Artificial intelligence is revolutionising how we approach clinical research in Australia. Today, AI systems are diagnosing leukaemia, matching lung cancer patients to appropriate clinical trials, and analysing medical imaging with remarkable accuracy. Walk into any major Australian clinical setting, and you’ll find AI tools calculating sepsis risk in emergency departments and predicting cancer outcomes.
Here’s the challenge, though: While innovation is accelerating, so is regulatory scrutiny.
The R&D Tax Incentive (or R&DTI) is the Australian government’s flagship program for supporting companies investing in research and development. It’s designed to de-risk innovation by providing a tax offset, essentially giving you money back for doing R&D that you might not have attempted otherwise. For companies with an annual turnover below $20 million, this can mean a refundable tax offset of up to 43.5% of eligible R&D expenditure. That’s real cash back, not just a reduction in tax owed, making it easier to invest in R&D and fuel economic growth in the sector.
But there’s a catch. The Australian Taxation Office (ATO) and AusIndustry, the two government bodies that administer the R&DTI, have specific concerns about R&D claims in the software development sector. These concerns are so significant that they’ve been formalised in Taxpayer Alert TA 2017/5, which essentially says “we’re watching this area closely.” Industry groups like FinTech Australia are even urging the government to ensure digital technologies like AI are explicitly recognised and protected in R&D policy discussions.
For clinical research strategists leading AI and machine learning projects, this creates a complex situation. Your work is genuinely innovative and exactly the kind of research the R&DTI was designed to support. But it also sits in a regulatory grey zone that requires careful navigation. Understanding the compliance landscape isn’t optional; it’s essential for protecting your claim and accessing the funding you need to keep innovating.
So what’s actually triggering these audits? And more importantly, how can you structure your claims to withstand scrutiny? Let’s break down what you need to know.
Why Your AI and Machine Learning Projects Are Under the Microscope
AI and machine learning projects attract more audit attention than most other types of R&D claims. This isn’t because regulators think you’re doing anything wrong; it’s because these projects sit at the intersection of several complexity factors that make them harder to assess. Understanding why helps you structure your work to address these concerns upfront.
The Blurry Line Between Genuine Experimentation and Smart Business Application
Here’s the fundamental challenge: the R&DTI requires you to clearly distinguish between Core R&D Activities, the actual experimental work, and everything else. “Core R&D” means the central experiment you’re conducting to resolve an unknown. “Everything else” includes supporting tasks, routine business operations, and the general work of building a product or service.
In AI projects, this boundary is often ambiguous. Let’s say you’re building a machine learning model to predict patient outcomes in a specific therapeutic area, work that’s advancing your organisation’s AI capabilities. Some parts of that project might qualify as AI R&D, while others definitely won’t. The key test is whether your work addresses what’s called technical uncertainty.
So what exactly is technical uncertainty? It means you’re trying to achieve something where the solution is genuinely unknown, not just unknown to you or your team, but unknown to the broader scientific community. The outcome cannot be determined in advance by a competent professional in the field, even if they had all your resources and expertise.
Critically, the uncertainty must be technical, not commercial. Commercial risk, rather than technical uncertainty, doesn’t qualify. Here’s the difference: “Will customers want this feature?” is a commercial question. “Can we develop a neural network architecture that maintains high predictive accuracy with significantly less labelled training data for rare disease classification?” is a technical question.
This distinction trips up many organisations. Using established AI tools to solve a business problem, even if it’s a clever solution, doesn’t automatically qualify as R&D. Training a standard convolutional neural network on your medical imaging dataset might be valuable work, but if you’re using well-understood techniques in a standard way, it’s likely a routine application. On the other hand, developing novel techniques to handle severely imbalanced medical datasets, or creating new approaches to improve model interpretability for clinical decision-making in high-stakes environments, could qualify as genuine R&D.
Proving You’re Creating New Knowledge, Not Just Finding New Uses
There’s another hurdle: Regulators scrutinise whether your work generates fundamentally new knowledge or is simply not just a new application of existing knowledge.
