Use this when

Any workflow where the AI might be unable to complete a case, and someone must be able to take over reliably.

The pattern

AI attempts the case
      ↓
CANNOT PROCEED (unsure, unusual, failed, unavailable)
      ↓
HAND OFF to a person, with context
      ↓
person completes the case in a working channel

There is always a path back to a human, and that path works even when parts of the system do not.

Why it works

AI will sometimes be unable to finish a case: it is uncertain, the case is unusual, an action fails, or a system is down. A human fallback means those cases are completed rather than dropped. It is the backstop behind escalation and failure design.

The important detail is that the fallback must not depend on the same thing that failed. If the AI's channel is down, the person needs another way to receive and handle the case.

How to apply it

  1. Decide who the fallback person or team is for each kind of case.
  2. Make sure the case reaches them with enough context to act, often via an exception queue.
  3. Confirm the fallback works when the primary systems are unavailable.
  4. Make sure the person has the information, time and authority to actually resolve it. See what human in the loop actually means.

Good fit

  • Every consequential workflow. A human fallback is close to mandatory once actions have real effects.

Poor fit

  • There is rarely a case against having a fallback. The mistake is usually having one that only works when everything is already fine.

What this does not solve

A human fallback ensures cases are completed; it does not scale infinitely. If the AI hands off too much, the fallback becomes the bottleneck. Use it as a safety net, and watch its volume as a signal about how well the automated path is really working.