Autonomy
Copilot vs. agent
Definition
A copilot proposes and a person disposes; an agent decides and acts. The distinction is not sophistication — it is who is accountable for the outcome.
Copilots are easy to approve because the human remains the decision-maker of record, and easy to over-credit because their value is diffuse and hard to isolate from everything else happening in a role. Agents concentrate value and concentrate risk in the same design. Conflating the two in a business case is how programs get funded as agents and governed as copilots, or the reverse, and either mismatch tends to surface at the worst possible time.
The accountability test
Ask who is named when the output is wrong. If the answer is a person who reviewed it, you have a copilot, and the governance model should reflect that a human remains the point of accountability. If the answer is the process, you have an agent and you need the controls that go with one — delegation boundaries, escalation paths, monitored authority — because there is no individual standing between the system's decision and its consequence in the way there is with a copilot.
Copilot value is hard to bank
Time saved that does not translate into fewer handoffs, faster cycle time, or higher capacity tends not to show up in a financial statement, no matter how genuinely useful the tool feels to the people using it day to day. Say up front which of those you expect — headcount avoidance, cycle time, capacity — and measure that specific claim, rather than defaulting to a vague sense that the tool is helpful, which is not a number anyone can hold the program accountable to later.
The transition from copilot to agent should be evidenced
Many programs start as a copilot to build trust and later consider removing the human review step to capture more value, which is a reasonable path but a real threshold to cross deliberately. That transition should be justified by review data — a sustained low override rate on a representative case set — rather than by schedule pressure or the assumption that a tool proven useful with review will behave identically without it, which is not a safe assumption for any workflow with real consequence.
In practice
- State in the business case which mode you are funding
- Do not claim headcount effects from a review-required copilot
- Promote a copilot to an agent only after the review data supports it
- Define in advance what review evidence would justify removing the human step
