AI Compliance Checklist

Ownership and Policy

0/5

One named person is accountable for AI decisions, not a committee that meets occasionally.You can say the name, and that person can stop a system going live.

A written policy states which AI tools staff may use and which data may be put into them.It names tools and data categories, not just principles like fairness.

Every AI system passes an approval gate before launch, signed off by someone outside the team that built it.A record of what was approved, by whom, and on what date.

Model, prompt and configuration changes are versioned and re-approved, not edited live.You can tell which version produced an output from three months ago.

Staff who use AI in their work have been trained on what the policy allows and where it stops.A dated record of who was trained, refreshed when the policy changes.

Risk and Human Review

0/5

Each use case has a written risk assessment covering what happens when the output is wrong.It states who is harmed, how the error surfaces, and how it is corrected.

You have defined which decisions require a human reviewer, and that reviewer can overturn the system.Reviewers reject outputs sometimes. A reviewer who never rejects is a rubber stamp.

Anyone affected by an automated decision has a route to question it and reach a person.The route is published, and there is a stated time to respond.

Outputs are checked for skewed treatment of the groups your system actually serves.You know which attributes you check against and when you last checked.

An incident plan says who is called, how the system is switched off, and who tells the people affected.It has been rehearsed or used at least once, not only written.

Data and Transparency

0/6

You have an inventory of what data each AI system reads, who owns that data, and where it is stored.One list, current, that names a data owner per source.

For every source you can state the lawful basis or consent it rests on, and the purpose it was collected for.You could show where that consent or basis is recorded.

You know where inference runs and what leaves your network: prompts, documents, customer records.For each system: hosted by a vendor or run in your own environment, and what is sent.

Inputs, outputs, model version and the person who acted on the result are logged, and kept long enough to reconstruct a decision.You could rebuild a decision made last quarter from the logs alone.

Someone can explain, in plain language, what drives a given decision and what its known failure modes are.A written explanation a non-technical colleague understands.

Accuracy is monitored against a baseline, so drift shows up before a user reports it.A measured baseline, a review interval, and a threshold that triggers action.

Vendors, Access and Exit

0/5

Vendor contracts state whether your data trains their models, how long it is retained, and who their sub-processors are.The terms are in the contract, not only on a marketing page.

Vendors must notify you of a breach or a material model change, within a period you agreed.A named notification route and a stated window.

You could leave a vendor: your data, prompts and configuration export in a usable format, and you know what would break.A test export, or a written account of what would need rebuilding.

AI systems use role-based access and are covered by your normal access reviews and leaver process.A leaver loses AI tool access on the same day as email.

Someone is responsible for tracking the AI rules that apply to your sector and jurisdiction, on a set review cycle.A named owner and a date the last review happened.