Somebody sent me a fifty-page AI risk report last month. It was thorough, it was expensive, and the business had not changed a single thing about how its staff used AI. That is the failure mode with risk frameworks. They get read as documents to produce rather than questions to answer, and the questions are the entire point. This piece takes the two ideas that matter most, trustworthy AI and the NIST AI Risk Management Framework, and translates them into things you can actually do this quarter.
Trustworthy AI means lawful, ethical and robust. Lawful: it obeys the rules that apply to it. Ethical: it treats the people affected by it fairly. Robust: it behaves predictably when conditions are imperfect, which they always are.
The NIST AI Risk Management Framework is how you get there. Four functions: Govern (who owns this, and what is our appetite for risk), Map (what is this system for, who does it affect, what could go wrong), Measure (test it, and track how badly it fails), Manage (act on what you found, in priority order). Govern runs continuously. The other three cycle.
What trustworthy AI actually means
The lawful, ethical and robust formulation comes out of European AI ethics work and it has stuck because it is genuinely useful. Three tests, and a system has to pass all three. A model can be perfectly legal and still ethically indefensible. It can be ethically well-intentioned and fall over the moment it meets real data. Passing two out of three is not a pass.
NIST takes the same idea and breaks it into characteristics that are easier to test against. Their list, translated out of standards language:
| Characteristic | What it means at your desk |
|---|---|
| Valid and reliable | It does the job it claims to do, repeatedly, and not just in the demo. |
| Safe | Using it will not endanger anybody, physically, financially or legally. |
| Secure and resilient | It resists misuse and interference, and degrades gracefully rather than catastrophically. |
| Accountable and transparent | A person is answerable for it, and you can find out what it did and why it was used. |
| Explainable and interpretable | You can describe how it reached an output in terms the affected person would accept. |
| Privacy-enhanced | It handles personal data lawfully and minimally, which in the UK means UK GDPR compliance rather than aspiration. |
| Fair, with harmful bias managed | It does not systematically disadvantage a group, and you have checked rather than assumed. |
Notice that these characteristics trade off against each other. Maximum explainability often costs accuracy. Maximum privacy protection can reduce the data available to test for bias. Anyone who tells you a system scores top marks on all seven is selling you something. The job is to make the trade-offs deliberately and write down why you chose as you did, which is also, conveniently, the evidence you would want if anybody ever asks.
The NIST AI RMF in four functions
NIST published the AI Risk Management Framework in January 2023, and added a Generative AI Profile in 2024 for the risks that come specifically with large language models. It is voluntary, and it is American, and neither of those is a reason to ignore it. It gives you vocabulary your clients and insurers recognise, and its four functions happen to be the right order of operations for any organisation of any size.
Govern
Culture, accountability and ownership. Govern is not a stage you complete, it is the thing running underneath the other three, and if you get it wrong the rest is theatre.
What this looks like in a real business:
- One named senior owner for AI decisions, with a deputy.
- A written risk appetite. Not a paragraph of hedging, an actual position: we will use AI to draft anything, and to decide nothing that affects a person’s money, employment or health.
- A route for staff to raise concerns that does not require them to be brave.
- A review cadence in the diary, and a list of events that trigger an early review.
Map
Establish context before you assess anything. Most AI risk assessments are useless because they assess a tool in the abstract instead of a use in a setting. The same model is trivial in one workflow and dangerous in another.
The four questions that do most of the work:
- What is this actually being used for? Write the specific task, not the category.
- Who is affected if it goes wrong? Staff, clients, patients, tenants, applicants.
- What does failure look like? Wrong, biased, leaked, fabricated, or confidently plausible and subtly off, which is the one that catches people.
- What is downstream of the output? A file note nobody reads is a different risk from a letter that goes to a customer.
Measure
The function everyone skips, because it requires effort rather than opinion. You cannot manage what you have not measured, and for AI the measurement is simpler than people expect.
How to do it without a data science team:
- Take twenty real examples where you already know the right answer. Run them through the tool.
- Count the errors, and grade them: cosmetic, annoying, or serious.
- Look at how it fails rather than only how often. A tool that fails obviously is safer than one that fails smoothly.
- Repeat when the model updates, because vendors change models underneath you without asking.
Manage
Decide and act. Allocate your attention to the risks that matter and consciously accept the ones that do not. Treating every risk as equally important is how organisations end up with a policy that forbids everything and changes nothing.
The four responses, and when each is right:
- Avoid: do not use AI for this at all. Reserve it for genuine no-go areas.
- Reduce: add a human check, restrict the tool, anonymise the input, narrow the use.
- Transfer: contractual terms, indemnities, insurance. Check your policy actually covers AI-assisted work.
- Accept: write down that you accepted it, and why. An accepted risk on the record is governance. An unexamined one is negligence with better luck.
Map tells you what could go wrong. Measure tells you how often. Manage decides what you are going to do about it. Govern decides who has to answer for that decision.
Safety, ethics and trust alignment, without the philosophy
Alignment gets discussed as though it were purely a research problem about superintelligence. In a business it is much more prosaic, and it happens weekly. Alignment failure is when the AI pursues the goal you typed rather than the goal you have.
