Applying the DSIT AI risk toolkit in a business of 15 to 200 people
Chris Duffy
Certified Chief AI Officer • • 5 Min Read
The DSIT AI Risk Management Toolkit is non-statutory guidance that the Department for Science, Innovation and Technology published on 8 September 2026 for government departments. No private business is required to follow it. A UK business of 15 to 200 people can still borrow its method: nine risk categories, a five-level risk appetite, and likelihood times impact scoring.
What the toolkit is, and who it was written for
The Department for Science, Innovation and Technology (DSIT) published the AI Risk Management Toolkit on 8 September 2026, under the Open Government Licence v3.0. It is the UK government's own method for finding, scoring and treating AI risk. It is built on the Orange Book and is designed to sit alongside the NCSC Cyber Assessment Framework.
It is guidance written for government departments. It places no legal duty on a private company, and following it does not discharge any UK GDPR, equality or sector-regulator duty. The guidance contains no statement that it applies to the private sector, so treat what follows as a method a small business can borrow, with no DSIT recommendation behind it for SMEs.
The publication page carries three attachments: the guidance itself as a web page, and a workbook in xlsx and ods formats. The guidance also mentions an AI Risk Monitoring Dashboard. This guide covers the guidance only, not the workbook.
This is general regulatory information, not legal advice.
When it is worth using in a small business
Three situations make it worth the effort:
- You supply the public sector, and the buyer asks how you manage AI risk.
- You want a risk method that a UK board, auditor or insurer will recognise without explanation.
- You need a scored AI risk register as evidence that your AI governance works in practice.
The nine risk categories
Six categories carry the UK's five cross-sector AI principles almost word for word, with safety, security and robustness split into two. DSIT added three business risks that any board already recognises.
- Appropriate transparency and explainability
- Fairness
- Accountability and governance
- Contestability and redress
- Technical robustness
- Security
- Financial (added by DSIT)
- Legal and regulatory compliance (added by DSIT)
- Risks to people and the environment (added by DSIT)
Appendix 1 of the guidance gives identification questions for each category. They name the laws to check as a minimum, including equality law, UK GDPR and intellectual property, and applicable AI law such as the EU AI Act.
The five-level risk appetite scale
Appendix 2 sets five levels of risk appetite for each category: Averse, Minimal, Cautious, Open and Eager. Appetite is set for the whole organisation and signed off at the top. The guidance calls it the hardest part of any risk method, and says its examples are there to be adapted.
How the guidance scores and treats a risk
Likelihood is scored 1 to 5 and impact 1 to 5, and the risk score is the product of the two. Likelihood carries probability bands: Rare under 5%, Unlikely 5 to 20%, Possible 20 to 50%, Likely 50 to 80%, and Almost Certain over 80%. You may weight categories, and where one risk has several outcomes you weight them or take the highest. The guidance sets no High, Medium or Low bands.
Appendix 3 sets impact thresholds for a government department. A loss of £10,000 to £50,000 counts as Minor, and Catastrophic needs more than £1 million. Those figures are the government's own calibration, so they make most of a small business's real risks look small.
There are four ways to treat a risk: avoidance, limitation, transference and acceptance. Appendix 4 lists example treatments against named risks such as poor accuracy, bias, privacy, security and hallucination, and calls them illustrative. The guidance also points to one option teams skip: a clear plan for what you do when the risk happens.
Applying it with 15 to 200 people
The steps below are IgniteAI's own adaptation for a small business. The nine categories and the five appetite levels stay exactly as DSIT wrote them.
- Name your AI governance officer. The guidance asks for a named AI governance officer leading a risk team. In a firm of fifty, that is usually the managing director, the operations director or the AI lead, with three or four people covering the roles the guidance lists.
- Set appetite on one page. The managing director picks one appetite level for each of the nine categories and signs it off.
- Re-set the impact scale for your size. Set financial impact bands as a share of your annual turnover, so a loss that would hurt you scores as one.
- Keep one register. Give each risk its DSIT category, appetite level, likelihood, impact, score and treatment type, so you can answer a buyer who asks whether you follow the DSIT toolkit with the evidence open.
- Reassess on change. The guidance runs risk management from the first use case to retirement. Review a row after a model update and keep reviewing once the tool is live.
A worked risk register row
This row is IgniteAI's own adaptation for a small business. It is an illustration, and DSIT did not write it.
| Risk | Staff paste client personal data into an AI tool outside the approved business account. |
|---|---|
| DSIT category | Legal and regulatory compliance |
| Appetite (set by the MD) | Averse |
| Likelihood | 3 (Possible) |
| Impact | 4, on bands set as a share of turnover |
| Score | 12 (3 x 4) |
| Treatment type | Limitation |
| Treatment | One approved business account, a written rule on which data may go in, and a monthly check of use. A response plan names who acts if client data leaves the approved account. |
| Owner | The named AI governance officer |
| Review | After any model or tool change, and quarterly while live |
Where it sits beside ISO/IEC 42001 and the NIST AI RMF
The NIST AI RMF is a voluntary structure built on four functions: Govern, Map, Measure and Manage. ISO/IEC 42001 is the certifiable route. The DSIT method can sit underneath either as the practical risk register, and a UK buyer will recognise it as their own government's method. Our AI governance for UK SMEs work keeps the register in this format.
Source: AI Risk Management Toolkit, Department for Science, Innovation and Technology, GOV.UK, published 8 September 2026. Read the guidance on GOV.UK, or download the toolkit guidance as a PDF (36 pages).
Contains public sector information licensed under the Open Government Licence v3.0. © Crown copyright.
Read next: AI governance for UK SMEs · What a DPIA is and when you need one · Book a call