This is particularly tricky in clinical AI research. You might be using machine learning in a therapeutic area where it’s never been applied before. That sounds innovative, right? And it is, from a clinical and business perspective. But for R&D tax purposes, applying known ML techniques to a new domain doesn’t necessarily constitute R&D. What matters is whether you’re generating new knowledge about how to solve technical problems, not whether you’re the first to apply existing solutions in your specific context.
There’s also a classification issue that catches many organisations off guard. Correct classification of your work is crucial. The Australian and New Zealand Standard Research Classification (ANZSRC) includes a code for “Artificial Intelligence” (4611), but here’s the thing: That code is specifically for research into the underlying techniques of AI itself, developing new algorithms, new approaches to machine learning, and fundamental advances in AI theory.
If you’re using AI as a tool to advance clinical research, your project should be classified under the relevant scientific domain, such as “Biomedical and Clinical Sciences.” This might seem like a technicality, but getting the classification wrong can undermine your entire claim by suggesting you don’t understand what type of research you’re actually conducting.
The Documentation Burden of Data-Driven Research
The R&DTI requires you to demonstrate a systematic progression of work, a clear journey from hypothesis and experiment to observation and conclusion. In traditional R&D, this might look like laboratory notebooks documenting each stage of an experiment. In software development, it needs to be adapted to agile frameworks and the iterative nature of how development teams actually work.
The challenge is that many AI and ML projects operate in a highly iterative, exploratory mode. Data scientists might try dozens of different approaches, tune hundreds of hyperparameters, and pivot quickly based on what they discover. That’s how good data science works. But for R&DTI purposes, you need to impose some formal structure on that process, documenting hypotheses before sprints begin, recording what you tried and why, and capturing results (including failures).
There’s one more critical requirement: Establishing the knowledge gap. You must prove that the outcome you’re seeking was unknown based on existing global knowledge. AusIndustry expects you to document this knowledge gap through a process of preliminary and secondary literature reviews. In practice, this means that before you start your R&D work, you need to have conducted and documented a proper search of academic literature, industry publications, and other sources to demonstrate that what you’re trying to achieve hasn’t already been solved elsewhere.
If you discover halfway through your project that someone else has already published a solution to your exact problem, that’s valuable information, but it also means that work can no longer be claimed as R&D, because the technical uncertainty has been resolved (just not by you).
Now that you understand why AI projects attract scrutiny, let’s look at the specific mistakes that turn that scrutiny into problems.
The Audit Traps That Catch Even Experienced Teams
Understanding the common pitfalls can help you avoid them. These are the issues that consistently trigger audits and lead to claims being reduced or rejected entirely.
Claiming Everything (When You Should Be Surgical)
The most significant red flag you can raise is making what regulators call a “whole of project claim”. This is when companies make the mistake of lodging their entire project as R&D, treating every hour spent and every dollar invested as eligible.
Here’s why that’s problematic: Almost no project is 100% R&D. Even if your overall goal is innovative, most projects include a substantial amount of work that doesn’t meet the R&DTI’s strict definition of experimental research.
Think about a typical clinical AI project. You might spend time on:
- Conducting literature reviews (could be supporting R&D)
- Developing novel machine learning architectures (likely core R&D)
- Building data pipelines using standard tools (routine work)
- Designing user interfaces for clinicians (not R&D)
- Project management and coordination (not R&D)
- Writing documentation (depends on what it documents)
- Deploying to production infrastructure (routine work)
- Training end users (definitely not R&D)
Regulators expect you to dissect the project with surgical precision and register only the specific activities that meet the eligibility criteria. Yes, this is more work. But it’s also more defensible. A carefully scoped claim that clearly articulates which specific activities were experimental and why is far more likely to survive audit than a blanket claim for the whole project.
When Your Records Tell the Wrong Story (or No Story at All)
The lack of contemporaneous records, meaning records created at the time the work is done, is one of the most common reasons for rejection.
