Scoring risk is the easy part. The toolkit's real value is the five levels of appetite it sets out for each of nine risk categories. Write down how much risk you'll accept in each one, keep the severe risks on a separate list, and name someone to own them.
On 8 September the Department for Science, Innovation and Technology published its AI Risk Management Toolkit. It’s written for public sector teams, but it’s one of the more practical things any organisation could read if it’s trying to use AI without either freezing or winging it.
For me, the most useful sentence in it is about appetite: “Defining your risk appetite is considered to be the most challenging aspect of any risk toolkit, especially in AI’s evolving field.”
It’s also the part I’d put money on most organisations skipping.
What's in it
The toolkit gives teams a common way to spot, score and deal with AI risk across the whole life of a tool, from picking the use case to switching it off.
It sorts risk into nine categories: financial, legal and regulatory compliance, transparency and explainability, fairness, accountability and governance, contestability and redress, technical robustness, security, and the effect on the environment and people. Each comes with a set of questions to work through.
Likelihood is scored from one to five, with percentage bands attached. “Rare” is under 5%. “Almost certain” is over 80%. Impact is scored the same way, and the two are multiplied to give a risk score. Once a risk is scored, you choose from four responses: avoid it, limit it, transfer some of it to a third party, or accept it.
It also says who should be in the room. Alongside the data scientists and security people, it lists senior leaders, legal, people who know the business, and end users, including those affected indirectly. Most private sector AI reviews I’ve seen stop at IT and legal. And it’s clear that this isn’t a one-off exercise. Risks should be looked at again after model updates and at each stage of release.
None of that is new. It’s sensible, joined-up and written in reasonably plain English, which is more than can be said for plenty of AI guidance.
Why appetite matters more than the scoring
The toolkit is blunt about what happens without a clear appetite. Teams “will invariably encounter a number of challenges and blockers when trying to manage risks for new AI tooling and solutions, which will eventually stifle innovation.”
That’s a government department saying that vague caution stops useful work. I’ve watched it happen. Nobody says no. Everybody asks another question. Months pass, and the people who wanted to help end up using the free version of something on their own laptop, which is a far bigger risk than the tool being debated. Blanket caution isn’t governance. It’s just delay with better vocabulary.
The toolkit offers five levels of appetite: averse, minimal, cautious, open and eager. It describes what each looks like for every one of the nine categories. That’s the useful bit. You don’t have one appetite for “AI”. You might be averse to fairness risk in anything that touches recruitment or credit decisions, and quite open to technical risk in a tool that summarises internal documents. Writing that down is what lets a team get on with the low-risk work without asking permission every time.
One thing to be careful with
Multiplying likelihood by impact is standard practice, and it’s fine as a way to sort a long list. But multiply two estimates together and you get something that looks like a fact. A rare but catastrophic risk can end up with the same score as a likely but trivial one, and slip quietly into the middle of the register.
The toolkit itself suggests combining numbers with judgement. I’d go further. Keep a separate short list of anything with a severe impact, whatever its score, and make sure someone senior has read it.
What to do with it this week
Pick one AI tool you’re using or planning. For each of the nine categories, write one sentence on how much risk you’re prepared to accept and why. Score the risks you can see, but pull the severe ones out on their own. Name an owner for each, and agree what will trigger a review: a model update, a move from pilot to live, or a change in who uses it.
Then ask the question the toolkit asks under contestability and redress. If this tool gets something wrong about a customer or a colleague, how would they know, and how would they challenge it? If there’s no good answer, that’s your first job.
The toolkit won’t make the decisions for you. It does give you a sensible structure and, usefully, official backing for the idea that being clear about risk is what lets you move.
Read the toolkit and not sure where to start?
Tell me which AI tool you’re trying to make a decision about, and we’ll work out what your risk appetite needs to say.