Skip to main content
All articles

Data Sovereignty

NSW AI Assessment Framework: What It Means for GovTech

NSW asks its agencies to prove their AI is safe before they use it. That question flows straight through procurement to every vendor selling AI-powered software to government.
  • September 2026 · updated October 2026
  • 10 min read

Short answer

The NSW AI Assessment Framework is how NSW Government agencies assess the risk of each AI use case before and while they use it, under the NSW AI Operational Policy. The assessment belongs to the agency, not the vendor, so no product is “approved” under it in the abstract. For GovTech, it means agencies will ask suppliers for specific evidence: what the AI does, where it runs, where people review its outputs, and how its contribution is logged.

Updated October 2026

NSW has replaced its AI Assurance Framework with a redesigned AI Assessment Framework (AIAF), launched in January 2026 and now applied under the NSW AI Operational Policy. From 30 September 2026, agencies must use the AIAF Platform for all new AI use case registrations and assessments. This article has been updated to reflect the current framework.

New South Wales was early among Australian governments to require its agencies to formally assess the AI systems they use, and its approach is a common reference point for AI governance in Australian government. For agencies, the framework turns "should we use this AI system" from a philosophical debate into a structured assessment. For vendors, it changes the sales conversation permanently: an agency that must document its AI risk position will ask its software suppliers for the evidence, and a vendor that cannot provide it becomes the agency's problem, which is a poor place from which to renew a contract.

What the Framework Actually Requires

Three NSW instruments now work together. The NSW AI Operational Policy sets mandatory requirements for the governance, assurance and acceptable use of AI across government. Circular DCS-2026-02 requires agencies to comply with that policy and to apply the AI Assessment Framework when it is required. And the AIAF Platform is where AI use cases are registered and assessed: from 30 September 2026, agencies must use it for all new registrations and assessments, with a controlled offline pathway only for agencies not yet able to access the platform.

The redesigned framework, developed by the NSW Office for AI with CSIRO's Data61, replaced lengthy self-assessments with a standards-aligned, risk-based process aligned with the Commonwealth's national framework for the assurance of AI in government. The NSW Government says an assessment now takes under 30 minutes. Low-risk uses move through quickly; uses assessed as high or critical risk must go to the AI Secretariat for review by the NSW AI Review Committee, and can trigger mandatory safeguards such as a privacy impact assessment. An AI feature that suggests how to word a newsletter is treated differently from one that influences decisions about individuals, and rightly so. The framework applies to AI use cases throughout their lifecycle, whether or not they sit inside a formal project.

The operational policy applies to NSW Government agencies. It has been reported as not mandatory for local government, universities or third-party suppliers, although councils and suppliers working with agencies will meet its questions through procurement. Check the current policy and circular for your organisation's position.

Two properties of the framework matter most for anyone selling into or buying for NSW government. First, the obligation sits with the agency, not the vendor. An agency cannot outsource its assessment by pointing at a vendor's marketing page; it needs real information about how the AI works, what data it uses, and what controls surround it. Second, the assessment is ongoing rather than a one-off gate. AI systems change, models are updated, and use cases expand, so the evidence a vendor provided at procurement needs to stay true, and stay available, through the life of the contract.

How It Flows Through to Procurement

The practical consequence is a new layer of questions in tenders and vendor assessments, sitting alongside the established security and hosting questions. Expect to be asked, or if you are the buyer, expect to ask: What AI capabilities does the product contain, precisely? What data do they process, and where? Which decisions do they make autonomously, and which merely recommend to a human? How is accuracy measured and monitored over time? What happens when the AI is wrong, and how would anyone know? Can AI-assisted outputs be distinguished from human ones after the fact?

For stakeholder engagement platforms, these questions have particular bite, because the data is community members' opinions, complaints, and personal information, and the outputs, including sentiment assessments, risk scores, and suggested responses, can influence how a government treats the people it serves. A vendor in this category should be expected to answer in specifics, not principles. The honest answers are checkable: they name the regions where inference runs, describe the human review points in the workflow, and explain what confidence information accompanies each AI output.

