a failure, the team may know what the system did but still cannot reconstruct why it was allowed to do it. Confidence scores and a handful of successful examples provide no agreed condition for restoring oversight.

Each change in oversight is a recorded decision tied to a task, supporting evidence, the authority granted, and a condition for taking it back.

Consider a support assistant that drafts replies. Removing approval for password-reset instructions and removing it for billing disputes are different risk decisions. Aggregate reply quality cannot justify both. Each task needs its own boundary, measurements, and fallback even when the same model performs the work.

Appropriate autonomy

Evidence from one task supports authority over that task. A system that reliably classifies routine refund requests has not thereby earned authority over exceptions, account changes, or payments. Scope keeps a local result from becoming a general claim about the system.

Some tasks should remain gated permanently. High-consequence actions may always require human approval. Other tasks may move toward monitoring after enough reliable operation.

Better output also does not automatically grant more authority. Worker quality and autonomy position are separate decisions. A strong worker can remain under review while it builds a history of results; an improved worker must qualify under the same governing rules.

Replacing a model, changing a prompt, or adding tools can invalidate evidence collected for the previous worker. A replacement can operate under tighter oversight while it qualifies.

Review produces useful evidence

Early oversight produces the labels and outcome data needed to build cheaper evaluators. Reviewed refunds become examples of routine approvals, required escalations, and policy failures. A classifier can eventually handle common cases while people review exceptions.

An evaluator measures performance. A separate governance decision changes authority. People remain available for uncertain cases and decisions that require permanent approval.

What adoption requires

A team needs an observable task boundary, consistent evaluation, retained evidence, and declared rules for both directions of change. It also needs an accountable owner who can approve less oversight and restore it immediately after a confirmed failure.

Teams without a reliable task boundary or review process should establish those before considering promotion.

Adoption can begin with one narrow, high-volume task. Its record supports a decision to advance, remain gated, narrow the scope, or improve the worker.