AI6 is the National AI Centre’s Guidance for AI Adoption, published in October 2025. It replaced the 2024 Voluntary AI Safety Standard, condensing its ten guardrails into six essential practices: decide who is accountable, understand impacts and plan accordingly, measure and manage risks, share essential information, test and monitor, and maintain human control. It’s voluntary, it covers organisations that use AI as well as those that build it, and it comes with free templates for an AI policy, an AI register and a screening tool.
Most of the mid-sized companies I talk to are in the same spot. Staff started using ChatGPT, Claude and Copilot before anyone wrote a policy. Vendors switched AI on inside software the company already pays for. Then the board, a big customer’s security questionnaire or the insurer asked what the company’s position on AI is, and nobody had a good answer.
AI6 is the best answer available in Australia right now. Not because it’s law, but because it’s the government’s own statement of what reasonable AI governance looks like. When someone asks whether you govern AI properly, “we’ve implemented the AI6 practices” is an answer they can check.
Two versions, and which one you need
The guidance comes in two versions. Foundations is short, written for organisations early in their AI adoption, and it’s where every mid-sized company should start. Implementation practices is the long version, for governance professionals, technical teams and anyone building AI or using it in high-risk settings. It’s aligned with ISO/IEC 42001 and the NIST AI Risk Management Framework, and a crosswalk maps it to both.
The guidance also makes a point that’s easy to skim past: governance happens at two levels. Some practices apply to the whole organisation, like having an AI policy. Others apply to each AI system in the way you use it. The guidance’s own example is ChatGPT: drafting marketing emails and assessing job applications are the same tool and very different risks. Each use case needs its own check.
The six practices, and what “done” looks like
Each practice has a “getting started” list and a “next steps” list. The numbers below are the guidance’s own. For a company of a few hundred people, here’s what I’d accept as done for the getting-started baseline.
| Practice | Getting started | Done looks like |
|---|---|---|
| 1. Decide who is accountable | 1.1 senior owner; 1.2 AI policy | One named executive owns AI. A two-page policy staff have actually been given. |
| 2. Understand impacts and plan accordingly | 2.1 stakeholder impact assessment; 2.2 contestability channels | A one-page impact assessment for each use case that touches people, and a way to challenge AI decisions. |
| 3. Measure and manage risks | 3.1 risk screening process | New AI tools and vendor AI features get screened before use, including the ones nobody asked for. |
| 4. Share essential information | 4.1 AI register; 4.2 disclose AI use | A register of every AI system, including AI embedded in SaaS. Disclosure where AI shapes decisions about people. |
| 5. Test and monitor | 5.1 ask suppliers for proof; 5.2 test before deploying; 5.3 monitor after; 5.4 extend data and security practices | Supplier test evidence on file, acceptance testing for your use case, someone watching outputs, and DLP and logging covering AI tools. |
| 6. Maintain human control | 6.1 meaningful human oversight; 6.2 human override points | Human review before high-stakes outputs take effect, and a documented way to pause or switch off each system. |
Where companies actually get stuck
The register is always bigger than they think. Practice 4.1 says the register should cover AI embedded in other systems, like HR and customer engagement tools. That’s where most of it is. Copilot in Microsoft 365, lead scoring in the CRM, CV ranking in the recruitment platform, the helpdesk’s auto-triage, fraud scoring in the payment gateway. A company that says “we only use ChatGPT” usually has a dozen AI features switched on that nobody chose.
Accountability stops at the committee. Practice 1.1 asks for a senior leader. Practice 1.3 then asks for a named person for every AI system. A steering committee that meets quarterly is neither. If something goes wrong with the CRM’s lead scoring on a Tuesday, someone specific has to own it.
Screening only catches what gets bought. Practice 3.1’s screening process usually gets bolted onto procurement, which misses every AI feature that arrives in a vendor update. The screen needs to sit in change management as well.
Agentic AI gets treated like a chatbot. Practice 3.2 is explicit that risk management should account for the differences between traditional IT, narrow AI, general-purpose AI and agentic AI. An agent that can send emails, change records or call APIs through MCP connectors carries risks a chat window doesn’t. Practices 6.1 and 6.2 matter most here: oversight scaled to how much the system can do on its own, and an off switch someone has actually tested.
Disclosure arrives with a deadline. Practice 4.2 is voluntary. The Privacy Act’s automated decision rules aren’t. From 10 December 2026, your privacy policy has to describe automated decisions that significantly affect people. If you’re building the AI register anyway, you’re most of the way there. I’ve written up what the new privacy rules require.
A 30-day AI6 baseline
The guidance says you don’t need to implement everything at once: get started across all six practices, then add more as your AI use grows. Here’s the order I’d do the getting-started baseline in, for a company of 100 to 1,000 people.
- Week 1. Name the AI governance owner. Build the AI register from licences, SaaS admin consoles, expense claims and a short staff survey, with an owner for each system.
- Week 2. Write the AI use policy. Add an AI screening step to procurement and change management.
- Week 3. Run short impact assessments for anything touching customers, staff or decisions about people. Put AI risks in the risk register you already run. Draft the automated-decision wording for the privacy policy.
- Week 4. Set up monitoring and override points for the higher-risk systems, extend DLP and logging to AI tools, and take a one-page summary to the board.
None of that needs new software. It needs someone with the authority to ask awkward questions, and a few hours a week from each system owner.
If you need ISO/IEC 42001 later
Most mid-sized companies don’t need ISO/IEC 42001 certification yet. Some will, usually because an enterprise customer or a government buyer starts asking. The good news is that the Implementation practices were built to align with 42001, so the register, the policy, the risk assessments and the monitoring records are the same evidence an auditor will want. And if you already run an ISO 27001 management system, 42001 sits on top of it rather than beside it.
AI6 isn’t hard. It’s just the kind of work that never gets done until somebody owns it. Start with the register: you can’t govern what you can’t list.
Questions people ask
Is AI6 mandatory?
No. The Guidance for AI Adoption is voluntary. But existing laws already apply to how you use AI, including the Privacy Act, consumer law, anti-discrimination law and directors’ duties, and from 10 December 2026 privacy policies must disclose automated decisions that significantly affect people.
What happened to the Voluntary AI Safety Standard?
AI6 is its first update. The National AI Centre condensed the standard’s ten guardrails into six essential practices, removed redundant language, and widened the audience to include developers as well as organisations that deploy AI.
Does AI6 apply if we only use tools like ChatGPT, Claude or Copilot?
Yes. The guidance applies to organisations that use AI, not just those that build it. Its own example is ChatGPT: the same tool needs different governance when it drafts marketing emails than when it helps assess job applications.
Where do I get the official templates?
The National AI Centre publishes an AI policy template, an AI register template and an AI screening tool alongside the guidance on industry.gov.au and ai.gov.au. They’re free and a sensible starting point.
How does AI6 relate to ISO/IEC 42001 and the NIST AI RMF?
The Implementation practices draw on and align with ISO/IEC 42001:2023 and the NIST AI Risk Management Framework 1.0, and the National AI Centre publishes a crosswalk between them. Work done for AI6 counts towards either.
Sources: National AI Centre, Guidance for AI Adoption: Foundations and Implementation practices (October 2025). This article is general information, not legal advice.