Your model is a policy. Govern it like one.
Regulators now ask who approved a model's behaviour, not only how accurate it is.
Most AI programmes treat governance as a gate at the end. By then the choices that matter are already made: which data, which trade-offs, which people can override the output.
We move those choices to the start, write them down in plain language, and test the model against them every week it runs.
Four documents
- A decision statement. What the model decides, for whom, and what happens when it is wrong.
- A data sheet. Where the training data came from and what it leaves out.
- An evaluation plan. The tests the model must pass before and after launch.
- An override policy. Who can overrule the model, and how that is recorded.
Two meetings
A monthly model-risk review with the business owner, and a quarterly review with your second line.
One dashboard
Drift, overrides and complaints on one page that the business owner reads, not just the data team.