AI risk management: a framework and a worked risk register.
The four-step method, a simple way to score AI risk, and a worked AI risk register for a business on Claude, mapped to Australia's AI6, ISO 42001, ISO/IEC 23894 and the NIST AI RMF.
The short answer
AI risk management is four steps repeated for each use of AI: identify what could go wrong and who it would affect, assess how likely and how serious it is, treat it with a control and a named owner, and monitor and review it on a schedule. The output is an AI risk register, one line per use. Every major framework asks for the same thing: it is practice three of Australia's Guidance for AI Adoption, Clause 6.1 of ISO/IEC 42001, the subject of ISO/IEC 23894, and the Map, Measure and Manage functions of the NIST AI RMF.
AI risk management in four steps
The same four steps sit under every AI risk management framework. What makes them work in a small business is doing them per use, in plain words, with a person's name on each line.
Identify
List each use of AI, not each tool: drafting customer replies is one use, summarising board papers is another. For each, ask what could go wrong and who it would affect.
Assess
Score how likely the problem is and how much it would matter, each from one to three. Multiply them for the rating. Score what happens with the controls you have today, not the ones you plan.
Treat
Put a control against anything rated medium or high: a person reviews before it goes out, the AI gets read-only access, certain data stays out, or the use is not approved. Name an owner.
Monitor and review
Check the control is followed and still works. Review each line on a schedule set by its rating, and straight away when the tool, data, access or use changes, or after an incident.
How to score an AI risk assessment
Three levels are enough. Finer scales invite arguments about whether something is a 6 or a 7 and add nothing to the decision, which is only ever: is this fine, does it need a control, or should it wait.
Likelihood, 1 to 3
1 if it would take an unusual mistake, 2 if it could happen in a normal month, 3 if it is happening already or will without a control.
Impact, 1 to 3
1 if it is caught internally and costs a little time, 2 if a customer notices or it costs real money to fix, 3 if it harms a person, breaches privacy or reaches the board.
Rating = likelihood x impact
1 or 2 is low: note it and review yearly. 3 or 4 is medium: a control and an owner, reviewed quarterly. 6 or 9 is high: a control before it goes live, reviewed monthly.
A worked AI risk register for a business on Claude
Six lines a typical small or mid-sized business on Claude might have, including one agent with write access. It is an illustration of the method, not a client's register: your uses, scores and owners will differ.
| Use case | Risk | Likelihood | Impact | Rating | Control | Owner | Review |
|---|---|---|---|---|---|---|---|
| Drafting customer email replies in Claude | A wrong or made-up detail is sent to a customer | 2 | 2 | 4 Medium | Claude drafts only; a person reads and sends every reply | Customer service lead | Quarterly |
| Staff pasting customer details into AI | Personal information ends up in a personal account outside the business's control | 2 | 3 | 6 High | Company plan only; policy says what may go in; personal accounts not used for work | Operations manager | Monthly |
| Monthly finance commentary from Xero data | A misread figure reaches the board pack | 2 | 3 | 6 High | Read-only Xero connection; finance checks every figure against the ledger before it goes out | Finance manager | Monthly |
| An agent with write access to the CRM | The agent updates or deletes the wrong records in bulk | 1 | 3 | 3 Medium | Permissions scoped to named fields; deletes and bulk changes wait for approval; off switch tested | Agent owner | Quarterly, and on any access change |
| Marketing copy drafted with Claude | A claim the business cannot back up is published | 2 | 2 | 4 Medium | A person checks every claim against a source before anything is posted | Marketing lead | Quarterly |
| Summarising job applications | An applicant is treated unfairly and nobody can explain why | 1 | 3 | 3 Medium | AI summarises only; it does not rank or reject; a person makes and records every decision | Hiring manager | Quarterly |
Two things stand out in a register like this. The high ratings come from data and money, not from the AI being clever or not, so the controls are about access and review. And the agent line is rated lower than you might expect only because its control is strong: take the approval step away and it becomes the highest risk on the page. The job application line also matters in Australia, where privacy policy changes due in December 2026 require disclosure of automated decisions that significantly affect individuals.
How it maps to AI6, ISO 42001, ISO/IEC 23894 and the NIST AI RMF
One register, four frameworks. Build it once and it is the risk evidence each of them asks for.
