An AI risk register is a living list of the ways AI tools could cause harm in your business, with a named owner, a rating and a control for each risk. Build it from the tools your people actually use, score each risk for likelihood and impact, assign one accountable person per line, and review it on a fixed schedule. This guide gives you the fields, example risks and a review rhythm that suit a small or medium Australian business.
CalibreforumSee How AI Is Used Across Your BusinessCalibre Forum’s AI readiness assessment gathers employee insights and produces a scored report with a clear 90-day action plan.Visit Calibre ForumTL;DR
- List every AI tool in use first, including the ones staff adopted on their own. A register of risks for tools you do not know about is empty.
- Each risk needs eight things: ID, name, description, impact, controls, likelihood, residual risk and a named owner.
- Scale the effort to the risk. Light checks for low-risk tools, deeper assessment for anything touching client data or decisions.
- Review quarterly for the first twelve months after a new deployment, then at least annually.
- Link the register to your existing risk, privacy and security processes instead of running it as a separate project.
Table of Contents
- Why an AI risk register is worth the effort
- What belongs in an AI risk register
- Build the register in six steps
- Where the register meets Australian law
- Choosing a template that fits
- Keeping it current
- Our perspective
- From survey findings to register lines
- FAQ
- Sources
Why an AI risk register is worth the effort
The National AI Centre is plain that AI can increase existing risks or introduce new ones: incorrect outputs, bias, privacy, security, legal and safety problems. Which of these matter to you depends on how the tool is used, who it affects and what oversight sits around it.
The same page notes that risk rises when AI deals directly with the public, influences important decisions, runs autonomously or handles sensitive information. A receptionist using a chatbot to tidy meeting notes is a different proposition from a clinic using a transcription tool on patient consultations.
The National AI Centre’s implementation guidance adds that AI risks shift with system complexity and use case, and that small issues can amplify quickly. It also asks organisations to run risk management that is fit for purpose and to decide what level of risk is acceptable. A register is the simplest record of those decisions.
What belongs in an AI risk register
A register has two halves. The first is an inventory: what tools exist, who uses them, what data goes in. The second is the risk list itself. Skip the inventory and the risk list will describe an imagined business rather than yours.
The fields
A widely used Australian AI risk register template records Risk ID, risk name, description, potential impact, example control measures, likelihood, residual risk and risk owner. That is a sound minimum. Here is how we would fill each one in:
| Field | What to write | Example |
|---|---|---|
| Risk ID | A short unique code | AI-007 |
| Risk name | Five words or fewer | Client data pasted into public chatbot |
| Description | What happens, with which tool and which data | Staff paste client emails into a free chatbot to summarise them |
| Potential impact | Who is harmed and how | Privacy breach, loss of client trust |
| Controls | What reduces the risk today | Approved tool list, staff guidance, blocked consumer accounts |
| Likelihood | Rating on your chosen scale | Likely |
| Residual risk | Rating after controls | Medium |
| Owner | One named person | Operations manager |
The GSDC Council’s register user guide recommends dropdowns for risk category and for likelihood and impact ratings. In a spreadsheet, that means data validation on those columns. It keeps ratings consistent when three different people edit the sheet.
Example risks to start from
The National AI Centre’s list of common AI risks covers accuracy and reliability errors, bias and fairness, privacy and data handling, security vulnerabilities, legal and compliance breaches, safety concerns in critical fields and reputational damage. The Australian template above adds vendor dependency, misinformation and workforce impacts.
For most small and medium businesses, these are the lines we would expect to see first:
- Sensitive data entered into unapproved tools. The most common one, and the hardest to see without asking staff. Our guide to AI data leakage covers how it happens.
- Wrong or invented output used without checking. A draft contract clause, a figure in a quote or a medical note that nobody verified.
- Supplier risk. The tool changes its terms, goes offline or is acquired. See our one-page AI vendor risk checklist.
- Bias in decisions about people. Screening applicants, ranking customers or scoring staff.
- Unclear ownership. Nobody is accountable for a tool because everybody uses it.
- Security gaps. Weak account controls, unreviewed integrations and prompt manipulation. Our AI security checklist goes through them.
- Reputational damage. An AI-generated message to a customer that is wrong, tone-deaf or clearly automated.
Operational and agentic risks deserve their own rows once tools start acting on their own: connecting to email, calendars or customer systems. The same goes for marketing teams who now write content with AI and track how AI search engines cite their brand, a field covered in this piece on generative engine optimization. Each new capability is a new line.
Build the register in six steps
1. Take inventory of tools and use cases
Record the tool, what it does, what data it handles, who is affected and who looks after it internally. The Digital Transformation Hub argues for exactly this kind of practical documentation rather than a theoretical exercise.
The awkward part is that IT rarely has the full list. Staff use personal accounts, browser extensions and AI features built into software you already pay for. Asking people, anonymously, works better than scanning for tools. See our guide to finding shadow AI.
2. Tier the tools by risk
The same Hub guide suggests a tiered approach: basic checks for low-risk tools, more rigorous assessment where tools handle client data or influence service delivery, and full safeguards for tools that serve vulnerable people, including human oversight, escalation pathways and independent review.
A practical way to tier: ask four questions of each tool. Does it touch personal or confidential information? Does it talk to customers? Does it influence a decision about a person? Does it act without a human approving each step? Zero yeses is a light-touch tool. Two or more is a full assessment.
3. Write the risks in plain language
A risk is an event with a consequence, not a topic. “Privacy” is a topic. “Client health details are pasted into a public chatbot and stored by the provider” is a risk. MITRITY’s register documentation points out that using a shared vocabulary, such as bias, robustness and safety, lets an initial assessment feed the register directly.
