Confidence-gated routing

Use confidence as a second axis. The answer tells you what; confidence tells you whether to act.

The problem it solves

The answer and the confidence are two independent pieces of information. Branching on the answer alone discards half of what the model told you.

Confidence-gated routing means: take the answer, then use confidence to decide whether that answer is reliable enough to act on. This is the basis for systems that are both reliable and safe.

Example: voice banking commands

Imagine building a voice interface for a bank account. You always want reasonable confidence in the interpreted intent, but some actions are riskier than others and therefore demand a higher threshold.

Step 1: determine intent

A single Choice call yields the intent, with candidates such as check_balance and approve_transfer.

Step 2: gate on confidence

action = response.answers["intent"]

# Below 0.6 confidence on any action, route to a human
if action.confidence < 0.6:
    route_to_support_agent(account_id)

elif action.choice == "check_balance":
    # Low stakes. 0.6 confidence is sufficient.
    show_balance(account_id)

elif action.choice == "approve_transfer":
    if action.confidence > 0.85:
        # High stakes, but high confidence. Proceed.
        ...
    else:
        # High stakes, moderate confidence. Confirm first.
        ask_user_to_confirm(account_id)

Why the thresholds must differ

Look at the three thresholds in that code:

Threshold	Purpose
`< 0.6` intercepts everything	When the model reports uncertainty, no action runs
`check_balance` at 0.6	Read-only; a wrong screen is recoverable
`approve_transfer` at 0.85	Moves money; mistakes are not reversible

This is “thresholds scale with risk” expressed in code. A single global threshold forces you to choose between over-interrupting on low-stakes actions and being under-cautious on high-stakes ones.

Practice notes

Set a hard floor first, then per-action thresholds. The floor (0.6 here) catches the model’s own reported uncertainty — it is the safety net. Action thresholds sit above it, graded by risk.

Make thresholds explicit configuration, not scattered magic numbers. Define every action’s threshold in one place so it can be audited and tuned. When someone asks “why does this transfer need human confirmation”, you can point at a specific number.

Do not let confidence replace business validation. Confidence is the model’s self-assessment, not a substitute for business rules. Amount limits and permission checks still belong in code.

Calibrate thresholds against real data. The official guidance is explicit: correct thresholds depend on your domain and how the model performs on your use case. Start conservative, test with your own data, adjust as you observe.

When not to use it

If a decision has no consequence when wrong — tagging a log line, say — adding a confidence gate only adds complexity and human cost. The value of confidence routing scales with how irreversible the decision is.

Confidence — how confidence relates to probabilities
Intent routing — usually paired with this pattern
Composite scoring — handling confidence in ranking