Why this matters for your business right now

If you run a business in Australia and you are using AI in any serious way, you have probably noticed the rules are catching up faster than anyone warned. Through 2025 and into 2026, regulators across Canberra, Sydney, and Melbourne have moved from position papers to actual enforcement. The framework is no longer theoretical. The question is no longer whether AI will be regulated in Australia. The question is whether the way you are using it today will pass scrutiny tomorrow.

I work with business owners across both sides of the Tasman, and the AU side of the conversation has a different texture. New Zealand tends to lean on the Privacy Act 2020 and its 13 Privacy Principles, with PP12 on offshore disclosure doing most of the heavy lifting. Australia has a thicker stack of regulators, each with their own patch of the AI risk landscape. ASIC watches financial services and credit. APRA watches banks, insurers, and super funds. AHPRA watches anything that touches patient information. AHRC weighs in on discrimination. The OAIC handles privacy generally. If your business touches more than one of those areas, which most do, you have multiple rulebooks running at once.

Below is a working map of what is in play, what is coming, and what the practical moves are for an AU business owner who wants to use AI to grow without inviting a regulatory headache.

The shape of the Australian AI regulatory framework in 2026

Australia does not yet have a single AI Act. The government has been deliberate about that. The approach in 2026 is layered, sector-by-sector, with a baseline of existing law plus new mandatory guardrails for high-risk settings.

The headline piece is the framework set out through the Department of Industry, Science and Resources. Industry estimates suggest around AUD 6 to 8 million organisations across the economy will be touched by AI risk obligations by the end of 2027. Whether that exact figure lands or not, the trajectory is clear. The government has introduced mandatory guardrails for high-risk AI systems. Those guardrails cover transparency, testing, data quality, human oversight, and contestability. If you are deploying AI in ways that affect housing, employment, credit, insurance, justice, or migration decisions about a person, you are almost certainly in scope.

For most of the business owners I speak with in Sydney, Brisbane, and Perth, the day-to-day exposure is less dramatic but still real. You might be using AI to screen CVs from Seek. You might be using it inside Xero or MYOB to flag anomalies. You might be routing customer conversations through a chatbot. Each of those carries a different weight under the framework. The smart move is to understand which weight applies to which use case before a regulator or a journalist asks the question for you.

ASIC and the financial services patch

ASIC has been the most aggressive of the Australian regulators on AI, and the most useful to understand first because its guidance tends to bleed into other sectors.

Regulatory Guide 265 remains the centrepiece. It sets out what ASIC expects of AFS licensees and credit licensees when they use AI or machine learning in their operations. The relevant obligations are the AFS licensee obligations under section 912A of the Corporations Act, the obligation to do all things necessary to ensure financial services are provided efficiently and honestly, and the design and distribution obligations for any product that uses AI in a way that materially affects consumers.

The practical impact for business owners is that ASIC has made it clear it will treat an AI-driven bad outcome the same way it treats a human-driven bad outcome. If your AI model denies a customer a credit limit and that denial is wrong, that is on you. If your AI model gives financial advice through a tool and it falls short of the reasonable knowledge and care standard, that is on you. Hallucination is not a defence. Vendor marketing is not a defence. The phrase I keep hearing from the Sydney compliance circles I move in is that AI is treated as an actor inside your business, not a piece of software you bought.

If you are an AFS or credit licensee, RG 271 on internal dispute resolution and RG 272 on reporting also interact with AI deployment. You need to be able to explain decisions, capture the data inputs, and route complaints into a human review path. If your current AI tooling cannot produce a decision log on demand, that is a gap. Verify any specifics on RG 271 and RG 272 reporting timelines with your lawyer or advisor, as ASIC does update them.

For business owners outside the financial services patch, ASIC’s example still matters. Many of the obligations that ASIC requires under RG 265 are becoming the de facto standard that AHRC, OAIC, and APRA expect when they review AI usage.

APRA CPS 234 and the data discipline layer

APRA’s CPS 234 on information security has been on the books since 2019, but in 2026 it is the data backbone of any AI risk story. The standard requires APRA-regulated entities, which includes authorised deposit-taking institutions, general insurers, life insurers, registrable superannuation entities, and licensed central counterparties, to maintain an information security capability commensurate with the size and extent of threats to those information assets.

The reason CPS 234 matters for AI is that an AI model is an information asset. The training data is an information asset. The prompts and outputs are information assets. The model weights, if you are fine-tuning, are information assets. APRA has signalled through its prudential practice guides that the standard covers all of these.

If you are a bank, insurer, or super fund, you are subject to CPS 234 directly and you have obligations around third-party data arrangements. If you are a vendor to those entities, you have indirect exposure because they will demand contractual terms that flow CPS 234 obligations down to you. We typically see APRA-regulated entities inserting AI-specific clauses into supplier contracts throughout 2026, particularly where the supplier’s AI system touches customer data or material information assets.

If you are a small or medium AU business that is not APRA-regulated and not a vendor to one, CPS 234 is still worth knowing. It is the cleanest articulation of how a regulator expects you to handle sensitive data inside an AI workflow, and it has become a reference point that AHRC and OAIC reach for when auditing non-financial businesses.

AHPRA and anything that touches a patient

If your business handles health information, or supplies a clinician or a clinic, AHPRA’s codes and the underlying Health Privacy Principles under most state and territory health records legislation apply. The Sixteen Principles under the Privacy Act 1988 are the baseline, including APP 6 on use and APP 11 on security, but healthcare carries additional state-level overlays in NSW, Victoria, and Queensland that are tighter.

