Does Your AI Startup’s Work Qualify for the R&D Tax Incentive?
Artificial intelligence is where most of the interesting technical work is happening in Australian startups right now. Companies are training custom models, building novel inference pipelines, and solving problems that didn’t exist five years ago.
They are also, in many cases, leaving money on the table.
The Australian R&D Tax Incentive refunds up to 43.5% of eligible development spend as a cash refund for companies turning over under $20 million. For an AI startup spending $300,000 a year on genuine development, that is potentially $130,000 back. But a lot of AI founders assume their work either automatically qualifies because it sounds impressive, or definitely doesn’t qualify because it’s “just software.”
Neither assumption is usually right.
Here is a clear breakdown of how the R&D Tax Incentive applies to AI and machine learning work, what qualifies, what doesn’t, and how to make sure you’re not missing a significant cash refund.
The Test That Applies to Every R&D Claim
Before getting into AI specifics, it helps to understand the test that applies to every R&D Tax Incentive claim, regardless of industry or technology.
The test is technical uncertainty. The work must involve an outcome that could not be known in advance, based on existing knowledge, information, and experience. Your team must have proceeded from a hypothesis, conducted experiments, made observations, and drawn conclusions. The goal must have been to generate new knowledge.
That’s it. No laboratory required. No academic research required. No patent required.
The question isn’t whether your AI product is impressive. The question is whether there was a point in your development where your team genuinely didn’t know if something would work, and had to experiment, fail, and iterate to find out.
That test applies to AI and machine learning exactly as it applies to hardware, biotech, or any other technical field.
What Qualifies for AI and Machine Learning Projects
AI and machine learning are explicitly recognised areas of technical R&D under the program. The question is which specific activities within those fields meet the technical uncertainty test.
Novel model development is where the strongest cases are built. If you are designing model architectures that don’t exist in the published literature, testing training approaches without a known outcome, or combining techniques in ways that haven’t been documented in the field, the technical uncertainty test is relatively straightforward to satisfy. The outcome has to be genuinely uncertain and the approach has to go beyond applying established methods.
Training on proprietary datasets with uncertain outcomes is another strong category. Many AI startups have access to unique datasets in specialised industries, from healthtech and agtech to legal tech and construction. If training on that data involved genuine uncertainty about model behaviour and required iteration to get useful results, including failed experiments along the way, that work often qualifies.
Building novel data pipelines for technically uncertain problems can be supporting R&D. If your data infrastructure was directly linked to a core experimental activity and was built specifically to enable that research, it may be included in the claim.
Solving engineering problems with no established solution is also worth examining. Latency at scale, inference efficiency, multi-modal systems, and real-time model serving regularly throw up genuine technical problems that can’t be resolved by following a documented approach. If your team designed and tested novel solutions, that is worth looking at.
What Doesn’t Qualify
This is where AI founders often misjudge their own work, in both directions.
Using an existing large language model via an API, or applying standard fine-tuning with documented techniques, is not R&D. If you are integrating an existing AI tool into your product in the way it was designed to be used, that is development, not research. Routine fine-tuning with well-established hyperparameter choices and standard architectures typically doesn’t involve the kind of uncertainty required.
Applying machine learning to understand consumer preferences or market behaviour is specifically excluded. The program excludes market research and research into consumer behaviour, regardless of the technical methods used. If your AI project is fundamentally aimed at figuring out what customers want rather than solving a defined technical problem, it needs careful consideration before you register it.
Building AI-powered product features using documented, established tools in their intended way does not qualify. The bar is whether the approach was genuinely uncertain, not whether you’ve used machine learning. A lot of good AI engineering is not R&D, and that’s fine. What matters is identifying the parts of your work that genuinely were.
A useful question to ask: if your lead engineer described this work to an expert in the field, would they say “we knew how to do this” or “we genuinely weren’t sure if this approach would work”? The honest answer to that question usually maps closely to what AusIndustry will decide.
The Classification Question That Trips Up AI Founders
There’s a technical piece that catches some AI startups off guard: how the work gets classified in the AusIndustry registration.
The application requires an ANZSRC (Australian and New Zealand Standard Research Classification) code. Many founders assume that AI or machine learning work should go under the Machine Learning category (code 4611). But that code covers the development of underlying machine learning techniques, not the application of those techniques to a specific domain.
Most AI startup work is better classified under the relevant domain. A startup training AI models on medical imaging data would typically be classified under Biomedical and Clinical Sciences. A startup applying ML to agricultural yield prediction would be classified under Agricultural, Veterinary and Food Sciences.
