Copilot
Suggest
- How it runs
- It suggests inside the person's own tool. The person acts.
- The human
- Does the work and owns the result.
- Example
- Write production code.
Where the human sits
Answer the questions below about the task. Your answers give one of four patterns. Write the pattern in field 8 of the Agent Canvas.
Steps: can someone write them down today?
- Predictable
- The steps are the same every time. You could hand them to a new colleague.
- Ambiguous
- The steps depend on what the agent finds along the way.
Stakes: what does a wrong answer cost?
- Low
- Minutes of rework.
- High
- Money, trust or compliance.
Predictable steps
Ambiguous steps
Low stakes
AI Pipeline
Autonomous
- How it runs
- A system event starts it. Nobody watches each run.
- The human
- Spot-checks results and handles the exceptions.
- Example
- Route inbound support tickets.
Autonomous Agent
Act and notify
- How it runs
- It plans, picks its tools and runs to the end in a sandbox.
- The human
- Reviews the result before it leaves the sandbox.
- Example
- Draft a research brief.
High stakes
Gated Agent
Act with confirmation
- How it runs
- It gathers context, proposes an action, then waits for a yes.
- The human
- Approves, edits or rejects each action.
- Example
- Process high-value refunds.
Autonomy levels, lowest first
- Suggest. The agent proposes. The person does the work.
- Draft for approval. The agent drafts. The person edits and sends.
- Act with confirmation. The agent acts after a yes.
- Act and notify. The agent acts, then tells the person.
- Autonomous. The agent acts. People check samples.
- The stakes set the highest level you allow.
- A condition from a works council, a regulator or a compliance owner overrules the grid.