What should an AI system register contain for an Australian organisation?
An AI system register should hold one entry for every use of AI, not every product, each recording its purpose, the system or model behind it, the one person accountable, the risk classification and its basis, whether it makes or informs decisions about individuals, the data it uses, where a human checks it, who approved it and when, the processes it sits inside and the next review date. The register exists so that what AI the organisation uses, why, and who answers for it can be read from the record rather than reconstructed from memory.
Written for risk, compliance and technology leaders who have to account for every AI use in their organisation. Reviewed by Mike Bentley, founder, .
One row per use, not per product
An AI system register is a list of every place the organisation relies on an AI capability, with enough recorded about each one that a person who was not involved can tell what it does, why, who answers for it and how it is controlled. The unit of the register is the use, not the product. A vendor's model that drafts customer replies in the contact centre and the same model that scores job applicants in recruitment are two entries, because everything that matters differs: the purpose, the people affected, the data flowing in, the person accountable and the point at which a human checks the output.
The obligations and frameworks that bear on a register are framed the same way. The test the Office of the Australian Information Commissioner applies to the automated decision transparency obligation attaches to a decision and the program arranged to make or inform it, not to a licence. The first subcategory of the Map function in the NIST AI Risk Management Framework asks for a system's intended purposes and the settings it will be deployed in to be understood and documented.
The Australian Government publishes an AI register template alongside its Guidance for AI Adoption, on industry.gov.au and ai.gov.au, and it is the right starting point for an organisation that wants the official shape. This article is about the fields that make a register useful beyond a list, whichever template you start from. The practical consequence of the one-row-per-use rule is that the register grows faster than the procurement list, and that the entries that matter most are often the ones nobody bought: a scoring formula in a spreadsheet, a feature switched on inside software the organisation already had, a chatbot a team started using on its own.
The fields that say what the use is and who answers for it
The first group of fields in an AI system register describes the AI use itself: what it does, what is behind it, who answers for it and where it acts.
A register that stops here is already useful. It answers the first question any review asks: what AI the organisation uses, and who is responsible for each use.
- Purpose: what the use does, in a sentence someone outside the team would understand, naming the output (a draft, a score, a ranking, a routing, a decision). "Improve efficiency" is not a purpose; "ranks inbound claims for assessor attention" is.
- System or model: the product, model or component behind the use, its vendor or internal team, and the version where known. When one component appears in several entries, this field ties them together.
- Owner: exactly one named person who answers for the use. Not a team, not a vendor, not "IT". Usually it is the person who owns the business process the use sits inside, because they can stop it, change it or explain it.
- Processes it sits inside: the named process steps where the use acts, linked to the process documentation where it exists. This field finds every AI use touching customers, or every one inside a critical operation, without another survey.
- Data it uses: the kinds of data that go in, with a plain flag for personal information and for sensitive categories such as health or biometric data. The OAIC's fact sheet says the kinds of personal information used must appear in the privacy policy where the transparency obligation applies, so this field is reused directly.
The fields that say how the use is controlled
The second group of fields in an AI system register records the governance position of each AI use. Each is a statement someone has made about the use, so each needs a date and a name beside it.
- Risk classification, with its basis: the level assigned, on whatever scale the organisation uses, and the assessment it came from, cited. "Not classified" is a legitimate value and more useful than a guess, because it says where work is still owed.
- Makes or informs decisions about individuals: yes or no, with what the decision is and whether the program decides it or informs a person who does. This is the field the APP 1.7 to 1.9 obligation turns on; the next section is about it.
- Human oversight point: where a person reviews, approves or can override the output, named by role and step. "Human in the loop" is not an oversight point; "the assessor reviews every ranked claim before contact" is.
- Approvals and dates: who approved the use, under which policy, on what date, with what conditions. Where no approval was ever given, the honest entry is "none recorded", which is itself a finding.
- Review date: when the entry is next due, set by risk and by what would change it (a model update, a vendor change, a new obligation, an incident).
