Read from the guidance
Searching for the NIST AI risk management framework? The UK quietly published its own
Every month, several hundred people in this country search for the NIST AI Risk Management Framework — an American instrument, from an American agency. On 8 September the UK government published one of its own, free, and as far as we can tell nobody noticed. It is not aimed at your business, and that turns out to be the most useful thing about it.
The short version
- The Department for Science, Innovation and Technology published an AI Risk Management Toolkit on 8 September 2026 — guidance plus a 280 KB Excel workbook, free, Open Government Licence.
- It is written for government departments, not for you. A private company has no duty here at all.
- Nine risk categories, most of which nobody thinks about when buying an AI tool.
- Impact and likelihood scored 1 to 5, with real probability bands rather than adjectives.
- Four treatment options, one of which is simply accepting the risk — named and legitimate.
- ⚠️ Two items on its compliance list do not apply to a private business. Skip those.
What it is, and who it is for
The toolkit exists, in its own words, “to support anyone involved in the design, operation, procurement and delivery of products empowered by or enabled by AI.” That sounds like everyone. Read on a little further and it is not.
It was built so that “the UK public sector” can realise benefits from AI. It was “created to assist departments”. And it “aligns with the Risk Management toolkit (RMF) found in the Orange book, which establishes risk management principles and processes for government departments and other UK public sector organisations.”
So: no duty, no obligation, nothing to comply with. If you run a plumbing firm or a recruitment agency or a dental practice, this document is not addressed to you and never will be.
Which is exactly why it is worth an afternoon. You are being handed a government-grade set of questions with none of the bureaucracy attached — no board sign-off, no evidence pack, no audit. Take the questions, leave the machinery.
The nine categories, and why they are the useful bit
Appendix 1 sorts its questions into nine risk categories:
| Financial | Legal and regulatory compliance | Appropriate transparency and explainability |
| Fairness | Accountability and governance | Contestability and redress |
| Technical robustness | Security | Environment and People |
Now think about the last time anyone in your business assessed an AI tool. The conversation was almost certainly about two things: is it accurate and what does it cost. Those are half of one category out of nine.
Contestability and redress is the one worth staring at. It asks, in effect: when this thing gets something wrong about a customer, how does that customer argue with it, and who can overturn it? Most small businesses adopting AI have no answer, because the question has never been put to them in those words.
Two questions from the financial category, quoted as written, give the flavour — and note that neither is about the purchase price:
“Is there a need for additional resources to fix and maintain the AI solution, and have the financial implications of this been considered?”
“How will a regular review of the legal and regulatory landscape be conducted, noting that it is currently evolving?”
The second one is quietly the hardest question in the document, because the honest answer for most businesses is “it will not be, until something goes wrong”.
The scoring, which is better than it needs to be
Impact and likelihood are each scored 1 to 5, and “The risk score is the product of the likelihood and impact scores.” Ordinary enough. What lifts it is that the likelihood scale comes with actual numbers attached rather than adjectives:
| Level | Description | Probability |
|---|---|---|
| 1 — Rare | Highly unlikely to occur | < 5% |
| 2 — Unlikely | Unlikely but possible | 5% - 20% |
| 3 — Possible | Possible to occur at some time | 20% - 50% |
| 4 — Likely | Likely to occur in many circumstances | 50% - 80% |
| 5 — Almost Certain | Almost certain to occur | > 80% |
(The percentage ranges are reproduced exactly as the guidance writes them, spacing and all.)
Attaching a number to “possible” is the whole trick. In an ordinary meeting, “that’s possible” and “that’s unlikely” mean whatever the most confident person in the room wants them to mean. Saying “somewhere between one in five and one in two” ends that argument, and often ends it in a direction nobody expected.
And then you are allowed to do nothing
The four treatment options are avoidance, limiting, transference and acceptance. Avoidance is “usually suitable when both the potential impact and cost of mitigating are too high”. Transference is “where all or a portion of risk liability and responsibility are shifted to the third-parties best placed to manage it” — insurance, or a supplier contract, in plain terms.
But the one worth pointing at is acceptance: “where risks are acknowledged, and potential impacts are accepted without taking further actions to mitigate or eliminate them”.
A government document explicitly naming acceptance — in plain words, deciding something is fine and moving on — as a legitimate, recorded outcome is genuinely useful to a small business, because the alternative most people default to is not mitigation. It is worrying vaguely and doing nothing while pretending otherwise. A written-down, dated decision to accept a risk is a completely different thing from having failed to think about it, even though they look identical from outside.
The parts to leave behind
Appendix 1’s legal checklist tells you to consider, among others, the Public Sector Equality Duty and the Social Value Act. Neither binds a private company. Work through the list as written and you will be evidencing duties you do not have, which is a tiring way to spend a Tuesday.
What does reach you from that same list: UK GDPR, the Equality Act, and intellectual property. Whether the EU AI Act reaches you is a separate question with a surprisingly crisp answer, and we have written about the single clause that decides it.
Likewise the governance apparatus — departmental boards, organisational risk appetite signed off at board level. The toolkit itself concedes that “Defining your risk appetite is considered to be the most challenging aspect of any risk toolkit”. For a business of nine people, the honest version of a risk appetite is one page and one person’s signature, which is roughly what we argued when we wrote about fitting an AI policy onto a single sheet.
Why you were searching for NIST
Because the American instrument got there first and has had years of people writing about it. That is the whole reason, and it is the same pattern as businesses googling whether their website is “ADA compliant” — an American statute — when the duty that actually binds them is the Equality Act 2010.
There is nothing wrong with reading NIST’s framework. It is thorough and it is free too. But if you are going to borrow a risk method from a government, borrowing one written in your own regulatory language, referencing the statutes that actually apply here, costs you nothing and confuses you less.
What argues the other way
It is not a standard. You cannot certify against it, a client cannot ask you to prove you follow it, and it carries no weight with a regulator. If you want something with letters after its name, this is not it.
It is heavy in places. Written for organisations with risk registers, boards and assurance functions. A nine-person company adopting one AI tool does not need the workbook, the scoring matrix and the appetite statement. It needs about forty minutes with Appendix 1.
And a scored risk register can become theatre. Numbers make a judgement look like a measurement. A 4×3 is still two guesses multiplied together, and writing 12 in a box does not make it more true — it just makes it harder to argue with. Use the scale to have the argument, not to end it.
Sources. Everything here is from the AI Risk Management Toolkit published by the Department for Science, Innovation and Technology on 8 September 2026, read directly at gov.uk rather than from any summary. Both the publication page and the full guidance were fetched raw on 13 September, again on 15 September, and again on the morning this was published, 16 September; every quoted fragment was re-asserted against the raw text each time — 26 of them on the final pass, with no misses — and the readable text came back identical to the character on all three occasions (80,546 characters of guidance, 5,086 of the publication page). If you want to check us, the pages are free and the quotations are searchable. Quotations preserve the guidance’s own spacing and punctuation, including the percentage ranges. This is a description of a free government document, not legal or risk advice.