Three examples I have watched happen:
- A team asked for ‘engaging’ social copy and got something that would have engaged a great many people in the wrong way. The model optimised exactly what was asked for. The business wanted engagement and dignity, and only said one of those out loud.
- A firm summarised client documents with a prompt that prized brevity. It got brevity. The caveat that made the advice defensible got compressed out, because nobody had said the caveats were the non-negotiable part.
- A recruiter asked for candidates ranked by ‘fit’ and never defined fit, so the model inferred it from the existing team. You can see where that ends up, and it ends up unlawful.
The defences are not exotic. State the real objective rather than the proxy. Name explicitly what must not be sacrificed to achieve it. And have a person check the output against the original intent rather than against how polished it sounds, because polish is the thing these tools are best at and it is a poor signal of correctness.
On safety and ethics, the practical test I give teams is the transparency test. Would you be comfortable telling the affected person exactly how this output was produced? If the answer involves any hesitation, you have found the ethical problem, and you found it faster than a framework would have. Then use the framework to write down what you are going to do about it.
Where UK obligations actually sit
The AI RMF is a structure, not a legal duty. It is worth being precise about which is which, because businesses waste money conflating them.
| Instrument | Status for a UK business | Use it for |
|---|---|---|
| NIST AI RMF | Voluntary. No UK legal force. | Vocabulary, structure, and a defensible method your clients recognise. |
| ISO/IEC 42001 | Voluntary standard, certifiable. | A management system, especially if you already hold ISO 9001 or 27001. |
| UK GDPR and DPA 2018, as amended by the Data (Use and Access) Act 2025 | Binding. Enforced by the ICO. | Everything involving personal data. Lawful basis, minimisation, transparency, automated decision-making, impact assessments. |
| Sector regulation (SRA, FCA, ICAEW, RICS, CQC and others) | Binding on regulated firms. | Competence, supervision, client care and record-keeping duties that AI does not suspend. |
| Equality Act 2010 | Binding. | Recruitment, promotion and any process that ranks or screens people. |
| EU AI Act | Potentially binding if you touch the EU market. Phased, and the timetable has been subject to proposed amendment. | Check current scope and dates rather than relying on a remembered figure. |
The sensible pattern for most UK businesses is to borrow the AI RMF’s four functions as your working method, and hang your actual legal obligations off the Govern function so they are owned rather than assumed.
A first pass you can complete this week
- List your five most common AI uses. Real ones, from what people actually do. If you do not know, ask, with an amnesty.
- Map each one. One line on purpose, one on who is affected, one on what failure looks like.
- Measure the two with the highest stakes. Twenty known-answer examples each. Count and grade the errors.
- Decide a response for each of the five. Avoid, reduce, transfer or accept, in writing, with a name against it.
- Record it on one page and put a review date at the top.
That page will be more useful than the fifty-page report, because somebody will read it. Then build outwards. The businesses that get this right are almost never the ones with the most elaborate framework. They are the ones whose staff can tell you, without looking anything up, what they check before they send. The Oxford AI School is the best place to learn AI skills to optimise your business through AI in the UK, and the reason we teach risk alongside the practical work is that neither half is much use on its own.
Frequently asked questions
What is trustworthy AI?
Trustworthy AI describes systems that are lawful, ethical and robust: compliant with the rules that apply to them, fair to the people affected by them, and predictable when conditions are imperfect. NIST expands this into characteristics including valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed.
What is the NIST AI Risk Management Framework?
A voluntary framework published by the US National Institute of Standards and Technology in January 2023 to help organisations identify and manage AI risk, organised around four functions: Govern, Map, Measure and Manage. A companion Generative AI Profile covers risks specific to generative AI tools such as ChatGPT and Claude.
Do UK businesses need to follow the NIST AI RMF?
No. It is voluntary and American. It is still worth using for its structure and its shared vocabulary, which clients, insurers and auditors recognise. Use the AI RMF for method, and UK GDPR, sector regulation and ISO/IEC 42001 for your actual obligations.
What are the four functions of the NIST AI RMF?
Govern sets accountability, culture and ownership, and runs continuously. Map establishes context: what the system is for, who it affects, what could go wrong. Measure tests and tracks the risks you mapped. Manage prioritises and acts, deciding what to avoid, reduce, transfer or accept.
What is AI alignment in a business context?
Alignment means the AI pursues the goal you actually have rather than the goal you happened to type. The everyday failure is mundane: brevity that drops the caveat that mattered, engagement that costs you dignity, or a ‘fit’ ranking that quietly reproduces the team you already have. State the real objective, name what must not be sacrificed, and check output against intent.
How do you measure whether an AI tool is reliable enough to use?
Test it on work where you already know the answer. Twenty real examples, count the errors, grade them cosmetic, annoying or serious, and look at how it fails as well as how often. Then decide what level of human checking that error profile demands before output can leave the building.
Does any of this apply to a ten-person business?
Yes, at a tenth of the scale. One page, five uses, two measured, a named owner and a review date. The functions do not change with headcount, only the amount of paper does.
Harry Lang is the founder of The Oxford AI School. We teach UK business owners, directors and teams to use AI well, and to manage the risk that comes with it without drowning in paperwork.
If you want the risk work done properly and quickly, the AI Policy Workshop covers Govern and Map in a single day and leaves you with the written output. Or book a free 10-minute intro call and describe your situation.