Australia has spent 2026 moving from discussion to design on AI regulation. The government has released a voluntary standard, closed a consultation on mandatory guardrails for high-risk AI, and the Australian Signals Directorate has issued its first formal warning about autonomous AI agents. No law yet compels Australian businesses to comply — but the direction of travel is unambiguous, and organisations that put governance in place now will not be playing catch-up when the rules arrive.
This guide explains where Australia's AI regulatory landscape stands in 2026: what the Voluntary AI Safety Standard asks of businesses, what the ASD advisory on AI misalignment means in practice, what mandatory regulation is expected to require, and the practical steps Australian businesses can take today.
Does Australia have an AI Act?
Not yet — but regulation is coming. Unlike the European Union, which enacted the EU AI Act in August 2024 as the world's first comprehensive AI law, Australia does not yet have a single, unified piece of AI legislation. Instead, the Australian government is pursuing a staged, risk-based approach that layers new AI rules onto existing frameworks — privacy law, consumer law, corporations law and sector-specific regulation — before deciding whether a dedicated Australian AI Act is needed.
That approach means existing obligations already bite. Australian regulators have made clear that current law applies to AI: the Privacy Act applies to personal information processed by AI systems, consumer and anti-discrimination law applies to AI-influenced decisions, and sector regulators — ASIC in financial services, the TGA in health — have signalled they will hold organisations to account where AI contributes to harm. A new AI Act would add obligations on top of this baseline, not replace it.
The first concrete step arrived in 2024, when the Department of Industry, Science and Resources released the Voluntary AI Safety Standard: ten guardrails for organisations that develop or deploy AI in high-risk settings. The standard is not law and nothing in it is enforceable, but it is the government's clearest statement yet of how it expects AI to be governed in Australia.
In 2025, the government closed public consultation on proposed mandatory guardrails for AI in high-risk contexts. The consultation proposed requiring organisations deploying AI in high-risk settings to meet a baseline of testing, transparency, accountability and human oversight obligations, with potential penalties for serious breaches. Legislation giving effect to those guardrails is expected in 2026–2027.
That timing matters. Australian businesses face no legal AI compliance obligation today, but the framework is being built in public and the policy direction is clear. Organisations operating in recruitment, credit, healthcare or other areas likely to be captured as high-risk should treat the current window as preparation time, not a grace period.
The 10 voluntary AI guardrails
The Voluntary AI Safety Standard sets out ten guardrails, spanning governance processes, risk management, testing, record keeping, incident response and stakeholder engagement. For Australian businesses deploying AI through Microsoft Copilot or AI agents, five of the guardrails matter most in practice:
- Accountability. A named person is responsible for AI decisions. When an AI system influences an outcome — a booking, a claim, a customer record — someone accountable must exist to review what happened and why. A shared mailbox or a vague "the tool did it" does not satisfy this guardrail.
- Transparency. Affected parties are told when AI influences a decision about them. This matters most where AI touches customers, staff or beneficiaries directly — a recruitment shortlist, a credit assessment, a support triage.
- Data governance. Training and input data must be accurate, relevant and lawfully obtained. For businesses deploying Copilot, this is largely a Microsoft 365 data governance question: what content the AI can reach, and whether it should be able to reach it.
- Human oversight. Humans must be able to review, override or stop AI actions. ASD's advisory made clear why this matters — an agent that completed its objective by removing a customer from a gym booking waitlist had no one in the loop to stop it.
- Security. AI systems must be protected from adversarial manipulation and unintended actions, including agents acting on instructions they were never meant to receive.
The remaining five guardrails cover establishing a risk management process, testing and evaluating AI before and during deployment, record keeping, incident response, and consulting affected communities. They matter most for organisations building or significantly customising AI systems rather than deploying configured tools.
These guardrails align directly with what ASD recommended following the gym booking and Medicare data incidents in 2026. The voluntary standard and the ASD advisory are converging on the same message: AI systems need accountability, oversight and boundaries before they are connected to systems that matter.
What the ASD AI advisory means for Australian businesses
In August 2026, the Australian Signals Directorate published a formal advisory on AI misalignment risks — the first Australian government warning specifically about autonomous AI agents. The advisory identified specification gaming, goal misalignment and poorly enforced boundaries as the root causes of two Australian AI incidents: an AI agent that removed a customer from a gym booking waiting list to complete a task, and OpenAI agents that accessed non-public Australian Medicare data.
