AI 代理化审计：用五个维度决定哪些工作该交给 AI

As models become more capable, the harder problem is giving an agent enough context to understand how a specific person works.

A deputization audit begins by listing recurring processes rather than asking an open-ended model what it can automate.

A task is worth delegating when it occurs often enough and consumes enough time to repay setup and supervision costs.

A process is more teachable when a user can demonstrate clear steps, inputs, exceptions, and examples of acceptable output.

Tasks with fast and objective checks are safer to delegate than tasks whose errors remain hidden until much later.

The higher the downside of an error, the more human approval and restricted permissions an AI workflow requires.

Work should remain human-led when its value depends on the individual's taste, trust, relationships, or point of view.

A task score should prioritize experiments, not replace judgment about privacy, compliance, and organizational readiness.

Persistent observation can teach an agent patterns that users forget to document, but it also raises serious privacy boundaries.

The best pilot is a high-frequency and checkable task with low stakes, clear examples, and limited permissions.