Australia’s medicines regulator has told the clinical AI industry that the technology inside a product is not what makes it a regulated medical device. What matters is what the maker says the software is for.
On 5 February 2026 the Therapeutic Goods Administration (TGA) published updated guidance on how AI-enabled software is regulated as a medical device, resolving a question that had left startups and hospitals unsure whether their tools sat inside or outside the therapeutic goods framework. The core message is that regulation is triggered by the manufacturer’s intended purpose, not the presence of AI. A diagnostic model and a spreadsheet macro are judged by the same test: what clinical job is the product built and marketed to do.
The distinction is not academic. If software is intended to diagnose, monitor, predict, prevent or treat a disease, injury or disability, it is a medical device under Australian law, and in most cases it must be entered on the Australian Register of Therapeutic Goods (ARTG) before it can be supplied. Get the classification wrong and a product can be pulled from the market.
Intended purpose is the trigger, not the algorithm
The TGA’s framework is deliberately technology-agnostic and risk-based. As law firm King & Wood Mallesons put it in its analysis of the guidance, the rules are “triggered by the manufacturer’s intended purpose, not the presence of AI features”. A large-language-model chatbot that summarises clinical notes and a conventional rules engine that flags drug interactions are assessed on the same basis.
Crucially, the assessment is objective. Lawyers at Allens, in an analysis titled “Same rules, smarter tools”, note that intended purpose is inferred from a product’s design, its marketing and how it is actually used, not merely from what a disclaimer says. A vendor cannot dodge regulation by labelling a diagnostic tool as “for information only” if everything about the product points to clinical use.
Where software is intended to influence, inform or replace a clinical decision, organisations should expect it to fall inside the framework and carry the obligations of a sponsor or manufacturer. That includes conforming with the Essential Principles, holding evidence to prove it, and running systems to monitor and report safety incidents and adverse events.
Standards, evidence and the problem of scope creep
The guidance sets out what compliance looks like in practice. Manufacturers must assess intended use and risk class and demonstrate conformance with recognised international standards, including IEC 62304 for software life-cycle processes and ISO 14971 for risk management, according to a summary by regulatory consultancy Pure Global. Related standards cover usability, quality systems and cybersecurity. Manufacturers are also expected to document how their models were trained and tested, assess data quality, and hold clinical evidence that the software performs as claimed.
The sharper theme running through the guidance is what happens after a product ships. Adaptive AI systems change, and the TGA warns against “scope creep”, where updates quietly expand what a tool does. King & Wood Mallesons notes that a software update which alters performance, or which turns unregulated software into a regulated device, needs approval before it is deployed. Updates, in other words, are regulatory events, not routine engineering.
Allens frames the same point bluntly: a sponsor’s responsibility for a device does not end at deployment. Manufacturers are expected to watch for off-label use and step in when they find their tools being used outside the approved purpose. The firm also flags a warning that will sting some developers: synthetic data is not a substitute for real-world data when it comes to satisfying clinical evidence requirements.
Why it matters for Australia
Regulation is the make-or-break gate for every clinical AI product sold in Australia. Until now, ambiguity over whether a tool counted as a medical device had stalled deployments, with hospitals and startups reluctant to commit to products of uncertain legal status. The guidance replaces that uncertainty with a clear test, which is exactly what a nascent local health-AI sector needs to raise capital, sign hospital contracts and plan a path to market.
The catch is that clarity cuts both ways. Companies that assumed a “decision-support” or “wellness” framing kept them outside the rules may now find themselves squarely inside, facing ARTG listing, conformity assessment and post-market surveillance obligations they had not budgeted for. Building regulatory assessment into product development from the start, as Allens urges, is cheaper than retrofitting it after a launch.
There is also a strategic dimension. Both Allens and Alignmt AI have observed that the TGA’s approach signals a direction that US and EU regulators may follow, positioning Australia as an early mover whose framework global peers are watching. For a country that often imports its tech rules, setting a reference point on clinical AI is a rare and valuable position, and one that could make Australian-developed tools easier to export if overseas regimes converge on the same intended-purpose logic.
The near-term test will be enforcement and interpretation. The guidance answers the threshold question of when AI software is regulated, but the harder calls, such as how much a model can adapt before it needs fresh approval and how aggressively the TGA polices scope creep, will be worked out product by product. King & Wood Mallesons advises developers to stay alert for further TGA guidance. For Australia’s clinical AI builders, the gate is now clearly marked. Walking through it on time is the next challenge.
Sources: Therapeutic Goods Administration, Allens, King & Wood Mallesons, Pure Global.
