Answer: Only when the action is explicitly authorized, low enough consequence, observable, recoverable and consistently validated. Autonomy should be assigned action by action, not to the entire agent.

What this means in practice

Read-Only Research

Reading approved public or internal information changes nothing, so it is usually the first place autonomy can be allowed.

Internal Classification

Sorting and tagging internal items can run on its own once testing shows it is consistently accurate and mistakes are easy to correct.

Draft Preparation

AI can prepare drafts without approval as long as a person reviews them before anything is sent or committed.

Approved Low-Risk Record Updates

Small, reversible internal changes can be narrowly pre-authorized, with logging, after the workflow has been validated.

Routine Monitoring

Watching for missed runs, stale data or overdue items and raising alerts is low-consequence work that can run unattended.

Pre-Authorized Reversible Actions

An action can run without case-by-case approval only if it was explicitly authorized in advance, is observable and can be undone.

A practical construction test

  1. Define the business outcome.
  2. Measure how the workflow performs today.
  3. Identify what AI is being asked to interpret or decide.
  4. Define what remains under human authority.
  5. Test with real examples and acceptance criteria.
  6. Measure results before expanding scale or autonomy.

Where governance fits

The AI model can change. The company’s operating rules should not change silently with it. Catha’s Governance Cage is designed to preserve the business authority layer above the model and workflow.