Many businesses are building valuable AI products on models they do not own, which raises a critical IP question: if a competitor can access the same model, what part of the system can the business actually protect?
We provide two useful downloads for innovators using AI systems below – an IP Health Checklist as well as a guide to patenting AI and machine learning inventions in Australia.
As AI becomes embedded in more products, platforms and processes, businesses are increasingly asking how to protect the value created around these systems.
In many cases, the protectable value may not sit in the model itself. It may sit in the surrounding system, in how:
- data is retrieved
- context is selected
- outputs are checked
- tools are invoked
- the system integrates with customer data or
- it is deployed in a technical or commercial environment.
That is where patent and IP strategy need to become more commercially focused. The question is not simply “can we patent AI?” The better question is: what would a competitor need to copy to achieve the same result?
Why model-focused protection may miss the value
A patent strategy that focuses only on the AI model can miss the part of the system that customers actually use and competitors may want to replicate.
For example, a business may use a widely available large language model but build a proprietary workflow around it. That workflow might retrieve technical documents, filter sensitive information, validate generated outputs, apply business or safety rules, and trigger actions in another system. In that case, the valuable asset might be the way the model is used within a broader system.
This is particularly important for businesses building AI agents, retrieval-augmented generation (RAG) systems, enterprise copilots, diagnostic workflows, automated support tools, industrial monitoring systems or technical decision-support platforms. These systems may appear, at a high level, to be “AI tools”. But the protectable value may lie in the way the tool is architected and deployed.
That matters for more than patent examination. It also affects valuation, due diligence, licensing and defensibility. If a business cannot clearly explain what it owns, what it has built, and what competitors cannot easily copy, its IP position may be weaker than its product story suggests.
The copying-risk test
A useful starting point is to ask a practical question: if a competitor wanted to reproduce the commercial benefit of the AI system, what would they need to copy?
Would they need the same model? The same training data? The same retrieval architecture? The same prompt-generation process? The same validation logic? The same integration with sensors, devices, enterprise systems or customer data?
Those questions help identify where the protectable contribution may sit.
In some cases, perhaps where the model is trained for a specialist technical task, uses proprietary data, meets particular performance constraints, or cannot be substituted easily with a general-purpose model, the answer will be the model itself.
In other cases, however, the more defensible asset may be the workflow that turns a general model into a useful product.
That distinction is important for patent drafting. A specification that says the invention “uses AI to generate improved recommendations” is unlikely to provide a strong foundation for protection. A stronger specification explains what the system does, how the workflow operates, and why that workflow produces a practical result.
Prompts, agents and RAG systems need more than a label
Prompt engineering, AI agents and RAG systems can all be commercially valuable. But the label does not do the patent work.
A claim directed merely to selecting or applying a prompt may look like instructions implemented using ordinary computing tools. A claim directed to an “AI agent” may still look like administrative automation if the agent simply triages emails, updates records or prepares summaries. A claim directed to “using RAG” may still look like information retrieval followed by text generation.
A stronger claim would answer the more detailed workflow questions: how are technical inputs obtained? How is information selected or filtered? How are model outputs validated? What happens if an output fails a safety, reliability or compliance check? Does the system merely present information, or does it change how a technical system operates?
The practical point is simple: the use of an AI agent, prompt workflow or RAG architecture is not enough. The IP strategy needs to identify the part of the system that creates value and can realistically be protected.
Why this matters for investors and transactions
These issues often become important during investment, acquisition or commercial partnership discussions.
A company may describe itself as having proprietary AI technology. But a due diligence team will want to know what that means. Does the company own the model or the training data? Can the model be replaced? Are the key workflows documented? Are they protected by patents, confidential know-how, contracts or a combination of those tools?
If the business relies heavily on a third-party model, that is not necessarily a problem. But it changes the IP analysis. The proprietary value may need to be found in the architecture, integration, data handling, validation process, user workflow or deployment environment.
That is also important for licensing. If the model is not the main proprietary asset, what exactly is being licensed?
Patent strategy and commercial strategy should answer the same question: what is the asset?
Drafting for defensibility, not just patentability
For businesses developing AI workflows, the application should focus on the part of the system that creates defensible value, not just the fact that AI is being used.
AI tools can help prepare technical descriptions, invention summaries and first drafts. But a patent application for an AI or software-based invention is not just a writing exercise. The way the invention is framed can materially affect whether the claims are accepted, whether competitors can design around them, and whether the patent supports the commercial position the business actually needs.
This is where experienced patent advice matters. A good filing strategy should identify what is most likely to be copied, what should be claimed, and what should instead be protected through confidentiality, contracts or controlled know-how.
If the filing is too high-level, it may not protect the part of the system competitors are likely to copy. If it discloses the wrong details, it may educate the market without delivering useful protection. The right balance depends on what the business is commercialising, how the product creates value, and where the durable advantage really sits.
AI IP Health Check
Before disclosing, deploying, licensing or raising capital for an AI system, businesses should ask whether they can explain what they actually own and what competitors would need to copy.
We have prepared an AI IP Health Check to help businesses identify where value may sit in an AI system and whether that value is being protected appropriately.
Related reading
For further context, see our earlier articles on patenting AI and machine learning inventions in Australia: Part 1 and Part 2 or download these below. This article also follows our discussion of IP Australia’s approach to computer-implemented inventions after Aristocrat ’25 and CQMS and the role of hardware in computer-implemented inventions.
To view the full article please click here.
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
