An AI risk register is the living list of the risks of an artificial intelligence system, with the likelihood and impact of each one, the existing controls, the risk left after them (residual risk), the owner and the agreed treatment. It shows where to act first, supports decisions by senior management and links AI risk management to the enterprise risk management of the organisation.
How to use
- Describe the system: source, whether it is generative, purpose, weight in the decision, personal data and people affected.
- Fill in the table with one risk per row, giving likelihood and impact from 1 to 5, the existing control, its effectiveness and the owner. With an empty table, the result suggests a starter register from the context.
- Read the top five priorities, the residual risk heat map, the warnings and the missing categories.
- Export as DOCX, CSV, PDF or Markdown and take the risks into the DPIA when personal data is involved.
What each field asks for
| Field | What to enter |
|---|---|
| AI system | The name the organisation knows the system by. |
| Source of the system | |
| Is it generative AI? | |
| Purpose and context | What the system does, who uses the output and for which decision. |
| Weight in the decision | How much the output decides about people. |
| Personal data | The most sensitive type of personal data the system receives or produces. |
| People affected | Who feels the effects of the output. Leave empty if nobody is affected. |
| Next review of the register | Date on which the register will be reviewed. |
| Category | |
| Risk description | E.g. a summary with an invented fact leads to the wrong priority |
| Likelihood (1 to 5) | 1 rare, 5 almost certain. |
| Impact (1 to 5) | 1 negligible, 5 severe for people or the organisation. |
| Existing control | E.g. a handler checks the summary against the documents |
| Control effectiveness | |
| Owner | E.g. claims manager |
Example
A fictional insurer records eight risks of its claims triage assistant, a vendor generative AI that recommends a review priority. Prompt injection through the free text of the claim notice (5 × 4 = 20, no control) is critical and goes to senior management; handlers accepting the suggestion without review (4 × 5, weak control, residual 15) is high and has no owner. Three risks are medium and three low, and the tool points to intellectual property and misuse as missing categories.
When to use
- Before putting an AI system into production or extending its use.
- When the AI inventory shows a system of medium or high risk.
- To take AI risks to the risk committee or senior management on a common scale.
- At the periodic review, after an incident or after a change of model or vendor.
Frequently asked questions
What should an AI risk register contain?
For each risk: category, description, likelihood, impact, existing control and its effectiveness, residual risk, treatment, owner and review date. The tool adds the heat map, the priorities and a link from each category to the NIST AI RMF, the EU AI Act and the GDPR.
How is residual risk calculated?
Inherent risk is likelihood times impact, from 1 to 25. A control lowers the likelihood by 0, 1, 2 or 3 levels depending on its effectiveness (none, weak, partial or strong), never below 1, and the impact does not change. Effectiveness set without a described control counts as none.
What are the bands and treatments?
From 1 to 4 the risk is low and may be accepted; from 5 to 9 it is medium and should be monitored; from 10 to 16 it is high and needs a mitigation plan; from 17 to 25 it is critical and should be escalated to senior management. Adjust the scale to your organisation’s risk appetite.
Where does the suggested starter register come from?
From the context you enter. Generative AI brings inaccuracy, intellectual property and misuse; a vendor brings third-party risks; personal data brings privacy; a recommendation or automated decision brings human oversight. The starting scores are estimates for you to adjust.
Does the register replace the DPIA or the FRIA?
No. The register organises all the risks of the system; the DPIA covers the risks to people when personal data is processed, and the FRIA, when the EU AI Act requires one, covers fundamental rights and may reuse the DPIA. The tool is educational; it is not legal advice or an audit.
Which DPCert certification covers this topic?
AI risk and impact assessment is a domain of RAIP Responsible AI Practitioner. Integrating AI risks into enterprise risk management is covered by RAIL Responsible AI Leader.