Questions That Separate Real Assurance from Marketing

  • Which specific outputs are AI-generated, and does each carry a confidence indicator a user can see?
  • Where does inference physically run, and is that enforced technically or just contractually?
  • What human review points exist before an AI output affects a decision about a person?
  • Are AI-assisted actions logged distinctly, so an audit can reconstruct what the AI contributed?
  • When models are updated, how is the change communicated and its impact measured?
  • Can the AI features be scoped or disabled per use case if an assessment requires it?

What Good Looks Like from a Vendor

The vendors who will thrive under AI assurance regimes are the ones who treat the framework's concerns as engineering requirements rather than compliance paperwork. That means human-in-the-loop review designed into workflows where AI outputs carry consequences, not bolted on. Confidence scoring exposed in the interface, so users can calibrate their trust. Audit trails that record what the AI contributed to each decision. Technical enforcement of data residency for AI processing, rather than policy promises. And plain documentation of what each AI feature does, honestly including what it does badly.

This is the approach we take at VivoAI: confidence scoring on AI outputs, human review points on consequential actions, and full audit trails on AI-assisted decisions, with inference running in the AWS Sydney region and enforced in code. In the spirit of the transparency this article argues for: no vendor's product, ours included, can be "approved" under the NSW AI Assessment Framework in the abstract, because the assessment belongs to the deploying agency and its specific use case. What a vendor owes you is the evidence that makes your assessment straightforward, and that is the standard we hold ourselves to. Our broader approach to data residency and onshore AI is documented in the same spirit.

The Wider Picture: Assurance Is Converging

NSW is not an island. The Commonwealth has published its own policy for the responsible use of AI in government, other states are moving along similar lines, and the direction across all of them is consistent: risk-based assessment, human accountability, transparency to the public, and ongoing monitoring rather than one-off approval. For agencies, that convergence is good news, because the evidence assembled for one framework substantially serves the others. For vendors, it means the window for treating AI governance as someone else's problem has closed. The questions are standard now. The differentiator is whether your answers are real.

For government buyers weighing platforms in our category, our guides to government stakeholder software and Australian data sovereignty cover the neighbouring questions of hosting, security, and procurement.

Where VivoAI fits

VivoAI is a stakeholder engagement platform whose AI transcribes and summarises calls, suggests links between records, clusters submissions and drafts replies, with people confirming each step. Outputs carry confidence levels, AI-assisted decisions keep a full audit trail, and inference runs in the AWS Sydney region, enforced in code. We supply agencies with the information their own assessment needs; we do not claim a status under the framework.

See how the AI keeps the record.

Frequently asked questions

What is the NSW AI Assessment Framework?

It is the NSW Government’s risk-based process for assessing AI use cases, which replaced the earlier AI Assurance Framework. It was launched in January 2026 and is applied under the NSW AI Operational Policy. From 30 September 2026, agencies must use the AIAF Platform for all new AI use case registrations and assessments.

Can a vendor’s product be approved under the NSW AI Assessment Framework?

No. The assessment belongs to the deploying agency and its specific use case, not to a product in the abstract. What a vendor can do is supply the evidence the agency needs: what the AI does, what data it uses, where it runs, where people review its outputs, and how its contribution is logged.

Does the framework apply to councils and suppliers?

The operational policy applies to NSW Government agencies. It has been reported as not mandatory for local government, universities or third-party suppliers, although councils and suppliers working with agencies will meet its questions through procurement. Check the current policy and circular for your organisation’s position.

What happens to high-risk AI use cases?

Use cases assessed as high or critical risk must go to the AI Secretariat for review by the NSW AI Review Committee, and can trigger mandatory safeguards such as a privacy impact assessment. Low-risk uses move through more quickly.

Sources

Next step

Ask us the hard questions

Bring your AI assurance checklist to a demo. We would rather show you the evidence than tell you a slogan.