Platform / Governance
Controlled autonomy
Do not trust an agent with everything on day one. Start with assistance, then progressively increase autonomy as the record shows it has earned it. The path is not manual to autonomous. It is five controlled levels.
The five levels
- L1
Assist
AI recommends. A person does the work and makes every decision.
Use when: New agents, high-risk or irreversible domains, anything touching people decisions.
- L2
Approve
AI prepares the action. A human approves before anything happens.
Use when: Actions are reversible only with effort, or the agent has little track record.
- L3
Supervise
AI acts within limits and is monitored. Humans approve exceptions above thresholds.
Use when: Routine, bounded actions with a recorded history of accepted work.
- L4
Autonomous
AI acts independently within strict policy and risk bounds.
Use when: Well-understood processes where limits, rollback and monitoring are proven.
- L5
Adaptive
AI improves its own behavior within controlled boundaries.
Use when: Mature use cases where changes are tested, bounded and reversible.
What changes as the level rises
| Level | Approval rules | Monitoring | Risk boundaries | Audit depth |
|---|---|---|---|---|
| L1 Assist | Human decides everything | Review of recommendations | Read-only access | Inputs and recommendations |
| L2 Approve | Every action approved | Approval and edit rates | Draft and staging writes only | Plus the action prepared and the approver |
| L3 Supervise | Thresholds and exceptions only | Live monitoring and alerts | Explicit limits per action | Plus policy applied at each step |
| L4 Autonomous | Policy-defined gates | Automated anomaly detection | Hard bounds enforced at runtime | Full trace, sampled human review |
| L5 Adaptive | Changes to behavior are gated | Drift and regression monitoring | Boundaries cannot be self-modified | Plus versioned behavior changes |
Worked example: the Procurement Agent, L1 to L3
L1 Assist
The agent reads approved supplier info, compares pricing and recommends a purchase. A buyer does everything else.
Track how often buyers accept the recommendation.
L2 Approve
It prepares draft POs. A buyer approves each one before it goes anywhere.
Track edits, rejections and contract issues it caught.
L3 Supervise
It creates draft POs within limits and routes them on. Purchases above threshold, contractual changes and sensitive external comms still require approval. It still cannot make payments.
Monitored live; limits tightened or relaxed from the record.
At every level the agent cannot access unrelated employee data, approve its own high-value transaction or modify restricted records. See the full procurement agent.
Questions
Why not go straight to full autonomy?
Capability without control is not enterprise-ready. Trust in an agent is earned from evidence, so autonomy is increased progressively and per action.
Can an agent raise its own autonomy level?
No. Level changes are made by the accountable owner and recorded in the agent Trust Profile.
Does every agent need to reach Level 5?
No. Many agents should stay at L2 or L3 permanently. The right level depends on risk, reversibility and the evidence you hold.