Here’s the key principle: Records must be created when the R&D is actually conducted, not reconstructed months later when you’re preparing your claim. If an auditor suspects you’ve created documentation after the fact to justify a claim, that’s a significant problem. It raises questions about whether the R&D was actually conducted the way you’re describing, or whether you’re retrofitting a narrative to fit the R&DTI requirements.
So what kind of records should you be maintaining? Evidence should include:
Literature reviews with dates: Document what you searched, where you searched, what you found, and what gaps remain. Do this before you start the experimental work.
Formally documented hypotheses: Before each sprint or major work phase, write down what you’re trying to achieve and why it’s uncertain. This doesn’t need to be a PhD thesis; a clear paragraph explaining “We hypothesise that X approach will achieve Y outcome because of Z reasoning” is sufficient.
Version control logs: Your Git commits (or equivalent) should tell a story. Commit messages like “tried CNN approach” or “testing LSTM for time series prediction” provide evidence of systematic experimentation.
Technical notes in project management tools: Implement Jira, Asana, or whatever tool your team already uses, and ensure technical challenges and solutions are documented in tickets and comments.
Test results, especially failures: It’s not just about what worked. Documenting what you tried that didn’t work is crucial evidence that you were dealing with genuine technical uncertainty.
Detailed timesheets: You need a time tracking tool that captures both R&D and non-R&D activities for everyone working on the project. This allows you to demonstrate exactly who worked on eligible activities and for how long.
All these records must be retained for five years. That’s five years from the date you lodge your claim, not from when you did the work.
Inflating Claims and Including Costs That Don’t Qualify
Another common problem is overstating claims by including ineligible costs. The R&DTI has specific rules about what you can claim, and including non-R&D costs is a fast track to an audit.
Common errors include:
- General business overheads that aren’t directly connected to the R&D
- Marketing and sales expenses
- Salaries for staff who didn’t actually work on the R&D activities
- Administrative costs
- General office expenses
The rule is simple but strict: all claimed costs must have a sufficient and direct link to your registered R&D activities. If you can’t clearly explain why a particular cost was necessary for the R&D, it probably shouldn’t be in your claim.
Some costs are legitimately shared between R&D and non-R&D activities. Your office rent, for example, or cloud computing costs, or certain equipment. For these situations, you need to use a reasonable and defensible apportionment method, a consistent way of splitting the cost between eligible and ineligible activities.
The method you choose needs to be documented and defensible. For instance, if your data science team uses 60% of their time on eligible R&D activities (as evidenced by timesheets), you might apportion 60% of their workstation costs and software licenses to the claim. What you can’t do is use unreasonable apportionment, applying a methodology that you know inflates the claim beyond what’s genuinely attributable to the R&D.
Why Getting This Right Opens Doors Beyond Tax Savings
Here’s something many organisations don’t fully appreciate: Getting your R&D compliance right isn’t just about avoiding audits or maximising your tax refund. Rather, it’s a strategic advantage that can strengthen your entire funding position.
Transforming Uncertainty Into Predictable Cash Flow
When you have robust, audit-ready documentation and processes, the R&D Tax Incentive stops being an uncertain windfall and becomes a reliable source of non-dilutive cash flow, meaning funding that doesn’t require you to give up equity or control of your organisation.
This matters enormously for financial planning. If you’re confident in your ability to self-assess if your R&D activities are eligible, you can incorporate expected refunds into your financial forecasts and budgets. For clinical research organisations that often operate with tight margins and long development timelines, this predictable R&D finance can be the difference between maintaining momentum and having to slow down innovation.
Think about it this way: If you’re conducting eligible AI R&D worth $1 million in expenditure during a financial year, a 43.5% refundable offset means you can expect $435,000 back. That’s not trivial funding. If you’re confident in your claim structure and documentation, you can factor that $435,000 into your cash flow planning for the year ahead. You can commit to research initiatives knowing the funding will be there. That confidence only comes from doing the compliance work properly.
