Use this when

An AI is right most of the time but not always, and you want it to handle the easy cases automatically while a person handles the uncertain ones.

The pattern

AI PRODUCES a result and a sense of how sure it is
      ↓
CONFIDENT ENOUGH?
  ├── Yes → proceed automatically
  └── No  → send to a person

You set a level of confidence. Above it, the workflow proceeds. Below it, the case goes to a human. The AI is allowed to say "I am not sure enough."

Why it works

Most AI errors cluster in the uncertain cases. By routing those to a person, you keep the reliable, confident cases automated and put human attention exactly where it is needed. This raises overall quality without giving up the efficiency of automation. It is a key part of escalation design and of triage and route.

How to apply it

  1. Get a confidence signal from the step, or use rules of unusualness (unexpected format, missing fields, rare category) as a proxy.
  2. Choose a threshold based on the cost of errors: the more costly a mistake, the higher the bar.
  3. Route below-threshold cases to a person or an exception queue.
  4. Review the human-handled cases to tune the threshold over time.

Good fit

  • Classification, extraction and routing, where some cases are clearly harder.
  • Workflows where a person is available to handle the uncertain minority.

Poor fit

  • Steps with no meaningful confidence signal and no good proxy for uncertainty.
  • Cases where even a confident answer must always be checked anyway.

What this does not solve

A threshold sends uncertain cases to people; it does not make the confident cases correct, and confidence signals can be imperfect. Combine it with sampling of automatic results and a real human fallback for the cases it defers.