The government’s AI Risk Management Toolkit: what it asks of public bodies and their suppliers

What happened

On 8 September 2026, the Department for Science, Innovation and Technology published the AI Risk Management Toolkit on GOV.UK, with an accompanying spreadsheet workbook.¹ The toolkit offers “guidance to help anyone involved in AI projects understand, assess and manage risks,” and describes itself as a “starting point for implementing good risk management practices.”¹

It is written for government departments and public sector bodies, builds on the government’s existing risk management guidance, the Orange Book, and applies across the life of an AI system, “from use case identification to retirement.”²

What it contains

The toolkit organises AI risk into nine categories:²

1. Financial, including the costs of running and fixing AI.
2. Legal and regulatory compliance, such as data protection and equality law.
3. Transparency and explainability.
4. Fairness, including biased or discriminatory outcomes.
5. Accountability and governance.
6. Contestability and redress, meaning whether people can challenge outcomes and obtain a remedy.
7. Technical robustness.
8. Security.
9. Risks to people and the environment.

It then sets out a process: agree a risk appetite for each category, signed off by the departmental board; identify risks continuously with a multi-disciplinary team; score each on likelihood and impact from one to five; choose a treatment (avoid, limit, transfer or accept); and monitor throughout the system’s life. Teams are asked to keep a central log of AI risks and share it with GDS and DSIT.²

Why it matters

The toolkit is guidance, not a requirement, and uses advisory language throughout.² Even so, it matters for three reasons.

It creates a common language. Nine named categories and a five-point scale give boards, auditors and project teams a shared way to discuss AI risk, and make gaps easier to spot.

Fairness and redress sit alongside cost and security. Several of the best-known public sector failures, from the 2020 exam grading algorithm to the Dutch childcare benefits scandal, were failures of fairness and contestability rather than of technology. The toolkit puts those risks on the same page as budget and cyber risk.

It will shape what suppliers are asked. Suppliers selling AI to the public sector should expect buyers to ask how their products perform against these nine categories, and for evidence.

What to do now

1. Public bodies: compare your current AI risk register with the nine categories, and note any you are not tracking. Contestability and redress are a common gap.
2. Boards and audit committees: ask whether a risk appetite has been agreed for AI, category by category, and who signed it off.
3. Suppliers: prepare a short, evidenced account of how your product addresses each category, before you are asked.

Questions for your leadership team

1. For our most significant AI system, which of the nine categories have we assessed, and which have we not?
2. Have we agreed how much fairness or contestability risk we are willing to accept, or have we only agreed financial and security risk?
3. If a person affected by one of our AI decisions wanted to challenge it, what would happen, and how long would it take?

This briefing summarises the sources listed above, with analysis by the editor. Read the original sources before acting on it.