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.
Monitor input distribution against training, output distribution over time, performance against whatever ground truth arrives later, and the rate of human override. A rising override rate is the best early warning you have, and it is usually already visible in a queue nobody reads.
Human oversight
Meaningful, not nominal. A reviewer approving hundreds of decisions an hour is not overseeing anything — they are laundering the model's output through a person.
Meaningful oversight requires the reviewer to see why the model decided, to have time to disagree, and to have their disagreement change the outcome and be recorded.
Documentation
Per model: intended use and exclusions, training data and its provenance, evaluation results by segment, known limitations, monitoring in place, and the owner. This is what you need when someone asks why a decision was made — and increasingly what a regulator expects to see.
Retirement
Have a way to turn it off. Know what happens to the process when you do, and confirm the fallback still works — a manual path that has not been exercised in two years is not a fallback.
Sources
references/sources.md in this skill lists the outside authorities that settle the questions
here — what each one is authoritative for, and what you may do with it. Check them before
answering on anything they cover, and cite what you used. Most are free to read and not free
to reproduce; the use note on each is binding.
Never
- Deploy without an evaluation set and a monitoring plan.
- Use a model outside its documented intended use because it seems to work.
- Train or fine-tune on customer data without confirming the lawful basis covers it. The basis for collecting it rarely extends to this.
- Let a model make a consequential decision about a person with no route to human review.