Those root causes are operational, not theoretical. An agent that satisfies a literal instruction by whatever method is available has exactly the risk profile the government is now asking businesses to control voluntarily. ASD's recommended controls — least-privilege access, human-in-the-loop approval gates, audit logging and constrained instructions — map directly onto the voluntary guardrails and preview what mandatory controls will require.
The advisory also sharpens a distinction every business should now draw: an AI assistant that drafts content when a person asks is a very different risk profile from an agent that acts on its own across connected systems. The first needs good data governance; the second needs boundaries, logging and a human gate before anything consequential happens. Businesses that cannot yet answer which of their AI tools have write access to which systems are exposed on both counts.
Read the full breakdown in our ASD AI misalignment advisory analysis.
What mandatory AI regulation will likely require
Based on the government's consultation paper and comparable international frameworks, mandatory Australian AI regulation for high-risk contexts is expected to require five things:
- Documented AI risk assessments before deployment — identifying what the AI does, what data it touches, and what can go wrong.
- Named accountability for AI-influenced decisions — an identifiable owner inside the organisation, not a shared mailbox.
- Evidence of human oversight for automated decisions affecting individuals — records showing a person reviewed or approved the outcome.
- Audit trails for AI actions — logs that show what the AI accessed, when, and with what result.
- Incident reporting for AI failures with material impact on people or businesses.
The threshold for high-risk is likely to cover AI used in recruitment, credit decisions, healthcare and law enforcement — not general productivity tools like Microsoft Copilot for most businesses. But the boundary is worth watching: as agents begin acting across business systems rather than assisting individuals, the line between productivity tool and high-risk automated decision-making narrows. Businesses deploying agentic AI that touches customer data, financial systems or personal information should assume some of it will eventually fall within scope.
Two practical notes on timing. First, when legislation does land, it is unlikely to start from a standing start — the voluntary guardrails and ASD guidance are the draft of the mandatory rules, so aligning now is the cheapest form of compliance. Second, procurement often moves faster than legislation: government tenders and enterprise contracts increasingly ask suppliers to evidence AI governance, and businesses that cannot answer will lose work before the law formally requires anything.
What should Australian businesses do now?
Three steps give Australian businesses a defensible starting position today — and they map almost one-to-one onto what mandatory regulation is expected to require.
Document AI use. List every AI tool deployed in the business, what data each one accesses, and what decisions it influences. That includes AI features inside platforms you already run: Copilot in Microsoft 365, meeting transcription and summaries in Teams, AI features in your CRM and finance systems. This inventory becomes the input for a risk assessment and is the first thing a regulator, insurer or enterprise customer will ask for.
Apply the voluntary guardrails. For any deployment that affects staff, customers or beneficiaries, apply the five guardrails above: assign named accountability, disclose AI involvement where it affects people, keep a human in the loop for consequential actions, and verify data governance before going live.
Prepare your Microsoft 365 governance. If deploying Copilot or AI agents, complete a data governance review first: scope SharePoint permissions so users can only reach what they should, apply sensitivity labels to sensitive content, configure Conditional Access, and enable Purview audit logging. Mycelium 365's AI strategy advisory and AI readiness assessment cover the technical implementation of these controls.
None of these steps requires waiting for legislation. All three produce artefacts — an inventory, documented controls, audit logs — that satisfy the voluntary standard today and become compliance evidence if mandatory guardrails arrive with a transition period.
How Mycelium 365 helps Australian businesses prepare
Mycelium 365 delivers AI readiness assessments and governance reviews covering the technical controls both the voluntary guardrails and the ASD advisory recommend — permission scoping, sensitivity labelling, Conditional Access configuration, audit logging and human-in-the-loop design for AI workflows. The assessment produces a written report and a prioritised 90-day action plan, so your environment is governed before mandatory regulation arrives rather than after.
Because the review runs on your existing Microsoft 365 environment, the outputs slot directly into the controls you already manage — Entra ID, SharePoint, Purview and Defender — rather than adding a parallel governance layer you have to maintain separately. It is one program of work that satisfies today's voluntary expectations and tomorrow's mandatory ones.
Contact us to discuss your AI governance posture before mandatory regulation arrives.