What does your AI business actually own? Finding the value around the model

More and more Australian businesses are building products on AI models that belong to someone else: a large language model reached through a paid interface, or an open-weights model downloaded and adapted. That raises a question that founders, investors and patent attorneys keep arriving at. If a competitor can use the same model tomorrow, what exactly is yours? This note sets out our view of that question for Australian technology businesses.

Where does the value in an AI product actually sit?

For many AI products, the value is not in the model. It sits in the system wrapped around it. Two companies can license the same model and deliver very different products, because one has engineered how the right information reaches the model, how its answers are checked before anyone relies on them, and how the whole thing connects to a customer's data, equipment and daily processes. That engineering is what customers experience and what a rival would have to reproduce.

Think of a document-review tool that uses a general-purpose model. The model is rented. The commercial advantage might be the way the tool pulls the relevant technical documents, strips out confidential material, tests each answer against business rules, and then triggers an action in another system. A patent strategy aimed only at "our AI" can miss all of that.

What would a competitor actually have to copy?

A useful test is substitution. Imagine a competitor who wants the same commercial result as you. List what they would have to reproduce: the model, the data it was trained on, your method of retrieval, your sequence of checks and your connections to sensors, devices or customer systems. Anything they can simply buy or download is a weak candidate for protection. Anything they would have to design, test and learn the hard way is where your defensible position is likely to be.

Sometimes the answer really is the model. A model built for a narrow technical job, using proprietary data and strict performance limits, and not easily replaced by a general one, can be the asset. The point of the exercise is to find out, rather than assume, which kind of business you are running.

Why are labels like "AI agent" or "prompt engineering" not enough?

A label does no legal work. Patent examiners in Australia ask whether a claimed computer-implemented invention is more than an abstract idea or the ordinary use of a computer. A claim that amounts to choosing a clever prompt can look like instructions given to a general tool. A claim to an "agent" that sorts emails or tidies records can read as administration done faster. A claim to "retrieval-augmented generation" can read as looking something up and then writing text about it.

The more persuasive position answers the questions an examiner will ask. Where does the technical input come from, and how is it screened before the model sees it? How are answers tested before anyone acts on them? What does the system do when a check fails? And does the result only inform a person, or does it alter the operation of a machine or process? IP Australia has updated its guidance on computer-implemented inventions following recent Federal Court decisions in this area. The practical lesson for a business is to describe the workflow, not the buzzword.

Why does this matter beyond patent examination?

Because it decides what you can say you own. In a funding round, an acquisition or a licensing negotiation, a company that calls its technology "proprietary AI" will be asked what that means. Who owns the model and the training data? Could the model be replaced next quarter? Are the key workflows written down? Are they protected by patents, by confidentiality, by contract, or by a combination? Relying on a third-party model is not a problem in itself. It does mean the proprietary value has to be found somewhere else: the architecture, the integration, the handling of data, the validation process or the deployment environment. It also raises a licensing question. If the model is not your main asset, what exactly are you offering to license?

How should an AI patent application be drafted?

Draft for defensibility, not just for a grant. A specification that says only that the invention applies artificial intelligence to give better recommendations gives an examiner little to work with, and a court even less. The better approach is to describe the steps the system performs, in what order, and the concrete benefit that follows from them. It also records the alternatives, so that a competitor cannot avoid the claims by changing one component.

AI tools can help with first drafts and invention summaries, and we use them ourselves. But framing the invention is a legal and commercial judgment. It influences whether claims are accepted, how easily a rival can work around them, and whether the patent backs the commercial position the business is trying to hold. A filing that is too high-level may not cover what competitors are likely to copy. A filing that reveals the wrong detail can teach the market something without giving you protection in return. Some elements are better held as confidential know-how or protected through contracts, and a good strategy decides which is which before anything is published.

What should an Australian business do now?

Before you disclose, deploy, license or raise capital on an AI system, work through five steps. First, write down what a competitor would need to copy. Second, identify which of those things you actually own, and which you merely use under someone else's terms. Third, decide which elements are suited to patent protection, which to confidentiality and which to contract. Fourth, check that your people and contractors have assigned their rights to the company. Fifth, avoid public disclosure of the technical workflow until the patent position is settled. Working through those five steps with an adviser before launch is far cheaper than trying to repair the position during due diligence.

Stellar IP Law advises technology businesses on patents, trade marks, designs and IP strategy for software and AI-enabled products. We work with clients across Sydney, Gold Coast, Brisbane, Sunshine Coast, Newcastle and Wollongong. Contact us to discuss what your business owns, and what a competitor could copy, before your next launch or funding round.