Autonomy is earned per action

The most useful operational idea in the framework comes from a case study: an IT automation company that scored every ticket type against three questions — how bad is a mistake, can it be undone, can a human realistically review each one — and let the score set the autonomy. Sixty percent of tickets became fully automatic with biweekly audits. Thirty percent became propose-and-approve. Ten percent the agent does not touch.

Notice what is not in that design: a decision about whether the agent as a whole is trustworthy. Trust is per action, not per agent. The same agent that resets passwords without asking should be structurally unable to delete an account, and both facts should come from the same risk table.

Human involvement has a matching ladder — the framework describes it as agent proposes and human operates, agent and human collaborate, agent operates and human approves at critical steps, agent operates and human observes. Tiering is how you choose a rung per action instead of arguing about the whole system at once. Larger organisations formalise this further; one pharma company in the framework runs five levels of agency, each with its own governance pathway, so low-risk agents move fast and high-risk ones meet real scrutiny.

The lab makes the tier gate physical. Six tickets run through Relay, the helpdesk agent this whole course instruments. With the gate on, tier-2 actions come back as queued for approval and tier-3 calls are refused with a reason. With the gate off — the demo build — whatever the model decides, happens. Run the offboarding ticket both ways and watch what changes.