It decides on a live freight pipeline. A person still sends the invoice.

Auto Billing is granted the documents the pipeline reads and a read only snapshot of open loads from your transport system, refreshed on a schedule. That is its whole world, and every sentence on this page about safety is downstream of that one fact.

By the time the agent sees anything, the work has happened. The paperwork is read, the fields extracted, each document matched to a load. The agent judges the assembled load with the transport system’s own row beside it, so every rule reads the matched value rather than the raw column.

Sit with the operator for an afternoon and have them say the rules out loud. The carrier rate matches to the penny or it is not a match. Weight always comes from the bill of lading. All three document types have to be matched to the same load before anything is decided. Two candidate load numbers in one batch is never automated. A reference that does not match is not a failure, because agents copy them and omit them.

Each rule goes on the specification as a card carrying the sentence as it was said. Four of them decide on every load, and the sentence stays on the card beside the check that runs.

A document that never carried a signature field is not an unsigned document. Treat those as the same thing and the project loses trust in its first week. It holds loads that were fine and clears loads that were not. So the outcome set has a third value, and it is never folded into the other two.

Three documents arrive at different times from different senders. None is the answer on its own. The artifact is the load: one row, one verdict, and when the verdict is no, the name of the thing that blocked it. A biller does not work a document. They work a load.

A load is auto billable only when nothing failed, nothing was indeterminate, and enough substantive checks actually ran. A load that passed every check because none applied is not clean. It is unexamined. A vacuous pass is the most dangerous output an automation system can produce, and the one a scoring model produces most easily.

Most of what stops a load in the first shadow reading is data rather than rules: stale reference data, incomplete batches where waiting was correct. Fix the data path first and the number moves on its own.

The agent emits its verdict. The load goes through manual billing anyway. An audit table records every field the biller changes. A load that went through untouched is a load that could have been billed automatically. That is a measurement, not an opinion.

Cohorts where billers consistently change nothing get switched to automatic, one slice at a time. Not because a number crossed a threshold. Because that cohort’s billers changed nothing. Where a biller does change something, the change is the finding. A field that always gets corrected is a rule nobody had written down.

There is no failure mode where work disappears. Anything the rules cannot settle goes to the biller who would have handled it anyway. That floor is why operations say yes.

The run history records what the agent would have decided, which documents it read, which rule produced which part of the verdict, and what a person did instead. Those four together make the eventual switch reversible.

That log is where a rule proves itself. A real rule on real data produces a result you can read back to the expression, the field and the value it saw, one record at a time.