Getting AI past governance in a regulated lab

When a medical laboratory group asked us to bring AI into their laboratory information system, the technical work was the easy half. Every previous attempt had died in the same place: a governance conversation nobody could finish, because nobody could answer who approves a model, who reviews its output, and what gets logged when it influences a result.

So we inverted the project. Before any model was selected, we sat with quality and clinical leadership and built the frame: named roles for every AI-touched decision, review gates a use-case must pass before and after deployment, and audit logging specified to the same standard as any other instrument in the lab.

That order changed the politics entirely. Compliance stopped being the department that says no and became the co-author of the intake process. When the first use-case arrived for review, the question was no longer "should AI be allowed in the lab" — it was "does this proposal pass the gates we wrote." One of those questions can be answered in a meeting; the other cannot.

The technical principle underneath is the one we hold for every regulated deployment: the system recommends, people decide. Clinicians keep final authority over every result, and the audit trail records both the recommendation and the human decision. That is not a limitation on the AI — it is the property that made production sign-off possible.

The lab now has AI in production and, more importantly, a repeatable path for the next use-case. If your organisation has an AI mandate stuck behind a governance wall, the lesson travels: stop trying to get the model approved. Build the thing that does the approving.