In 2026 the question most healthcare-adjacent business owners are asking is whether they can use a general-purpose AI tool, or even a healthcare-specific one, to draft clinical notes, summarise patient records, or assist with triage. The short answer is yes, with controls. The longer answer is that you need an AI policy that defines the use case, a data protection impact assessment, a logging layer, a human-in-the-loop requirement for any clinical decision, and a clear statement that the AI is not the clinical decision-maker. A Sydney practice manager I spoke with recently described the workload as six weeks of policy writing followed by two days of tool configuration. That order matters.

Verify with your lawyer or advisor on specific obligations under the AHPRA codes and the relevant state or territory health records legislation, as the regulatory detail here shifts quickly.

The privacy layer under the Privacy Act 1988

Privacy is the floor that everything else sits on. The Australian Privacy Principles under the Privacy Act 1988 cover collection, use, disclosure, storage, and cross-border transfer of personal information. They are technology-neutral on their face, but in practice they have become the tool the OAIC uses to assess AI deployment.

APP 3 on collection and APP 6 on use are where most AI projects get tripped up. If you are collecting personal information to feed into an AI model, the collection has to be reasonably necessary for your functions or activities. If you are using personal information to train a model or to generate outputs about an individual, that use has to fall within the purposes you told the individual about, or within a permitted general situation.

Cross-border disclosure is the specific point where AU and NZ diverge. Under New Zealand PP12, you can disclose personal information offshore if the recipient is subject to a comparable privacy regime or if the individual is authorised. Under the Australian APP 8, you must take reasonable steps to ensure the overseas recipient does not breach the APPs, and you remain liable for that recipient’s conduct. If you are using a US-hosted AI tool that processes the personal information of Australian customers, APP 8 is doing real work.

High-risk use cases that regulators are watching in 2026

Across ASIC, APRA, AHPRA, and AHRC, the use cases that draw the most attention are predictable. Employment screening tools that rank candidates from Seek. Credit decisioning tools used by lenders and fintechs. Tools that triage or deny access to government services. Insurance pricing and claims handling. Biometric identification. Content moderation at scale. Each of these sits inside a specific regulator’s remit.

Below the high-risk threshold, the regulators are still watching but the enforcement priority is lower. Marketing copy generation, internal summarisation tools, code assistance, customer service agents with a human handoff, all of these are live and in use across thousands of AU businesses. They are not the targets of the moment. The risk for a small business is not that the regulator knocks on your door because you used AI to draft a tender response. The risk is that you build a workflow that quietly tips into a high-risk category, and you did not notice the line being crossed.

Practical moves for an AU business owner in 2026

The first move is to write an AI policy that fits your business, not one copied from a US vendor template. The policy should list your approved tools, your prohibited uses, your data classes, and your human-in-the-loop requirements. For a business of under 20 staff, this is a two-page document. For a business of 200, it is a 15-page document with appendices. Both are useful.

The second move is to do an AI inventory. List every tool that touches your customer data or your staff data, note what it does, where the data goes, who the vendor is, and whether the data leaves Australia. We typically see AU businesses run this exercise and find three or four tools they had forgotten about, often in marketing and HR.

The third move is to put logging and decision-record capability in place for any AI workflow that affects a customer. This is where RG 265 and CPS 234 habits cross over. If your AI tool cannot produce an audit trail on request, raise that with the vendor. If they cannot, find a different vendor.

The fourth move is to check your contracts. AI-as-a-service contracts vary wildly on liability, data ownership, and disclosure. Vendor terms drafted in 2023 routinely give the vendor rights to your inputs for model improvement. That is not acceptable under APP 6 for most use cases.

The fifth move is budget for the work. For an AU small business the policy and inventory work runs about AUD 5,000 to 15,000 with an external advisor, give or take. For a mid-market business, industry estimates suggest the figure lands between AUD 40,000 and 120,000 once you include tooling upgrades and staff training. Those numbers are approximate, and the range is wide because every business is different. Treat them as a starting point for a conversation with your advisor.

Finally, keep watching the horizon. The mandatory guardrails under the high-risk framework are likely to be enacted through 2026 and 2027, with compliance dates phasing in. The exact start dates and the exact definition of high-risk use cases have been subject to consultation and may have shifted by the time you read this. Verify the current status with your lawyer or your industry association.

What good looks like for the next 12 months

A business that handles AI regulation well in 2026 treats it the way a finance team treats BAS. Not exciting, not optional, but built into the operating rhythm. The team knows the rules. The tools are inventoried. The contracts are clean. The outputs are logged. When ASIC, APRA, OAIC, AHRC, or AHPRA asks a question, the answer is already in the file.

If you are starting from scratch, expect six to ten weeks of focused work to get the foundations in place. If you already have data governance and privacy compliance, you are looking at three to four weeks to extend those practices into the AI layer. Either way, the work is finite and the upside is real. The businesses that get this right will be the ones whose AI deployments scale, and whose competitors struggle to justify theirs.

Enterprise DNA works with NZ and AU businesses on this challenge every week, from sole traders through to multi-office practices. Book a 60-min Omni Audit — https://calendly.com/sam-mckay/discovery-call?utm_source=edna-landing&utm_medium=blog&utm_campaign=nzau