Classic automation copied fields from one system into another and stopped at the first decision. A human read the document, judged the case, chose the route, and only then did the robots resume. The queue in front of that human was the real cost of the process.
Language models changed which steps can be automated. Reading a document, classifying a case, and proposing the next action are now machine tasks, with accuracy that can be measured against people doing the same work.
The escalation contract
Judgment automation works under a contract: the system acts alone only above a confidence threshold, sends everything else to a person, and records every decision either way. The contract is what makes automation auditable, and audit is what makes it acceptable to risk, compliance, and leadership.
In practice the threshold starts conservative. As the measured accuracy record accumulates, the threshold earns its way up, and the human queue shrinks case class by case class.
Start with the queue, not the technology
The best first project is rarely the most exciting one. It is the queue where volume is high, the decision is narrow, the ground truth is checkable, and the cost of a mistake is recoverable. Document intake, case triage, and reconciliation flows fit that shape.
One well-instrumented queue, automated with an escalation contract and run in parallel until proven, does more for an automation program than any platform decision.