Know exactly what the AI touched, on which shipment, and when

The question that decides whether an operations team will let AI anywhere near live freight is not how clever the model is. It is simpler: if it does something wrong, can I see what it did? On MoveTheWheels the answer does not depend on the AI cooperating, because the recording happens underneath it.

Every operation the logistics service executes passes through an audit layer that derives a stable name from the permission the field itself declares — the resource and the action, prefixed with the product — and writes it with the actor and actor type, the workspace and tenant, the specific record touched, whether it succeeded, and the error text if it did not. An agent deleting a shipment and a dispatcher deleting the same shipment produce the same shape of record, which is exactly what makes the trail worth reading.

Redaction is built in rather than left to the caller. Field names that look like credentials or tokens, driver licence numbers, tax IDs, business licences, email addresses and phone numbers are replaced before the entry is written, and the captured arguments and results are capped by depth, by array length, by number of keys and by string length — so the audit trail stays a trail and never quietly becomes a second copy of your customer and driver data.

Two things are still missing, and they are the AI-specific half. There is no purpose-built run ledger yet — the model used, the prompt, the latency, the credits burned, and the decision an automated run reached — and there is no undo path for a write an agent initiated. The operation record tells you what happened and who caused it; the ledger that tells you why an autonomous run chose it arrives with the shared AI runtime. An evaluation set that regression-tests the assistant’s answers before a prompt or model change is likewise on the list, not in the repository.

Illustrative scenario for a pre-launch platform. The per-operation audit ledger with redaction and actor attribution is built and runs on every request; the AI run ledger, the undo path and the evaluation set are roadmap.