This matters because some domains have exclusions. Research in social sciences, arts and humanities, and market research into consumer behaviour is excluded from the program regardless of the methods used. Getting the classification right from the start protects the claim.
How to Demonstrate Technical Uncertainty for AI Work
AusIndustry has published guidance on demonstrating that an AI activity’s outcome was genuinely unknown in advance. Their framework suggests a two-stage process.
First, run a preliminary literature review. Search the research and technical literature for published work addressing the specific problem your project is trying to solve.
Second, develop a hypothesis based on what you found, then do a second review to check whether any published research has already tested that hypothesis or solution approach.
If the second review finds no existing research testing your proposed solution, that supports the conclusion that the outcome was genuinely unknown based on current knowledge.
This doesn’t need to be a formal academic exercise, but it does need to be documented. A record of the searches you ran, the articles or papers you reviewed, and what you concluded is exactly the kind of contemporaneous documentation AusIndustry looks for.
What Costs Can You Claim?
If your AI activities qualify, the eligible costs are broader than most founders realise.
Staff and contractor costs for Australian developers, data scientists, and engineers who work on qualifying activities are eligible. Their time needs to be documented and clearly linked to the R&D work.
Cloud compute costs, including GPU instances used for training runs, test environments, and development infrastructure, are eligible. Production serving costs are generally not eligible, but experimental training and testing compute is.
Hardware depreciation is eligible for equipment used specifically for R&D purposes, including specialised servers and workstations.
Software licences and tools used specifically for R&D activities can be included.
For all of these, the key is a clear, documented link between the expenditure and the registered R&D activities.
What Your Records Should Look Like
AI development naturally generates good records, which is an advantage for claims in this sector.
Experiment logs and model training results, including records from failed approaches, are strong contemporaneous evidence. If you track experiments in a tool like Weights and Biases or even a shared spreadsheet, those records already exist and tell a clear story.
Version control history, particularly commit logs showing iteration, changes in approach, and failed builds, demonstrates that the work required experimentation rather than straightforward execution.
Technical design documents, architecture decisions, and notes from engineering discussions that explain why a particular approach was chosen, and what alternatives were considered, add significant depth to a claim.
The records AusIndustry and the ATO want are ones created at the time the work was done, not reconstructed later. Most AI development teams already produce those records. The task is making sure they’re collected and organised for the claim.
The New Financial Year Starts in One Week
The financial year ending 30 June 2026 closes in a few days. The year starting 1 July 2026 is the right moment to set things up properly.
For AI startups beginning or continuing development work in the new financial year, the most valuable thing you can do right now is establish good record-keeping from the start. Document the technical uncertainty at the beginning of each experimental stream. Log your experiments as you run them, including the approaches that didn’t work. Separate R&D spend from routine development costs in your accounting from day one.
The current rules apply in full for FY2026-27. The 43.5% refundable offset remains available for companies turning over under $20 million. The minimum spend is $20,000. The R&D Tax Incentive changes announced in the 2026 Budget, including the removal of supporting activities and the higher minimum spend, don’t apply until 1 July 2028.
If you’re an AI startup spending meaningful money on genuine experimental development in Australia, there is very likely a real cash refund available. The question is whether you’re set up to capture it.
Where to Start
Most AI founders underestimate their own eligibility. The work feels routine to them because they do it every day. The technical uncertainty that seems normal from the inside is exactly what the R&D Tax Incentive is designed to reward.
If you’re not sure whether your work qualifies, that’s exactly the kind of thing worth a quick conversation. Granton does free 15-minute eligibility calls. We’ll give you a straight answer on whether we think your work qualifies, what the claim might be worth, and what documentation you need. If it doesn’t stack up, we’ll say so.
You can book a call at go.granton.io. No cost, no obligation, just a clear picture of where you stand.
Are you ready to turn your funding aspirations into reality? At Granton, we specialize in helping individuals and businesses navigate the world of grants, offering expert guidance on grant applications and finding opportunities that best suit their needs. Whether you’re seeking funding for a startup, nonprofit, or a specific project, our team is here to assist you every step of the way. We take the guesswork out of Grant Applications, R&D Tax Incentives, and Accelerator Programs, making the process smoother and increasing your chances of success. Ready to take the next step? Book a free consultation with us today, and let’s explore how we can help you secure the grants you deserve. Visit our website at granton.io to learn more or use our contact form to get in touch. Your grant journey starts here!