Australia's Voluntary AI Safety Standard: a practical guide for SMEs in 2026

Australia confirmed in December 2025 that there will be no standalone AI Act. The current framework is the Voluntary AI Safety Standard with 10 voluntary guardrails, an Australian AI Safety Institute funded at AUD 29.9 million launching in early 2026, and full continuing application of existing law including the Privacy Act. Voluntary does not mean ignorable. It means the regulation lives somewhere other than an AI-specific Act.

The short version

Australia decided in early December 2025 not to introduce a standalone AI Act. The framework that replaces the originally proposed mandatory guardrails for high-risk AI is three things: the Voluntary AI Safety Standard with 10 voluntary guardrails, the new Australian AI Safety Institute (allocated AUD 29.9 million in funding per IAPP reporting on the National AI Plan announcement, 2 December 2025), and the continued full application of the existing legal framework (Privacy Act, consumer law, IP law, sector-specific regulation).

For Australian SMEs, this means there is no AI-specific compliance regime to meet, and there is no AI Act to register under, certify for, or audit against. It does not mean AI is unregulated. The Privacy Act applies to any AI system that handles personal information. Australian Consumer Law applies to any AI system that makes consumer-facing decisions. Sector regulators (APRA, ASIC, ACMA, OAIC, TGA) apply their existing remit to AI used inside their sectors.

The right framing for a business deploying AI in 2026 is not "do I need to comply with an AI Act?" The answer to that is no. It is "have I documented how this AI system meets the standards I am already legally obliged to meet, in a way the relevant regulator would find credible?" That is what the 10 voluntary guardrails are for.

How Australia landed here

Through 2024 and most of 2025, the public expectation was that Australia would follow Europe and parts of the United States in introducing prescriptive AI regulation. A set of mandatory guardrails for high-risk AI settings was actively consulted on through 2024 (the Voluntary AI Safety Standard was released in September 2024 alongside a proposals paper on mandatory guardrails).

In December 2025, the Australian Government''s National AI Plan confirmed a different direction: no standalone Act, no mandatory AI-specific licensing, reliance on existing law plus voluntary guidance plus a new AI Safety Institute (per the IAPP roadmap reporting).

The reasoning, summarised charitably, is that Australia''s existing law already covers most of the harms an AI Act would address (privacy, consumer protection, anti-discrimination, sector-specific safety duties), and that creating a parallel AI regime would mostly produce regulatory duplication rather than additional protection. The reasoning, summarised critically, is that an AI Act is hard to draft well, expensive to implement, and politically risky during a productivity-focused electoral cycle. Either way, the framework is what businesses now operate under.

The 10 voluntary guardrails

The Voluntary AI Safety Standard published by the Department of Industry, Science and Resources sets out 10 guardrails covering accountability, risk management, data governance, monitoring, transparency, human oversight, contestability, supply chain accountability, stakeholder engagement, and ongoing compliance. Read the full text of the guardrails directly at the Department''s page.

For practical purposes inside an SME, the 10 guardrails translate into roughly 10 questions you should be able to answer in writing for any AI system you put into production:

1. Who in the business is accountable for this AI system and its decisions?
2. What is the risk it could go wrong, and how would you know if it did?
3. Where does the data come from, and is the business legally allowed to use it for this purpose?
4. How is the system monitored once it is live, and how often?
5. What is disclosed to users and customers about how the AI is being used?
6. What is the role of a human in the loop, and where is that human empowered to override the AI?
7. How can a user, customer, or staff member challenge a decision the AI has made about them?
8. What due diligence has been done on third-party AI providers in the supply chain?
9. Have affected stakeholders (staff, customers, vendors) been consulted on material AI deployments?
10. What is the cadence at which this system is reviewed for continued safety and fitness for purpose?

A written answer to each of these is not a compliance certificate. It is the document you would produce in response to a regulator query, an internal incident, or a customer complaint. For most SMEs, a clear one-page response per guardrail is the practical target.

The Privacy Act still applies in full

The single most consequential point of confusion about Australia''s voluntary framework is the treatment of personal information. The voluntary framework does not change Privacy Act obligations. It sits on top of them.

If your AI system ingests, processes, or generates personal information (customer names, contact details, behavioural data, voice recordings, written communications that identify individuals), the Australian Privacy Principles apply in full. The Office of the Australian Information Commissioner has been clear in its guidance on AI and the Privacy Act that AI-related privacy breaches will be assessed under the existing Privacy Act, not under a separate AI regime.

The practical implications are concrete. A documented privacy impact assessment is a baseline expectation for any AI system handling personal information, not a nice-to-have. The default position on consent and notification is that customers should know if their data is being processed by AI in a way they would not reasonably expect from a non-AI service. The reasonable-steps test for security in APP 11 applies to AI providers in the supply chain.

A business that has not done this work is not protected by the fact that there is no AI Act. It is exposed under the law that exists.

