Short answer
Decision autonomy is the level of authority a system has over a specific decision. There are five levels — human, augmented, on-the-loop, automated and agentic — and each decision earns its level through policy based on consequence, evidence, confidence, reversibility and track record, with the ability to step back down at any time.
The five levels
01 · HumanA person decides. AI assembles context and flags gaps.
02 · AugmentedAI recommends options with evidence; a named owner approves.
03 · On-the-loopAI acts inside a time window; humans monitor and can override.
04 · AutomatedPolicy-bounded, reversible decisions run end-to-end; exceptions escalate.
05 · AgenticMulti-step goal pursuit inside earned authority, budgets and audit trails.
What changes at each level
| Level | Who decides | Who acts | Typical use |
|---|---|---|---|
| 01 Human | Person | Person | Novel, strategic or irreversible decisions |
| 02 Augmented | Person, from AI options | System after approval | Supplier reallocation, prior-auth triage |
| 03 On-the-loop | System, within a window | System; human can override | Truck rerouting, maintenance scheduling |
| 04 Automated | System, within policy | System; exceptions escalate | Fraud holds, robot task allocation |
| 05 Agentic | System pursuing a goal | System across multiple steps | Multi-step procurement, fleet-wide rebalancing within budget |
How a decision earns autonomy
- ConsequenceLower consequence permits higher autonomy.
- ReversibilityReversible actions can move up faster than irreversible ones.
- ConfidenceCalibrated confidence above threshold on real outcomes, not test sets.
- Track recordShadow-mode agreement and outcome history over a defined period.
- PolicyA named owner signs off the level; risk and compliance agree.
- RevocabilityA trigger that moves the decision back down automatically — drift, incidents, novelty.
Autonomy by industry
| Decision | Level today | Why |
|---|---|---|
| Robot mission reallocation | Automated in envelope | Frequent, reversible, low consequence |
| Truck exception reroute | On-the-loop | Time-critical, reversible, dispatcher can override |
| Maintenance timing | Augmented | Costly, schedule-coupled, engineer owns |
| Prior-auth approval | Augmented | Clinical judgment required |
| Ground an aircraft | Human | Airworthiness authority |
Autonomy for AI agents
Agents change the question from “can the model do this?” to “which decisions is this agent allowed to take, with what budget, and how do we undo them?” Treat each agent capability as a decision type with its own autonomy level, budget, rate limit and audit trail. See decision safety and decision observability.
Key takeaways
- Autonomy is granted per decision, not per model.
- Five levels: human, augmented, on-the-loop, automated, agentic.
- Consequence, reversibility, confidence and track record decide the level.
- Every level needs a way back down.
Frequently asked questions
What are the levels of AI decision autonomy?
Human, augmented, on-the-loop, automated and agentic.
What does human-on-the-loop mean?
The system acts within a time window or envelope while a human monitors and can override before consequences land.
When should a decision be fully automated?
When it is frequent, bounded by policy, reversible, well calibrated on real outcomes and has exception escalation.
Can autonomy be revoked?
Yes. Drift, incidents, novelty or policy changes should automatically move a decision back to a lower autonomy level.