Escalate to a human
Make uncertainty an explicit operational path.
When the API returns ask_human, create a review item containing only the evidence the reviewer needs: the original situation, question, option descriptions, suggested answer, probability distribution, pinned model ID, and threshold.
Use your own request or case ID for deduplication. Do not use a new decision response ID as the sole deduplication key: a retry can produce another ID for the same business case.
Review flow
- Persist the pending case before acknowledging it to the caller.
- Present the evidence and all options; label the model’s answer as a suggestion.
- Require a human decision before any irreversible action.
- Record who approved the action and apply your normal authorization checks.
- Expire stale reviews and provide a visible status to the user.
A queue with no reviewer is a blocked workflow. Set an ownership and response-time policy before enabling it. If inference fails entirely, use the same review path or return a temporary failure; never fabricate a confident answer.
Use human outcomes for workflow evaluation only under your data policy. Do not assume customer data is automatically training data.