- Status: proposed, in pilot, in production or retired, with the date each change took effect, so a retired use is recorded as retired rather than deleted.
- Evidence: a reference to where the assessment, the approval and the test results live, rather than a copy of them.
The flag that matters most from 10 December 2026
From 10 December 2026, according to the Office of the Australian Information Commissioner's APP 1 Guidelines, an organisation covered by the Australian Privacy Principles whose computer programs make, or substantially and directly inform, decisions that could significantly affect an individual's rights or interests must say in its privacy policy which kinds of personal information those programs use and which kinds of decisions they make. The related article on automated decision transparency covers the obligation; this section covers the register field it turns on.
Three points in the OAIC's fact sheet on the obligation, published on 30 September 2026 according to the OAIC's announcement, decide how the field should be designed. A computer program includes pre-programmed rule-based processes as well as machine learning and generative AI, and the fact sheet's first worked example is a spreadsheet formula that ranks people, so the flag applies to the whole register, not only to entries with "AI" on the box. A decision may be in scope even where the program's output does not replace the entire decision-making process or is subject to human review, so the flag cannot be cleared by writing "a person signs off". And the fact sheet counts procuring a third-party program and configuring off-the-shelf software among the ways an entity may have arranged for a program, so vendor-supplied uses carry the flag as in-house ones do.
Designed well, the field has three parts: whether the use makes or informs a decision about an individual, what the decision is in plain words, and whether the program decides alone or a person decides with it. The fact sheet also says an entity in doubt about whether a use is in scope should take a cautious approach and include the information, so record the doubt in the register rather than let it default to "no".
Why a compliance percentage misleads
An AI system register often has to answer to a framework as well as to the law, and the temptation is to summarise the answer as a percentage. That is a mistake for three reasons. A percentage hides which items are unmet, and the unmet items are the point. Its denominator is arbitrary, since frameworks let organisations choose which parts apply. And it invites self-assessment to drift upwards, because a score is something to improve and a position is something to defend.
The pattern that works is one position per framework item per use: met, partly met, not met or not applicable, with the evidence reference and the date it was last confirmed, and "not assessed" as the honest default. A count of items at each position is a fair summary. A single figure is not.
What to record against NIST AI RMF and ISO/IEC 42001
The NIST AI Risk Management Framework, released on 26 January 2023 and described by NIST as voluntary, is a useful public reference because it is not a checklist. Its core has four functions, Govern, Map, Measure and Manage, and the tables in NIST AI 100-1 list 19 categories and 72 subcategories beneath them, with NIST stating that the actions do not constitute a checklist and are not necessarily an ordered set of steps. Among the Govern outcomes are that mechanisms are in place to inventory AI systems and that processes exist for decommissioning them safely. NIST's framework page, read on 9 November 2026, says AI RMF 1.0 is being revised, so a register that cites subcategory identifiers should record the version it cites.
ISO/IEC 42001:2023 is a different kind of document. It is a management system standard, adopted identically in Australia as AS ISO/IEC 42001:2023 and published by Standards Australia on 16 February 2024, which describes it as providing requirements for establishing, implementing, maintaining and continually improving an AI management system. Its text is sold, so a register records it by citation only: the clause a position relates to, never its wording. Standards Australia's own post on the standard says organisations may adopt it for internal conformance or seek third-party assessment where available; a register entry is neither.
Populating the register from what you already hold
The usual way to build a register is a survey: a form goes to every team asking what AI they use, some of them answer, and the register starts life incomplete and already distrusted. Surveys fail because the people who could answer do not think of what they do as using AI, because the question arrives without context, and because nothing in their week depends on answering it.
