Season 2 · Episode 5 · Design
Graduated autonomy: where the agent decides and where a person does
The question isn't whether to trust an agent, but how much, for what and with what safety net. Autonomy is designed in levels, like self-driving cars.
At one company, the pilot of a reconciliation agent ends in a debate that sounds philosophical. Finance wants the agent to post the obvious matches on its own, without bothering anyone. Audit wants nothing posted without review. IT doesn't know what to configure. Three weeks pass without a decision, until someone draws four ways of operating on the whiteboard: inform, propose, act after approval or act with notice. Forbidden actions sit outside the scale.
The scene is a composite, and the whiteboard scale is the central idea of this episode. An agent's autonomy isn't a single decision. It's a design.
The lesson from self-driving cars
SAE J3016 defines levels 0 through 5 for driving features, from no automation to automated driving under all conditions in which a person can drive. A level describes the active feature and the roles of the person and system; one vehicle can offer features at different levels. The useful analogy is defining scope and responsibility for each function, not applying the automotive standard to business AI.
The four levels below are a design guide for business agents, not SAE levels. One application can assign different controls to each task.
Four practical levels
- Informs: queries, analyzes and reports without modifying business records.
- Proposes: prepares an action and its evidence; a person decides and executes it in the system.
- Acts with prior approval: the agent applies the action only after explicit sign-off from the authorized role.
- Acts with notice: acts automatically within strict limits, records the result and notifies for later review. Only for low-impact, reversible cases.
Some actions remain forbidden to that agent, whatever its level on other tasks. The scale does not replace technical permissions or the ability to stop it.
The question isn't whether to trust the agent. It's how much, for which task and with what safety net.
How to assign the level
To assign a starting level, begin with the impact and reversibility of an error. An email draft is usually less serious than a payment to the wrong account; reclassifying a ticket is usually easier to undo than posting a journal entry in a closed period. Add data sensitivity, how quickly failures can be detected and the process's obligations. High-impact or hard-to-reverse tasks remain at “proposes” or “prior approval”.
Autonomy is earned with evidence
Assigning autonomy after a good demo is risky. Start with human review and expand scope when representative tests and difficult cases meet agreed quality and safety thresholds. A high unchanged-approval rate alone is not enough: examine rare failures, monitor in production and retain the ability to return to a lower level.
That takes two things that recur throughout this season: records of proposals with business impact and control decisions, plus measures of the outcomes. Without that, the autonomy discussion becomes philosophical again.
For Monday
- Take a candidate process and list every action an agent could take in it.
- Classify each action by the cost of an error and whether it's reversible.
- Assign each a starting level and define what evidence would be needed to raise it.
Sources and context
The scenes in these episodes are composites of common situations, not customer cases. Any figures in them are illustrative.