Case: data governance as the foundation
A retailer's demand-forecasting AI had quietly degraded for months. Forecasts drifted, stock decisions worsened, and no one noticed until the losses showed up in inventory. The model hadn't broken. Its data had.
A supplier had changed how a category was coded upstream. The model, trained on the old coding, kept confidently producing wrong answers from newly-meaningless inputs. Nobody was watching the data feeding the system, because governance had focused entirely on the model and treated its inputs as a given.
ISO 42001 keeps returning to data governance because AI systems are only as trustworthy as what flows into them. A perfect model on drifting data is a confident liar. The failure is rarely dramatic; it is silent, gradual, and invisible until it's expensive.
We instituted data-quality monitoring at the model's inputs — schema checks, distribution alerts, an owner for each critical feed. The next upstream change triggered an alert in a day, not a quarter. Governing AI, it turns out, is mostly governing the data beneath it. Everything else sits on that foundation.