1. Pick work with an observable result
Choose a task that someone already understands well enough to review. Fixing a small defect, drafting a reply from approved material or preparing a change proposal are easier to assess than an open-ended request to improve a whole department.
Describe the result in ordinary language. Name who will review it, what a good result looks like and which actions remain outside the exercise. If you cannot tell whether the work succeeded, narrow the task before connecting more tools.
- Name the task and the person responsible for its result.
- Record the current way of doing it so you have a useful comparison.
- Choose a limited set of tools, repositories and people.
- Decide what must be reviewed before the result is used.
2. Map the information and actions
Write down the route from the person to the AI tool, model provider and any connected systems. Include files, retrieved material and tool outputs that may become part of a request. The exact route depends on the tool and its settings.
Keep two questions separate: what information may be sent, and what the tool may change. A tool allowed to read a project does not automatically need permission to publish a release.
| Question | What to record |
|---|---|
| Information | Allowed data, restricted data and who receives it |
| Actions | Operations the tool may propose or execute |
| Approval | Which person may approve which operation |
| Records | What to retain, who can read it and when to delete it |
3. Test the normal path and the boundaries
A successful example only tells you about that example. Include work that should be permitted, a request that should be refused, an action requiring approval and a connection that becomes unavailable. These are suggested engineering checks, not a certification programme.
| Test | Question to answer |
|---|---|
| Permitted work | Can the person complete the task with the agreed rules? |
| Restricted request | Is the intended information rule applied, and is the reason understandable? |
| Approval | Does the proposed operation wait for the right person? |
| Changed proposal | Is a changed operation checked again rather than covered by an earlier approval? |
| Interruption | Can the team identify what happened and what remains uncertain? |
4. Keep a short written scope
Use this worksheet with the delivery, security and business owners. A written scope makes it easier to spot disagreements before they become integration work.
Task and useful result:
People and reviewer:
AI tools and connection route:
Information allowed / restricted:
Actions allowed / approval required:
Records, access and retention:
Permitted and refused test cases:
Interruption and recovery checks:
Known gaps and unresolved questions:
Decision, owner and next review:5. Decide what to change or expand
Review the useful result and the controls separately. The task may work while a data route remains untested; the controls may work while the task produces little value. Record both findings.
Expand only the parts you have agreed to use. Keep unsupported routes, missing records and unresolved results visible. NIST’s AI risk-management resources provide broader guidance for organising this work; this worksheet is our practical starting point.
NIST AI Risk Management Framework
NIST generative AI risk profile
A checklist for human-reviewed platform decisions
For a digital labour platform, list the decisions an automated system takes or supports. Identify which need a human decision, who can change the outcome and what the person affected is told.
Our one-page working checklist helps structure that conversation. It is a review aid; check the applicable national law and the actual workflow with your legal and operational owners.