Change Control: Lessons from Budget Modifications

In grants, you couldn't just move money between budget lines because it felt right. Significant changes needed prior approval — a written request, a justification, a sign-off. It wasn't about distrust. It was about making sure changes were deliberate, because quiet drift is how budgets (and programs) come apart.

Live AI systems drift the same way, and nobody asks permission.

A model gets swapped for a newer version. The prompt gets tweaked by whoever's on shift. A new data source gets connected. Each change feels small. Together, they can transform what the system does — with no record and no review.

Borrow the budget modification discipline. Define which changes need approval before they happen: model swaps, new data sources, changes to what's automated. Keep it lightweight — a short written note and one sign-off — but make it real.

The principle is simple: the system running today should match the system that was approved. In grants, undocumented changes were called unallowable. In AI, they're called incidents waiting to happen.

Control the changes.

I explain frameworks; I don't give legal counsel.