Using Compliance as a Forcing Function for Research Excellence
There’s an unexpected benefit to the R&DTI’s strict documentation requirements: they can actually improve the quality of your research.
When you implement systematic experimentation and follow the required systematic progression of work, you’re not just satisfying regulators, you’re compelling your teams to adopt more rigorous, scientific methods. The act of writing down hypotheses before experiments, documenting what you tried and why, and analysing results systematically makes you a better researcher.
Many organisations find that the discipline required for R&DTI compliance actually helps them:
- Catch flawed assumptions earlier in the process
- Avoid pursuing dead ends for too long
- Build institutional knowledge that persists even if team members leave
- Replicate successful experiments more reliably
- Communicate technical achievements more clearly to stakeholders
This isn’t compliance for compliance’s sake; it’s using regulatory requirements as a framework for research excellence.
Making Your R&D Documentation Work Harder
Here’s another advantage: The detailed technical documentation you prepare for an R&DTI claim is directly reusable for other government grants and funding opportunities.
Grant applications typically require detailed technical descriptions, evidence of innovation, demonstration of capability, and clear milestones. If you’ve already documented all of this for your R&D claim, you’re not starting from scratch when a relevant grant opportunity appears. You’re adapting existing, high-quality documentation.
There’s also a signalling effect. A history of successful R&D claims serves as third-party validation of your technical capabilities and innovative approach. When you’re seeking expert advice or pitching to investors, being able to say “We’ve successfully claimed R&D tax incentives for three consecutive years, with all claims approved on first submission” sends a powerful message about your organisational sophistication and the genuine innovation you’re delivering.
Venture capitalists and private investors understand the R&DTI process. They know it’s not easy, and they know AusIndustry doesn’t approve questionable claims. Successfully navigating the R&DTI demonstrates that you’re not just talking about innovation, you’re doing real, systematic R&D that meets government standards.
Which raises an important question: How do you actually build this level of compliance capability? Most organisations don’t have it in-house, and that’s where specialist partners become essential.
The Reality of AI R&D Compliance: What Comes Next
Here’s what you need to understand: The regulatory grey zone surrounding AI R&D isn’t going away. If anything, scrutiny is likely to increase as more organisations make claims in this space and regulators refine their interpretation of what qualifies.
But here’s the opportunity hidden in that challenge.
Organisations that take compliance seriously, that implement proper documentation systems, structure activities appropriately, and work with specialists who understand both the technical and regulatory sides, will find themselves at a competitive edge. Not just in terms of securing their R&D claims, but in building better research practices overall.
They’ll have predictable access to R&D finance that competitors are leaving on the table. They’ll build stronger AI capabilities through more rigorous research practices that yield more reliable results. They’ll be better positioned for additional grants and investment because they’ve already done the documentation work. And perhaps most importantly, they’ll be able to focus their energy on innovation rather than audit defence.
The work you’re doing, developing smarter diagnostic tools, improving patient matching for trials, and creating more effective treatment protocols, has the potential to save lives and advance medical science. The R&D Tax Incentive exists precisely to support this kind of innovation. It’s designed to de-risk the experimental work that leads to breakthroughs and drives economic growth in the sector.
But accessing this funding requires more than just doing innovative work. It requires understanding the compliance framework and structuring your activities accordingly.
Turning Compliance Into Your Competitive Edge
The good news? With the right documentation, structure, and support, you can claim R&D finance confidently. You can access the non-dilutive funding your innovation deserves, and you can do it in a way that not only satisfies regulators but actually improves the quality and repeatability of your research.
AI and machine learning are transforming clinical research, opening possibilities that seemed like science fiction just a few years ago. You don’t have to navigate the compliance complexity alone, and you definitely shouldn’t let regulatory concerns slow down your innovation.
Ready to maximise your AI R&D claim while staying compliant? Learn more about how Rocking Horse Group supports AI-driven research organisations with expert R&D finance solutions designed specifically for the clinical research sector. We help innovative teams access the funding they need to keep pushing boundaries, without the compliance headaches.