Confidence Needs a Cost Function

A routing model can answer a narrow question quickly: which queue should receive this request, which tool should run next, or whether the factory should proceed. Measure enough answers and you get an accuracy number. That number looks like a decision rule. It is not one.

The same error can carry wildly different costs. Sending a billing question to technical support creates delay. Sending a permission change down a low-scrutiny path can create damage. If both count as one miss in a benchmark, the measurement has erased the part the business actually needs to govern.

Start with the consequence, not the score. Name the cost of a false approval, a false block, an unnecessary escalation, and a silent misroute. Then set a policy for each class of work. A model result becomes useful only after the factory knows what that result is allowed to do.

Confidence matters when it means something repeatable. If a route marked highly certain fails as often as a coin toss, the number is decoration. If lower scores actually collect the ambiguous cases, the factory can spend more reasoning there and let routine work move without a person watching every turn.

But calibration is conditional. A model can behave well on the examples used to test it and fail when the language, request shape, or operating context changes. A neat probability does not announce that the input came from a world the model barely knows. It can remain precise while the ground moves underneath it.

That is why the route needs an abstain condition. Unfamiliar language, missing context, conflicting signals, or a score near the decision boundary should change the path. The factory can ask for clarification, choose a stronger evaluator, narrow the available action, or stop. Uncertainty should operate machinery, not merely appear in a log.

A multilingual system is not one router with translated labels. Requests that look equivalent on paper can carry different phrasing, politeness, shorthand, and ambiguity in practice. Performance can hold in one language and collapse in another. One global score averages away the place where the failure lives.

Track outcomes by the conditions that can change the decision. Language is one. So are request type, consequence class, route, model version, and the presence of required context. You do not need a dashboard full of vanity slices. You need enough state to answer a blunt question: where does this system stop earning the authority we gave it?

The policy should follow that evidence. A route can run automatically in conditions where it has repeatedly held and require more proof elsewhere. That is not unfairness dressed up as caution. It is refusing to turn an average into permission.

When a decision changes what happens next, preserve why it happened. Record the candidate routes, the chosen route, the confidence, the policy version, and the conditions that affected the choice. Then connect that receipt to the outcome: accepted, corrected, escalated, reversed, or failed later.

Without that chain, the team can see that the router is wrong without learning how it is wrong. A person fixes the queue, an agent retries with a stronger model, and the evidence disappears into the successful result. The system looks healthy because rescue erased the miss. The factory pays for the same rescue again tomorrow.

The receipt turns correction into a control change. Repeated overrides can move a threshold. A failure cluster in one language can restrict authority there. A costly false approval can force independent verification for that route. The lesson survives the person who noticed it and reaches every later run.

People should decide which mistakes the business can tolerate. They should not inspect every routine routing choice forever. Once the consequence classes, escalation rules, and proof requirements are explicit, agents can apply them more consistently than a rotating queue of people working from instinct.

That will eliminate coordination work. Triage, assignment, follow-up, and routine approval are jobs today because the decision rules live in people's heads and the state lives across inboxes. Put the rules and state into the factory and much of that labor stops being a role. Human judgment moves to defining the policy and owning the exceptions that can actually change it.

Our position is simple: confidence should never create authority by itself. Price the mistakes. Watch the conditions. Keep the receipt. Let the machinery handle decisions it has earned, and make uncertainty choose a safer route before production teaches the lesson for you.