Australia: Guidance for AI Adoption (AI6)
This is practice three, measure and manage risks. The register above is the evidence for it, and the review column feeds practice five, test and monitor. Our page on AI governance in Australia sets out all six practices.
ISO/IEC 42001
Clause 6.1 asks for an AI risk assessment, an AI risk treatment plan and an AI system impact assessment; Clause 8 asks you to carry them out in operation and repeat them when things change. The register, with owners and review dates, is the record an auditor samples.
ISO/IEC 23894
Guidance on AI risk management built on ISO 31000. It is not certifiable on its own; it shapes the method you use inside a 42001 management system. The four steps on this page follow its shape.
NIST AI RMF
Four functions: Govern, Map, Measure and Manage. Map is identify, Measure is assess, Manage is treat and monitor, and Govern is the accountability around them. It is voluntary and cannot be certified.
For the Australian practices in full, see AI governance in Australia. For the US framework against the certifiable standard, see ISO 42001 vs the NIST AI RMF. For the management system the register sits inside, see ISO 42001.
How often to review AI risks
Review a line whenever its use starts, whenever the tool, the model, the data or the access changes, and after any incident. Between those, let the rating set the schedule: high monthly, medium quarterly, low at least once a year. Put the dates in the calendar of the person named as owner, not in a document.
The reviews go faster when there is a record of what the AI actually did, so the owner is checking evidence rather than memory. That is also what an ISO 42001 auditor or an Australian customer asking about AI6 will want to see.
Where Sentry AI fits
We set the register up where the work happens
Sentry AI sets up the AI register, the policy and the decision trail on your own Claude or ChatGPT workspace, so the risk register stays current as people work. Start with the one-page AI policy template if you do not have a policy yet. We are not a certification body, and this page is general information, not legal advice.
Frequently asked questions
What is AI risk management?
AI risk management is the routine of finding what could go wrong with each use of AI in a business, judging how likely and how serious it is, putting a control in place where it matters, and checking on a schedule that the control still works. For most businesses the output is a short AI risk register with an owner and a review date on every line.
What is an AI risk management framework?
A published method for doing that routine consistently. The best known are the NIST AI Risk Management Framework, ISO/IEC 23894 (guidance on AI risk management) and the risk clauses of ISO/IEC 42001. In Australia, the Guidance for AI Adoption makes it practice three, measure and manage risks. They share the same core steps, so a register built under one serves the others.
How do you do an AI risk assessment?
Take one use of AI at a time, not the tool as a whole. Ask what could go wrong, who would be affected, how likely it is and how much it would matter. Score likelihood and impact from one to three, multiply them, and put a control and a named owner against anything that scores three or more. Write it down with a review date.
What is the NIST AI Risk Management Framework?
A voluntary framework from the US National Institute of Standards and Technology, organised around four functions: Govern, Map, Measure and Manage. It is widely named in tenders and cannot be certified. Map, Measure and Manage line up with identifying, assessing and treating risk; Govern is the accountability around them. Our comparison of ISO 42001 and the NIST AI RMF covers the differences in full.
What is ISO/IEC 23894?
ISO/IEC 23894:2023 is international guidance on AI risk management, built on the general risk standard ISO 31000. It is not certifiable on its own. Organisations normally use it to shape the risk process that sits inside an ISO/IEC 42001 management system.
What goes in an AI risk register?
One line per use of AI: the use case, the risk, the likelihood, the impact, the resulting rating, the control, the named owner and the next review date. Agents get their own lines, with what they can access and what they may do without a person. Keep it somewhere the business can see, not in one person's files.
How often should AI risks be reviewed?
Whenever a use starts, whenever the tool, model, data or access changes, and after any incident. Between those, on a fixed schedule set by the rating: high-rated lines monthly, medium quarterly and low at least once a year. A review that is on the calendar and actually happens is worth more than a long register nobody opens.
Start the register with what is actually in use
Every AI risk register starts with a list of the AI people use, and most businesses find more than they expected. The free AI Readiness Check is the first pass.
Ask an AI about this page
Opens the assistant with this page loaded: read it, summarise it, cite it.