Decision ownership assigns one named role accountable for each decision type — its model, policy, autonomy level and outcomes — even when AI recommends and systems execute. Owners approve changes to autonomy, review overrides and outcomes, and answer for the decision in audits.
Roles around a decision
| Role | Responsibility |
|---|---|
| Decision owner | Accountable for outcomes, policy and autonomy level of the decision type |
| Approver | Approves individual decisions at augmented level |
| Model steward | Maintains models, data and calibration behind recommendations |
| Risk & compliance | Sets envelopes and reviews high-consequence changes |
| Operator | Executes, monitors and overrides with reasons |
Decision rights for AI
- One owner per decision typeA role, not a committee; the owner can delegate approval, not accountability.
- Autonomy changes need the ownerMoving from augmented to automated is a signed policy change.
- Overrides require reasonsReasons become evidence for model and policy improvement.
- Periodic outcome reviewThe owner reviews quality, calibration and incidents on a fixed cadence.
- Accountability cannot be delegated to a model.
- One owner per decision type.
- Owners control autonomy levels and review outcomes.
A RACI for AI-assisted decisions
| Activity | Decision owner | Approver | Model steward | Risk & compliance | Operator |
|---|---|---|---|---|---|
| Define objective & constraints | A | C | C | C | I |
| Set autonomy level | A | C | C | R | I |
| Approve individual decisions | A | R | I | I | I |
| Maintain models & calibration | A | I | R | C | I |
| Execute & override | I | I | I | I | R |
| Review outcomes & incidents | A | C | R | C | C |
R = responsible, A = accountable, C = consulted, I = informed. The pattern that matters: exactly one A per row, and it is almost always the decision owner.
Owners in practice
| Decision | Typical owner |
|---|---|
| Maintenance timing on critical rotating equipment | Reliability engineering lead |
| Quality hold / release | Plant quality manager |
| Supplier reallocation | Category or procurement manager |
| Prior-authorization triage | Utilization management medical director |
| Exception rerouting | Transportation operations manager |
| Robot fleet task allocation | Automation / intralogistics lead |
| Fly, restrict or ground an aircraft | Maintenance officer with engineering authority |
Ownership when agents act
When an AI agent can take actions across systems, ownership attaches to each capability, not to the agent. An agent that can both reorder parts and reschedule work touches two decision types with two owners, two autonomy levels and two budgets. Treat the agent as an executor operating inside decisions that people own. See decision autonomy.