They are asking me to declare what a machine decides and nobody has the list.

Use this today, without hiring anyone

Two exercises that get done in an afternoon and give you half the diagnosis. The first tells you whether you can prove it; the second, how exposed you are.

The provenance test. Take a piece your organization published, or a decision it made, last week and try to answer four things.

1. Which system produced it and in what version.
2. What went in, with what instruction and from whom.
3. Who reviewed it before it went out and what they changed.
4. Where all of that got recorded and for how long.

Time it. That number is your reconstruction time and it is what will be measured on you.

The four classification questions. Run them over the three systems that worry you most.

1. Who does it decide about, and does it affect their access to a job, credit, an essential service or a right?
2. What happens if it gets it wrong, and can the affected person notice and challenge it?
3. How much autonomy does it have in practice, not in the manual?
4. In what context is it used today, not what it was bought for?

Two warnings from experience. The generation nobody recorded is in customer service, communications, internal reports that go out to third parties and the code someone pasted from an assistant. And the systems that show up late are not the big ones: they are the spreadsheets with a model behind them and the tools one area contracted on its own.