Core controls
- Data boundaries: which fields and systems the model can read.
- Least-privilege tools for any agent action.
- Human approval for sends, spend and customer-impacting changes.
- Logging of prompts, tool calls and outcomes.
- Evaluation sets for accuracy, bias and safety.
Risk questions before launch
- What is the worst plausible failure and who is affected?
- Could personal data leak into prompts or outputs?
- How will we detect drift or misuse?
- Who can switch the system off quickly?
Honest communication
Say plainly when content or decisions are AI-assisted, and do not claim human review unless it happened.
Frequently asked questions
Is a policy document enough?
No. Policies need technical enforcement, testing and named owners to matter.
Which framework should we use?
Start with a recognised framework such as NIST AI RMF and adapt it to your sector and regulators.
Last updated: