New South Wales AI assessment framework: Making AI risk assessment mandatory across state government

The problem

As AI adoption grows across government, so does the need for a consistent way to assess the risks it introduces. AI systems can raise questions about privacy, fairness, security, transparency, and accountability, and those questions look different depending on the system, the context, and the people it affects.

The New South Wales (NSW) Government took the view that the safe and responsible use of AI across its agencies needed more than published principles or general guidance. It needed a structured process that every agency would follow, one that would identify the risks posed by a specific AI system, determine the required safeguards, and ensure that the highest-risk systems received independent oversight.

The solution

The NSW Government developed the AI Assessment Framework (AIAF), a mandatory risk assessment tool that every state government agency must use whenever it designs, develops, deploys, procures, or uses any system containing AI.

This is not optional guidance. Under Circular DCS-2024-04, all NSW Government agencies must apply the AIAF alongside the NSW AI Ethics Policy. Every use of AI, whether or not it is part of a formal project, must go through the process.

The framework is structured as a workbook with four sections.

The first section makes sure the right people are involved and records basic information about the AI system: what it does, who is responsible for it, and when the assessment was completed.

The second section is the core of the process. Agencies answer the 16 assessment questions about their AI system, which are aligned with the state's five AI ethics principles: community benefit, fairness, privacy and security, transparency, and accountability. Based on the answers, the framework generates the overall risk level and a list of mandatory assurance activities the agency must complete. These can include a Privacy Impact Assessment, a Human Rights Impact Assessment, or a cybersecurity review. The framework sets these requirements from the responses; the agency does not choose them.

The third section goes deeper. It contains additional questions that help agencies identify further risks and connect them to practical safeguards. This section is optional but recommended. Any risks that surface are added to the risk register the agency maintains.

The fourth section covers what happens after the assessment. It records sign-off, requires the completed assessment to be saved in the agency's records system, and requires any AI system rated high or critical risk to be registered in the agency's AI register and referred to the NSW AI Review Committee for independent review.

The framework is not a one-off exercise. Agencies must re-run the assessment whenever the AI system's features, datasets, purpose, or decision context changes. A system that was low risk at launch can become higher risk if its scope expands or it is used in a new way.

The impact

1. Every NSW Government agency must now assess every use of AI through a consistent process

The AIAF applies across the entire state government, regardless of the agency, the size of the project, or whether the AI is part of a formal initiative. This creates a common baseline: no agency can adopt AI without working through the same questions and following the actions the framework requires.

2. Agencies cannot set their own requirements

Because the framework generates the risk level and the required follow-up actions from an agency's answers, an agency cannot complete the assessment and then decide for itself that nothing further is needed. This reduces the chance that an agency underestimates what its AI system requires.

3. The highest-risk AI systems receive an independent review

Systems assessed as high or critical risk are referred to the AI Review Committee, which reviews them alongside the agency's own assessment. Projects with budgets above $5 million, or those funded through the Digital Restart Fund, receive additional central oversight under the NSW Digital Assurance Framework. Low-risk applications undergo a proportionate self-assessment, while the highest-risk systems receive dedicated scrutiny.

4. The framework applies across the AI system's whole life

The AIAF is built around the full lifecycle of an AI system. Because agencies reassess when a system materially changes, governance keeps pace as the system evolves, well beyond the initial check at launch.

5. AI governance is embedded within the state's existing oversight structures

The AIAF is integrated into the NSW Digital Assurance Framework and sits alongside the NSW AI Strategy, the AI Ethics Policy, and the state's other AI guidance, including its guidance on agentic AI. AI oversight is part of how digital projects are already managed, so agencies handle it within their existing processes.

Lessons learned

• Making the framework mandatory closes the gap between principles and practice. Many governments publish AI ethics guidelines that agencies are encouraged to follow. NSW made the framework mandatory. AI governance does not depend on whether individual agencies choose to engage; every AI use case by a government agency must go through the process.

• Letting the framework determine the requirements reduces the risk of inconsistent self-assessment. The 16 assessment questions automatically generate the risk level and the mandatory actions. An agency cannot answer them and then decide for itself that nothing further is needed. For governments designing similar frameworks, building the requirements into the assessment's structure makes the process more reliable because it does not depend on each assessor's judgement.

• A comprehensive framework needs to stay proportionate to avoid overburdening small teams. Reviewers of the framework have noted that its thoroughness, a strength for high-risk systems, can be demanding for smaller teams and low-complexity use cases, and that closer mapping to international AI risk-management standards could improve its use beyond NSW. The tiered design helps, since low-risk uses receive a lighter self-assessment, and the broader lesson for governments is to keep the process proportionate so that simple uses are not overburdened.

This case study was written with assistance from artificial intelligence.