An autonomy decision matrix sets, in advance, the highest depth an AI agent may ever reach for each kind of action, based on two things: how reversible the action is and how much harm a mistake would cause. Fill in the grid once, apply it to every action, and use it to settle requests for more autonomy.
What does the matrix decide, and what does it not?
The matrix sets ceilings. It says that a reversible, low-impact action may reach depth 3, while an irreversible, high-impact one may never pass depth 2. It does not say any particular agent has reached that ceiling. That still needs evidence from certification on your own cases. Depth is earned; the matrix only says how far it can be earned.
Without a matrix, each autonomy change request gets argued from scratch, and decisions drift with whoever is in the room. With one, the request becomes a narrower question: has the evidence reached the level the matrix allows? The broader reasoning is in the guide to regulated autonomy.
How do you build the grid?
Use two axes.
Reversibility, from your action reversibility register: reversible, compensable, window-bound, irreversible.
Impact if wrong, judged on the worst plausible single error: - Low: minor inconvenience, corrected without the customer noticing. - Medium: customer notices, small financial or service effect, corrected quickly. - High: financial loss, a regulatory breach, harm to a vulnerable customer, or disclosure of personal data.
In each cell, write the maximum depth. Depth 1 is read, 2 is draft for a person, 3 is execute routine work within bounds, 4 is move money or take hard-to-undo actions. Leave depth 5 out; it applies only to orchestrators that coordinate other agents.
Some cells should read "withheld": actions no agent may take at any depth. Those become platform guardrails.
The autonomy decision matrix template
Part A: the grid
| Reversibility / Impact | Low | Medium | High |
|---|---|---|---|
| Reversible | |||
| Compensable | |||
| Window-bound | |||
| Irreversible |
Part B: applying it to actions
| Action | Reversibility | Impact | Matrix ceiling | Current depth | Bounds at current depth | Owner |
|---|---|---|---|---|---|---|
Part B is where the matrix earns its keep. The gap between "matrix ceiling" and "current depth" is your roadmap: every row where current depth is lower is a candidate for a future request, and every row where it is higher is a finding.
A filled sample: a regional insurer's claims operation
Part A
| Reversibility / Impact | Low | Medium | High |
|---|---|---|---|
| Reversible | 3 | 3 | 2 |
| Compensable | 3 | 2 | 2 |
| Window-bound | 3 | 2 | 1 |
| Irreversible | 2 | 2 | Withheld |
Part B (extract)
| Action | Reversibility | Impact | Ceiling | Current depth | Bounds |
|---|---|---|---|---|---|
| Record first notice of loss | Reversible | Low | 3 | 3 | Policy must be active |
| Request missing documents from claimant | Irreversible once sent | Low | 2 | 2 | Templated wording only |
| Pay a glass repair claim | Compensable | Medium | 2 | 2 | Claims up to 500, approved by handler |
| Decline a claim | Irreversible once communicated | High | Withheld | Not available | Always a person |
Two things stand out. Paying a small glass claim feels routine, yet this insurer caps it at depth 2, because a wrong payment is recovered only with effort and customer friction. And declining a claim is withheld outright, which matches the insurer's view that refusing cover is a judgement no agent should own.
The insurer's matrix never shows depth 4. That is a choice, not a template rule. A business with strong reversal controls on its own rail might allow depth 4 in the reversible, medium-impact cell.
How does the matrix relate to supervised and bounded autonomy?
Cells at depth 2 describe supervised autonomy: the agent prepares the work and a named person releases it. Cells at depth 3 describe bounded autonomy: the agent acts, but only inside the limits in the bounds column. If you need a refresher on how the levels are defined, see how autonomy levels for AI agents work.
After an incident, the matrix tells you where depth should go back to. The autonomy reduction plan template uses it as the reference point.
Questions people ask
Who should approve the matrix?
The risk committee or its equivalent, because the grid expresses risk appetite. The business owner of each process signs Part B for their actions. Review both at least once a year.
Can different products have different matrices?
Yes, if their customers or harms differ. A matrix for business customers may allow more than one for retail customers. Keep them separate documents and record which agents fall under each.
Should impact be scored per action or per customer?
Per action, using the worst plausible customer. Then use bounds to narrow it: an action that is high impact for vulnerable customers may be allowed at depth 3 only where the customer is not flagged.