AI and ML governance

Regimes governing automated decision-making differ by jurisdiction and sector and are changing quickly. Anything affecting credit, employment, housing, insurance, healthcare, or education carries specific legal obligations, involve Legal & Risk and qualified counsel rather than treating it as an engineering question.

Define intended use before evaluating anything

Write down what the system is for, what it is not for, who is affected by its output, and what happens when it is wrong. Most AI failures are use outside intended scope by someone who did not know the scope existed.

Then decide the consequence tier, because it sets everything after it:

• Advisory, a human decides, the model suggests. Lightest oversight.
• Assistive, the model acts, a human reviews before effect.
• Autonomous, the model acts with effect. Highest bar, and rarely appropriate where a person is materially affected.

Evaluation

A held-out evaluation set that reflects real inputs, including the awkward ones. Built before deployment and kept stable, or you cannot compare versions.

• Measure the failure that matters. Aggregate accuracy hides the errors you care about. A model that is 95% accurate and wrong disproportionately on one group is not 95% good.
• Evaluate by segment, always. This is where fairness problems and quiet degradation appear.
• Both error directions. False positives and false negatives usually have different costs, and the threshold should reflect that ratio rather than a default.
• Establish a baseline. Compare against the current process, often a simple rule, not against zero. Plenty of models fail to beat the heuristic they replaced.

Monitoring

Models degrade silently: the world moves, inputs drift, and accuracy falls without any error being raised.