Sector regulators have not stood down

The decision not to pass an AI Act does not affect sector regulators. APRA continues to issue prudential guidance on AI use by regulated financial entities. ASIC has been clear that AI-generated misleading representations are still misleading representations under consumer law. ACMA''s remit over telecommunications and online safety applies to AI-mediated services in those sectors. The TGA regulates AI in medical device contexts under the existing therapeutic goods framework.

For an SME in a regulated sector, the practical compliance picture is unchanged: meet your sector regulator''s expectations, and treat the voluntary AI framework as supplementary documentation rather than the primary compliance obligation.

For an SME outside the regulated sectors, the picture is different: the Privacy Act and Australian Consumer Law are the main constraints, and the voluntary guardrails are the documentation that demonstrates you have thought about the AI-specific dimensions of those constraints.

What the AI Safety Institute actually does

The Australian AI Safety Institute, allocated AUD 29.9 million in funding and scheduled to launch in early 2026 per IAPP reporting, has three broad functions in its public framing: technical AI safety research, advice to government on emerging AI risks, and supporting voluntary best practice across Australian industry.

It is explicitly not a regulator. It does not issue licences, fines, or compliance certificates. It does not assess individual AI systems for safety. It does not certify businesses as AI-safe.

What it will do, over time, is shape the technical standards that regulators, courts, and counterparties consider reasonable. The Voluntary AI Safety Standard is the current expression of that work. Future Institute outputs (technical guidelines, sector-specific guidance, model evaluation protocols) will be expressed as voluntary best practice, but will become the de facto floor for what reasonable conduct looks like. For an SME, the practical takeaway is to track Institute publications and update internal AI policies in line with them.

A practical compliance posture for an Australian SME

For most Australian SMEs deploying AI in 2026, a defensible position can be assembled in approximately a week of focused work. The components:

A one-page AI policy. Covers acceptable use, prohibited use, data handling expectations, accountability for AI-assisted decisions, and the process for adopting new AI tools. Signed off at the senior leadership level.

A documented AI inventory. A simple register of every AI tool in use across the business, with named owner, purpose, data inputs, data outputs, and review date. Updated quarterly.

A response to each of the 10 voluntary guardrails. One page per guardrail describing how the business''s AI use addresses the guardrail in practice. Honest where compensating controls are limited, rather than aspirational.

A privacy impact assessment for any AI system processing personal information. Standard structure: what data, what processing, what disclosure, what risk, what controls. The OAIC publishes a usable PIA template.

A vendor due diligence record for third-party AI providers. Captures what each provider has committed to under their terms, where the data sits, and what the business''s recourse is if the provider changes its model or terms.

A review cadence. Quarterly is reasonable for most SMEs. Annual is acceptable for businesses with low AI exposure. Monthly is appropriate for businesses with customer-facing AI in critical decision paths.

Together, these six artefacts cost roughly a week of senior time to assemble and a half-day per quarter to maintain. They are the practical answer to "what does AI compliance look like for an Australian SME in 2026."

What this does not cover

The framework above is fit for most Australian SMEs. It is not fit for purposes that genuinely sit in the high-risk bucket: AI systems making material employment, credit, insurance, healthcare, or law-enforcement decisions; AI systems processing biometric data; AI systems operating in critical infrastructure; AI deployed at scale to make consumer-facing decisions without human review.

Businesses operating in any of those categories should not rely on the voluntary framework alone. The applicable sector regulator''s guidance, the OAIC''s specific high-risk advice, and a documented legal review of the relevant duties are necessary. The voluntary guardrails are the floor in those settings, not the ceiling.

The honest summary

Australia''s AI regulatory position in 2026 is permissive but not absent. There is no AI Act, and there is unlikely to be one in the near term. The 10 voluntary guardrails, the AI Safety Institute, and the unchanged application of existing law together form a coherent framework that is lighter than Europe''s and heavier than the United States federal baseline.

For an Australian SME, the practical implication is to adopt AI fast, document well, and treat the voluntary guardrails as the standard of care a competent operator should be able to demonstrate. The cost of doing this work is modest. The cost of not doing it, in the event of a privacy breach, a customer complaint, or a sector regulator query, is meaningfully higher.

Sources and references

• Department of Industry, Science and Resources: Voluntary AI Safety Standard (primary publication)
• Department of Industry, Science and Resources: The 10 guardrails (primary publication of the guardrails)
• IAPP: Australia unveils AI policy roadmap (2 December 2025; confirms no standalone AI Act and AUD 29.9 million funding for the AI Safety Institute)
• MinterEllison: Responsible use of AI: new Australian guardrails released (5 September 2024; legal commentary on the release of the standard and proposed mandatory guardrails)
• Office of the Australian Information Commissioner: Guidance on privacy and the use of commercially available AI products
• Office of the Australian Information Commissioner: Privacy Impact Assessment guidance
• Office of the Australian Information Commissioner: Australian Privacy Principles

Where claims could not be tied to a primary or quoted source they have been removed rather than retained without citation. Numbers of guardrails and funding figures are quoted exactly from the Department of Industry, Science and Resources and IAPP respectively.