Nothing ships unaudited: how we govern a workforce of AI workers

Every conversation about AI workforces eventually arrives at the same question, usually from the most senior person in the room. What stops it doing something stupid?

Our answer is a single rule with teeth. Nothing ships unaudited. No article publishes, no message sends, no record changes, no money moves, until the output has cleared a review. The interesting part is not the rule. It is how it is enforced.

Withhold the tools, not just the permission

You can tell an AI worker not to publish. It will comply almost every time, and almost is not a governance standard. So we do not rely on instructions. Workers that draft content are not given the publishing tool at all. Workers that prepare payments cannot reach the system that sends them. By default, every worker in the system produces recommendations, drafts and prepared actions, and the ability to make something real lives one layer up, behind the review.

This is the difference between a policy and an architecture. A policy asks the worker to behave. An architecture makes the unwanted action impossible from where the worker stands.

The review layer

Review takes one of two forms, chosen per step according to risk. A human sign-off, for anything with a large blast radius: money leaving, public statements, personnel matters, commitments to clients. Or an AI auditor operating within defined parameters, for the high-volume work where the criteria can be written down: does the draft follow the style rules, do the figures reconcile, are the claims sourced, is the schema valid.

Defined parameters is the load-bearing phrase. An AI auditor is not a second opinion from the same model in a different mood. It is a separate worker with a written brief, explicit pass and fail criteria, thresholds for escalation, and its own error rate tracked over time. When it is unsure, the case routes to a person. When it fails repeatedly, a circuit breaker halts that line of work entirely until someone looks.

The structure around it

Workers sit inside an organisational chart, the way people do. Departments own domains, escalation paths are explicit, and coordination happens through shared records rather than free-form chat between agents, because free-running agent conversations drift and a drifting system is an ungoverned one.

Underneath everything sits the log. Every action, every decision, every approval and every rejection is recorded with who did it, when, and on what basis. When something goes wrong, and occasionally something does, the question "why did the system do that" has an answer measured in minutes of reading, not weeks of forensics.

None of this slows the work as much as people expect. Most output clears review in seconds because most work is routine and the criteria are clear. What the structure buys is the thing no model card can offer: an operation you can let run overnight and still explain to an auditor in the morning.