If you sell software to a NSW agency, work in one, or are on the receiving end of its decisions, the useful question is no longer whether government uses AI. The auditor’s answer to that is yes, everywhere it looked. The question is whether the scaffolding has caught up. By the Auditor-General’s own count, the people side mostly has: someone owns AI at every agency examined, and most train their staff. The machinery around it mostly has not. Risk frameworks, IT testing, contracts, the register and the budget lag behind, and buying AI comes last.
The count, from the auditor’s table
Source: NSW Auditor-General, Internal controls and governance 2026, chapter 9, governance table on page 44 (number of agencies that have considered each element, out of 10 selected agencies). “Regular reporting” is the 7 agencies that report regularly to senior management or governance bodies; the other 3 report ad hoc at project level. The 10 agencies were selected for the highest AI use or greatest potential benefit (page 40) and are not named in the report. This is a different and smaller sample from the 2025 report’s 26 agencies, so the two years are not a trend line. Read 4 October 2026.
The table is short and blunt. Accountability, meaning an overall owner responsible for
deploying and maintaining each AI solution, is in place at all 10. Training in responsible
use is in place at eight. Seven report regularly to senior management. Then the line falls
away. Five agencies have looked at their risk management framework in light of AI, four have
looked at their IT policies and testing protocols (the
report’s example is testing for
unintended biases and vulnerabilities
), and one has considered AI in procurement,
which the report notes is not mandatory for the NSW public sector although the national
framework calls for careful attention to contracts.
Policy sits at the same halfway mark. The
report
heads its finding All selected agencies use AI but only half have an AI policy
, and
says agencies leaning on the whole-of-government ethics policy instead are not covered:
While the AI Ethics Policy sets out overarching principles, it is not sufficient
.
Agency-level policy is what says who owns AI, who reviews a new use case, and how and to whom
staff report misuse. Half of the 10 also have a formal AI strategy.
Registers that miss what was never assessed
All 10 agencies have an AI register. Six capture all their AI use. The other four capture
only AI that has been through the NSW assessment framework, which, in the words of the
report’s executive summary,
means increasing the risk of unmanaged and unidentified AI use
. The report adds two
harder cases: AI that third parties use on an agency’s behalf, and unapproved AI
tools. Last year’s report had already recommended a central inventory of every AI tool;
this one restates the case that a complete inventory would
increase their ability to confirm that governance arrangements are fit for purpose
.
Money is the other blind spot. Most of the 10 did not centrally track the cost of AI or set
budgets for it, and the report explains why that matters more for AI than for ordinary
software: pricing is moving from per-user licences to charges per token of compute used.
As the
report
puts it, Token prices and usage can be volatile which may result in unexpected costs to
agencies.
On assessment itself the picture is better. Of 15 projects the auditor selected, 10 under
$5 million and five over, 13 had been through the AI Assessment Framework. The two that had
not were implemented before the framework existed, and four of the 15 had no cyber risk
assessment. The report accepts that retrospective assessment is not required, then points
out that the framework applies across a system’s whole lifecycle, so
Agencies need to assess existing AI solutions against the current guidance.
That is
the report’s one AI recommendation: by 30 June 2027, agencies should assess all
required AI solutions against the framework and use it to manage AI through its lifecycle.
The case study that shows it can be done
The report’s worked example is Transport for NSW’s seatbelt cameras. Since July 2024 the agency has enforced seatbelt offences automatically, building on its mobile phone detection cameras, with AI from a private provider making probabilistic classifications of whether a photo shows an offence. It was among the first projects to go to the AI Review Committee. The auditor records that the system enables scanning of over 100 million trips and aims to reach close to every NSW driver each year, and that Monash University Accident Research Centre, engaged by the agency to model the outcomes, estimated it would prevent 17 to 26 deaths and 41 to 62 serious injuries over five years.
What earns it a case study is the design around the model. According to the
report,
Most images are never seen by a human
; images showing no offence are deleted, and
those escalated are cropped and pixelated. The AI only filters: At least 3 human
decision-makers review all potential offences before an infringement is issued
, the
final decision to issue a penalty sits with Revenue NSW officers, and the usual review and
court rights remain. Transport for NSW has published how the process works, and the report
says early data suggests seatbelt compliance is improving. The auditor’s own takeaway:
The practical lesson for agencies is to build these controls into program design from
the start.
What has changed since the audit
The rulebook the auditor tested against has been replaced. Circular
DCS-2026-02,
issued 30 July 2026 and updated 1 September, makes the new
NSW AI Operational Policy
mandatory. It replaces the NSW AI Ethics Policy and adopts Australia’s AI Ethics
Principles instead. The report acknowledges the change, and records that the Office of AI
(since renamed NSW GovAI) told the auditor the new policy will improve some of the matters
raised in this report
. Here is how the counts line up against what the policy now
requires.
| The auditor found, at 10 agencies | The Operational Policy now requires |
|---|---|
| 10 have addressed accountability | Each agency head nominates an Accountable Official, a senior executive, and notifies NSW GovAI within three months of the policy’s issue; a governance board must oversee AI |
| Half have an AI policy | No standalone agency AI policy as such, but ICT acceptable use policies must align with the policy’s acceptable use standard; agency rules may add requirements, never remove them |
| 4 registers capture only assessed AI | Register AI use cases in the AI Assessment Framework Platform, with routine productivity use generally exempt (see below) |
| 5 have considered AI in risk management | Oversight proportionate to each use case’s risk band; high and critical risk referred to the AI Review Committee; annual audits for high and critical risk |
| 1 has considered AI in procurement | Third-party supplier risk managed by the system’s owner, with contracts that let suppliers disclose AI use and notify material changes |
| 8 have training | AI literacy and policy training made available to staff who use AI, using modules NSW GovAI provides |
| Most do not centrally track AI costs | We found no cost or budget requirement in the policy |
| 7 report regularly to senior management | A signed annual attestation to NSW GovAI by 31 October each year, the first due 31 October 2027 |
The register row needs the detail. The policy’s fifth mandatory requirement reads
in the policy itself
Register all AI use cases in the AIAF Platform
. Its Attachment A then narrows what
“all” means: drafting, summarisation, coding support and other routine productivity
use of approved agency tools do not generally require registration
where a human stays
meaningfully in charge and the use stays within agency guardrails. Registration is required
where AI influences consequential decisions, sits inside operational or public-facing
systems, acts autonomously, or handles highly sensitive information. And for AI already
running: If AI is already in use, registration must occur as soon as practicable
.
Two other changes landed in the same window. The report notes that Cyber Security NSW revised its Cyber Security Policy in July 2026, adding guidance covering AI. And the Audit Office has a dedicated performance audit, NSW government agencies’ use of artificial intelligence systems, among the 2026–27 audits in its Audit Work Plan 2026–2029, which says that list includes audits that start in the year but publish in the next.
Our read
This is our view, built on the figures above. The first thing worth noticing is that the
shape has repeated. Last year’s report, covering 21 AI-using agencies, also put
accountability and training at the top and procurement at the bottom, at 5 per cent (we
covered it in NSW agencies run 357 AI tools).
Different samples, so not a trend. But the same order in two different samples says
something about where the effort goes: agencies have named owners and run courses, the
visible work, while the contract is where AI quietly arrives. The
report
itself says adoption is driven in part by vendors integrating AI capabilities into existing
off the shelf platforms and cloud services
. An agency that has not considered AI in its
procurement is still adopting AI through its software renewals. It just has not decided to.
On paper, the new policy answers most of the table, and its supplier disclosure clause is the most direct answer to that procurement count of one. Two things it leaves alone. It says nothing we could find about tracking what AI costs, which the auditor flagged at most of the 10 agencies. And its register is a filter by design. Exempting routine AI drafting from registration is defensible; registering every summary would bury the uses that matter. But the auditor’s complaint was registers that miss AI nobody assessed, and its 2025 recommendation was a central AI inventory. The platform will produce a list of AI that carries risk, which is not the same thing. Its count should never be read as the number of AI tools NSW runs.
Then there is the clock. The auditor wants existing AI assessed by 30 June 2027. The first
attestation, signed by each agency head but reported by the agency itself, is due 31 October
2027, and the policy offers extensions for that first round. The performance audit is the
next independent count, and the report has already said, Governance is not keeping pace
with the speed at which agencies are adopting AI
. The seatbelt cameras show NSW can build
AI governance properly when it starts from the design. The table shows how much of the
governance around everything else is still being retrofitted.
Sources
- Internal Controls and Governance 2026: grants, consultants, purchasing cards and technology, NSW Auditor-General’s Report to Parliament, financial audit, 12 August 2026 (PDF, 50 pages, downloaded and read in full 4 October 2026): the report snapshot (page 1); the AI key findings on registers, costs, strategy and assurance (pages 5 and 6); recommendation 4 and its 30 June 2027 date (page 6); the practical lessons on vendor-integrated AI (page 9); the July 2026 Cyber Security Policy revision (page 33); chapter 9 in full, covering the selection of the 10 agencies, the new policy framework and the Office of AI’s advice (page 40), the Transport for NSW seatbelt case study (pages 41 and 42), registers, costs, the 15 projects and strategy (page 43), and the governance table (page 44).
- Report landing page, audit.nsw.gov.au (read 4 October 2026): the 12 August 2026 publication date and the report snapshot.
- Circular DCS-2026-02, Use of Artificial Intelligence by NSW Government Agencies, Department of Customer Service, issued 30 July 2026, updated 1 September 2026, mandatory (read in full 4 October 2026): the mandatory agency requirements, the 31 October 2027 first attestation, and that it replaces DCS-2024-04.
- NSW AI Operational Policy, final version 1.1, Department of Customer Service (PDF, 25 pages, downloaded and read in full 4 October 2026): the seven mandatory requirements (section 4), the Accountable Official, governance board and assurance accountabilities including supplier disclosure (section 5), attestation and the first-period extension pathway (section 7), version history (section 9), and use case registration and its exemptions (Attachment A).
- NSW AI Operational Policy page, Digital NSW (read 4 October 2026): the summary of agency and employee responsibilities and the link to the policy PDF.
- Audit Work Plan 2026–2029, Audit Office of NSW (read 4 October 2026): the planned performance audit of NSW government agencies’ use of artificial intelligence systems, and the note that 2026–27 audits may publish in the following year.
- Internal controls and governance 2025: Procurement and technology, NSW Auditor-General, tabled 29 October 2025 (PDF read in full 11 July 2026 for our earlier story): the 2025 figures referred to in Our read.
How we checked this. Every count in the chart and the table is from the governance table on page 44 of the 2026 report or the findings on pages 5, 6 and 43, and page numbers are the report’s printed numbers. The report describes the agencies as “selected” and does not name them, so we do not either; it is a financial audit report and carries no agency responses. The report calls the framework both the “NSW AI Assessment Framework” and the “NSW AI Assurance Framework”; we read both as the same assessment framework the circular mandates. The right-hand column of the table is our summary of the policy, read in full, set against the auditor’s categories; matching a finding to a requirement is our judgement, not the auditor’s. The statement that the policy has no cost requirement rests on a full-text search of the policy for cost and budget terms. Our read is opinion based on the documents cited.
Work in one of these agencies, or read the report differently? Tell us and we will check it against the documents and log the outcome here.