Use this when

An action is serious enough that a single approval is not enough protection: large payments, irreversible changes, anything with major consequences.

The pattern

AI PREPARES the action
      ↓
FIRST PERSON reviews and approves
      ↓
SECOND PERSON reviews and approves
      ↓
SYSTEM EXECUTES and logs both approvals

Two different people must approve before the action happens. Neither can complete it alone.

Why it works

For the highest-risk actions, one reviewer is a single point of failure. They may be rushed, mistaken, or pressured. A second, independent approval makes a costly error much less likely, because both people would have to miss the same problem. It is the strongest form of draft, approve, execute, reserved for where it is warranted.

This is a long-standing control in finance and operations, and it applies cleanly to AI-prepared actions.

How to apply it

  1. Identify the actions serious enough to justify it, usually high-cost or irreversible ones.
  2. Require two distinct approvers, ideally with different perspectives.
  3. Give both enough context to judge, not just a button.
  4. Record both approvals with the action. See the checklist before giving an agent write access.

Good fit

  • Large financial transactions and irreversible commitments.
  • Actions with legal, safety or reputational weight.
  • Early operation of an agent on a high-stakes action, while trust is being built.

Poor fit

  • Routine, low-risk, reversible actions, where it adds friction without real benefit.
  • High-volume work, where requiring two approvals per item is impractical. Reserve it for the cases that warrant it.

What this does not solve

Two-person approval reduces the chance of a costly mistake; it does not eliminate it, and it is too heavy for everyday actions. Use it selectively, for the actions where the cost of being wrong justifies the extra effort.