Playbook: Autonomy Ladder; what AI is allowed to do
Do you know what your AI/Agent is allowed to do? Use this playbook to get familiar with understanding how much action the system is actually allowed to do.
The Autonomy Ladder: What AI Is Allowed To Do
Use this playbook to choose the right level of autonomy before AI moves from answering questions to acting in workflows.
In plain English
Many teams talk about “AI agents” without defining what the system is actually allowed to do.
As agents move across tools, the biggest confusion is often not the model. It is decision rights. Can the AI answer? Draft? Recommend? Trigger? Update? Approve? Each step changes the human role.
How this connects to the sequence
The Readiness Gate checks if a workflow is ready. The Autonomy Ladder defines how much action the system is allowed to take.
Signal linked to this playbook
Job fear starts where decision rights get fuzzy
This Signal connects job anxiety to decision-rights clarity. People fear AI most when they cannot see what it is allowed to decide.
The ladder
- Level 0 — No AI: the workflow stays manual or non-AI for now.
- Level 1 — Retrieve / explain: AI helps find and explain information.
- Level 2 — Draft / prepare: AI creates a draft or work packet for a human.
- Level 3 — Recommend: AI proposes a next action with rationale.
- Level 4 — Trigger with approval: AI prepares the action, human approves.
- Level 5 — Act within limits: AI takes bounded action under defined rules.
- Level 6 — Autonomous exception handling: only for mature, low-risk, highly monitored workflows.
Use this when
- People are unsure whether AI is assisting or deciding.
- An agent is being given tool access.
- A team says the human is still in the loop but nobody knows where.
- Role anxiety is rising because decision rights are unclear.
- A workflow could affect customers, finance, safety, HR, or operations.
The operator questions
- What can the AI read?
- What can it draft?
- What can it recommend?
- What can it trigger?
- What can it update?
- What can it approve?
- Where does the human approve, override, or escalate?
- What actions are explicitly forbidden?
What good looks like
- Each workflow has one current autonomy level.
- Every higher level has a reason, not just a vendor demo.
- Human approval is placed where risk, accountability, or judgment require it.
- The autonomy level is reviewed as the workflow matures.
The first move
Choose five workflows. Assign the current autonomy level and the maximum allowed autonomy level. If nobody can agree, the workflow is not ready to scale.
Human work signal
The higher the autonomy, the more clearly the human role must be designed: reviewer, approver, exception handler, supervisor, trainer, or outcome owner.
What to capture in the worksheet
| # | Field | Why it matters |
|---|---|---|
| 1 | Workflow name | Identifies the workflow being governed. |
| 2 | Current autonomy level | Shows what the AI is allowed to do today. |
| 3 | Maximum allowed level | Sets the upper limit on AI decision rights. |
| 4 | Human role | Defines human accountability and oversight. |
| 5 | Forbidden actions | Prevents unacceptable or high-risk behaviour. |
| 6 | Approval points | Identifies where human sign-off is required. |
| 7 | Escalation path | Defines how exceptions and issues are handled. |
| 8 | Review cadence | Ensures autonomy is reassessed over time. |
| 9 | Decision | Records the agreed autonomy outcome. |
Get the lightweight workbook
The public playbook gives you the method. The member workbook gives you the simple working sheet across the Playbooks.
