What is an AI risk register?
By Nick Jang, founder of MINERVA Studio ·
The short answer
An AI risk register is a list of the specific risks each AI system carries, with how likely and how serious each one is, who owns it, which controls are in place, what risk is left over and when it will be looked at again. It sits beside the AI inventory: the inventory says which AI you have; the register says what could go wrong with each and what’s being done about it.
Register versus inventory
People mix these up, and it costs them. An inventory has one row per AI system: what it is, who owns it, who supplies it, its risk rating. A register has one row per risk: the AI chatbot might give wrong product information; the same chatbot might leak personal data; the same chatbot might be talked into ignoring its instructions. Three rows, one system.
You need both. The inventory tells you where to look; the register tells you what you found and what you decided. At a bank, the inventory side is what OSFI’s E-23 calls the model inventory.
The columns
| Column | What goes in it |
|---|---|
| Risk ID | A stable reference, e.g. AIR-014. |
| AI system | The system from the inventory, by its ID. |
| Risk description | One sentence: cause, event, consequence. |
| Category | Accuracy, bias, privacy, security, transparency, third party, misuse, legal. |
| Likelihood and impact | Each on your institution’s scale, before controls. |
| Inherent rating | Likelihood times impact, or your matrix’s cell. |
| Controls | What reduces it: testing, human review, limits, monitoring, contract terms. |
| Residual rating | The rating after controls. |
| Owner | One named person, not a team. |
| Treatment | Accept, reduce, transfer or avoid, and who approved that choice. |
| Actions and due dates | What still has to be done, by when. |
| Last reviewed / next review | Dates, set by the residual rating. |
An example row
| Field | Example |
|---|---|
| Risk ID | AIR-014 |
| AI system | MDL-031, customer service chat assistant (vendor LLM) |
| Risk | The assistant states a wrong fee or rate, a customer acts on it and complains or is harmed. |
| Category | Accuracy, conduct |
| Inherent | Likely × moderate = high |
| Controls | Answers limited to approved product pages; fees and rates pulled from the product system, never generated; weekly sample review of 200 conversations; hand-off to a person on any pricing question. |
| Residual | Medium |
| Owner | Head of digital servicing |
| Treatment | Reduce; accepted at medium by the CRO on 14 Oct 2026 |
| Next review | 14 Jan 2027 |
Where to find the risks
You don’t have to invent the list. The MIT AI Risk Repository collects more than 1,700 AI risks drawn from 65 published frameworks, grouped into seven domains: discrimination and toxicity; privacy and security; misinformation; malicious actors and misuse; human-computer interaction; socioeconomic and environmental harms; and AI system safety, failures and limitations. NIST’s Generative AI Profile adds twelve risks specific to generative AI, and OWASP lists the top ten for AI agents.
Use them as a checklist against each system, then write the register in your own words about your own use. A copied generic risk with no owner isn’t a register; it’s a reading list.
Keeping it alive
Registers die in spreadsheets. Nobody updates the row after the review, actions slip past their dates, and the copy the board saw is three versions old. What keeps one alive: an owner per risk, review dates driven by the rating, and a link from each risk to its AI system and to the evidence for its controls.
MINERVA keeps each AI system’s risks, findings, owners and decisions on one record with the inventory, so the register and the inventory can’t drift apart. See AI governance software.
Related questions
What is the difference between an AI risk register and an AI inventory?
An inventory has one row per AI system: what it is, who owns it, its rating. A risk register has one row per risk: what could go wrong with a system, how bad it is, the controls, the owner and the review date.
What should an AI risk register include?
A risk ID, the AI system, a description, category, likelihood and impact, inherent rating, controls, residual rating, owner, treatment decision, actions with dates, and review dates.
What is the MIT AI Risk Repository?
A public database from MIT of more than 1,700 AI risks collected from 65 frameworks, grouped by cause and by domain. It is a useful checklist when building an AI risk register.
How often should an AI risk register be reviewed?
Each risk on a cycle set by its residual rating, and the whole register whenever a system changes materially, a new system arrives or an incident happens.
Who owns the AI risk register?
Each risk has a named owner, usually in the business that uses the system. The register as a whole is typically maintained by the risk function or the AI governance lead.
Sources
- MIT — AI Risk Repository
- NIST — AI Risk Management Framework: Generative AI Profile (NIST AI 600-1)
- OWASP GenAI Security Project — Top 10 for Agentic Applications (2026)
- ISO — ISO/IEC 42001:2023, AI management system
New answers are posted on MINERVA on LinkedIn. Follow along, or send the question you’d like answered next to minerva@minerva.studio.
A plain-language explanation, not legal advice. The regulator’s own text is the source; read it, and take advice on how it applies to your institution.