4. Rate likelihood and impact
Keep the scale small. Three or five levels each for likelihood and impact are enough. Match the scale to the one your business already uses for other risks so the AI lines can sit in the same report. When assessing, the Australian template suggests weighing business context, legal obligations, data sensitivity, human impact, technology reliability and reputational trust.
Rate twice: once as things stand (inherent), then again after controls (residual). The gap between them shows whether your controls are doing anything.
5. Name one owner per risk
The Digital Transformation Hub says every risk needs a named internal owner with the authority and responsibility to act. “IT” or “management” is not a name. Pick the person who can actually change the process, which for a data-handling risk is often the operations lead rather than the IT contractor.
6. Decide, and record the decision
For each risk you either reduce it, transfer it, avoid it or accept it. If you accept one, write down why. MITRITY recommends documenting the reasoning for accepted risks, and it protects you later when someone asks why a known risk stayed open.
Where the register meets Australian law
There is no AI-specific statute that tells a small business what its register must contain. The Department of Industry, Science and Resources lists general laws that may apply to AI risks: directors’ duties, privacy law, the Security of Critical Infrastructure Act, Australian Consumer Law and anti-discrimination law. A register is a sensible place to note which of these each risk touches.
Privacy is the one most businesses meet first. The OAIC’s guidance on commercially available AI products says privacy obligations apply to personal information put into an AI system and to AI output that contains personal information. It also expects due diligence before adopting a product, including how human oversight will be built in. Each of those points translates into a register line. For specific legal questions, speak to a lawyer or your privacy officer.
Choosing a template that fits
You will find spreadsheets, databases and builders. They suit different jobs:
| Type | Good for | Watch out for |
|---|---|---|
| Simple spreadsheet with the eight fields | Small teams starting out | Needs someone to keep it alive |
| Large risk databases | Brainstorming risks you have not thought of | Too much to use as a working register |
| Enterprise GRC platforms | Larger organisations with existing risk systems | Cost and setup time for a small business |
The MIT AI Risk Repository is a good example of the second type: a living database of more than 1,700 risks drawn from 65 frameworks, and the database itself is free to copy and use. Use it as a prompt list. Do not paste 1,700 rows into your register.
Whatever you pick, re-score the risks for your own environment. A rating written for a bank does not describe a four-person physiotherapy clinic.
Keeping it current
A register that is filled in once and filed away is worse than none, because it suggests control that no longer exists. The Digital Transformation Hub recommends revisiting it quarterly for the first twelve months after a new AI deployment, and at least annually after that. Each review asks three things: are the listed risks still current, have new ones appeared, and are the controls working?
Two habits help. Set a review date on every line, not just on the document. And add a trigger: any new AI tool, any new use of an existing tool, or any incident means a register update within the week.
Our perspective
We see the same pattern in Australian and New Zealand businesses. The register is usually not the hard part. The hard part is the first column: nobody knows which AI tools are really in use, with what data, doing which jobs. Leaders often guess low.
That is why our assessment starts with the people. We survey every staff member anonymously, so the answers are honest, and combine them into a scored report with a 90-day action plan. We install nothing and collect no individual data. What comes out is an organisation-wide map of tools and workflows, which is the raw material for a good register.
Our advice: build a small register in a week rather than a perfect one in a quarter. Ten real risks with owners beat a hundred generic ones with none.
— Anthony
From survey findings to register lines
If you want the inventory done for you, our AI readiness report costs A$1,990 as a one-off. It turns anonymous staff responses into a scored view of capabilities and risks, plus a 90-day plan. You can see what is in the report or open the sample report first. If you want the picture refreshed every quarter, the membership is A$2,388 per year and covers recurring assessments, which fits the review rhythm above.
Sector matters too. We have pages for medical clinics, aged care and NDIS providers and managed service providers, where the risk lines look quite different.
FAQ
Does AI have risks for a small business?
Yes. The National AI Centre lists accuracy errors, bias, privacy, security, legal and reputational risks, and says they depend on how the tool is used. Small businesses are exposed mainly through staff using tools with client data, often without a policy.
Which items normally appear in an AI risk register?
Each entry usually has an ID, name, description, impact, controls, likelihood, residual risk and owner. Typical risks include data privacy breaches, biased outcomes, incorrect output, vendor dependency and reputational harm. Add your own tools and workflows to the list.
How often should we update it?
Quarterly for the first twelve months after deploying a new AI tool, then at least once a year. Also update it whenever you adopt a new tool, change how an existing one is used or have an incident.
Who should own the register?
One person should maintain the document, but each risk needs its own named owner who can act on it. Ownership should sit with the person running the affected process, not only with IT.
Sources
- Know the risks, National AI Centre
- Guidance for AI adoption: implementation guidance, National AI Centre
- AI risk register template
- AI inventory and risk register user guide, GSDC
- Managing risk: how to prepare your organisation, Digital Transformation Hub
- AI risk register documentation, MITRITY
- The legal landscape, Department of Industry, Science and Resources
- Guidance on privacy and commercially available AI products, OAIC
- MIT AI Risk Repository
- MIT AI Risk Repository database
- Generative engine optimization guide
Recommended
- AI Vendor Risk: A One Page Checklist for Business Teams
- AI Model Risk Management for Australian Businesses
General information only. Articles on this blog are written with the help of AI writing tools and published by Calibre Forum. They are not legal, insurance, tax or financial advice, so check anything important with a qualified adviser.