The material that does answer the question already exists. Contracts and purchase orders name the vendors, the products and the AI features inside them. Policies name the uses the organisation has decided to allow or forbid, which is a list of expected entries. Process maps and procedures show where a system acts inside a workflow, which is where a use lives. Meeting notes, project documents and wiki pages record the pilots, the tools a team adopted and the decisions nobody wrote up elsewhere. Privacy impact assessments and risk assessments already hold the classification and its basis for the uses they covered.
Reading that material produces candidate entries with their provenance attached: this use, found in this document, owned by this person according to that page. Each candidate is then confirmed or rejected by the person named, a far smaller ask than a blank form. Omni by Valstra holds a register of this shape in the same model as the processes the uses sit inside, built from the documents, meeting transcripts and wiki pages an organisation already has, and nothing is written to it until a person accepts it. It is one way to do the work; a disciplined analyst with the same documents and a spreadsheet is another.
Keeping the register current
A register is accurate on the day it is finished and decays from there. What keeps it current is not a policy that says it must be, but a small set of events that each trigger a change to a named entry, and one person who is accountable for the register as a whole.
Every change should be a new dated line, not an overwrite, so the register can say what was recorded at the time a decision was made. A register that can show its own history is evidence; one that shows only its current state is a claim.
The test of a current register is that a question from a board, an auditor or the OAIC can be answered from it in an afternoon, with the documents it points at, and without a phone call to find out what is actually running.
- A new vendor contract or renewal: the entry is created, or its system field updated, before the contract is signed, so the register is a gate rather than an afterthought.
- A change to a process the use sits inside: the process owner's change prompts a check of every entry linked to that process.
- A model or vendor change, including a vendor turning a feature on: the system field, the risk basis and the oversight point are re-confirmed.
- An incident or a complaint about an outcome: the entry records it and its review date moves forward.
- A regulatory date: every entry flagged as making or informing decisions about individuals is reviewed, and the privacy policy wording it supports confirmed, before 10 December 2026.
- The review date itself: the owner re-confirms each statement, or marks it "not confirmed since" with the date.
Sources
- Australian Government, Guidance for AI Adoption: AI register template
- OAIC, APP Guidelines, Chapter 1: APP 1 Open and transparent management of personal information
- OAIC, APP 1.7 to 1.9 transparency obligation fact sheet, September 2026 (PDF)
- OAIC, New resources on transparency for use of AI and automated decision-making, 30 September 2026
- NIST, Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1, January 2023 (PDF)
- NIST, AI Risk Management Framework
- Standards Australia, AS ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system
- Standards Australia, Standards Australia adopts the international standard for AI management system, AS ISO/IEC 42001:2023, 16 February 2024
- Standards Australia, Spotlight on AS ISO/IEC 42001:2023, 5 September 2025
Common questions
- Is a spreadsheet enough to start an AI system register?
- Yes, if it has one row per use, the fields described on this page, a single accountable owner per row and a dated history of changes rather than overwrites. It stops being enough when entries need to link to processes, approvals and evidence held elsewhere, when more than one team edits it, or when you need to show what it said on a past date.
- Does an AI system register make us compliant with anything?
- No. The APP 1.7 to 1.9 obligation is met in the privacy policy, the NIST AI RMF is voluntary and says its actions are not a checklist, and ISO/IEC 42001 is a standard for a whole management system, not for a list. The register is what lets you know what to disclose, what to assess and what to show; it is the record, not the compliance.
- Who should own an AI system register?
- One named person owns the register as a whole, usually in risk or compliance, and answers for its completeness. Each entry is owned by exactly one named person, usually the owner of the business process the use sits inside, because they can stop or change the use; a vendor, a committee or a team is never the owner.
- How often must an AI system register be reviewed?
- None of the sources on this page sets a fixed interval; the NIST framework leaves the frequency of periodic review for the organisation to determine. Set a review date per entry by risk, re-check an entry whenever its system, process, vendor or outcomes change, and review every entry flagged as making or informing decisions about individuals before 10 December 2026.
See it against your own processes
A demo runs on your material, not a canned dataset. Bring a process you find hard to explain.