Quick answer: Routine use of existing tools, APIs, libraries or pre-trained models — including standard training, fine-tuning an LLM, prompt engineering and API integration — will often fall outside core R&D unless it involves resolving a genuine technical uncertainty through systematic experimentation. In contrast, R&D requires a technical unknown investigated through hypothesis-driven experimentation, often where existing approaches are insufficient.
25 June 2026 — the 2026–27 Federal Budget announced R&DTI changes scheduled to start 1 July 2028; this article describes the current rules unless stated otherwise.
The AI boom has put a hard question in front of a lot of Australian companies: we're building with machine learning — can we claim the R&D Tax Incentive (R&DTI) for it? It's a fair question, and the honest answer is "sometimes". Using AI is not the test. The test is whether your work involves a genuine technical unknown that you resolve through systematic experimentation — or whether you're applying existing tools, libraries, pre-trained models and APIs to a known outcome.
That distinction matters more in AI than almost anywhere else, because so much modern ML is, by design, easy to apply. Calling an API, training a standard model on your data, or fine-tuning an off-the-shelf large language model (LLM) are now routine engineering tasks for many teams. As a Registered Research Service Provider (RSP000047) based at Lot Fourteen — alongside the Australian Institute for Machine Learning — we see a lot of strong AI work that is genuinely experimental, and a lot that simply isn't. The mechanics below come from business.gov.au, AusIndustry and the ATO; how they apply to you depends on your facts, so self-assess and seek your own advice.
The eligibility test, applied to AI/ML
The R&DTI funds core R&D activities: experimental activities whose outcome cannot be known or determined in advance on the basis of current knowledge, information or experience, and which are conducted by applying a systematic progression of work — based on principles of established science, proceeding from hypothesis to experiment, observation and evaluation — for the purpose of generating new knowledge.
business.gov.au publishes a dedicated AI sub-guide, "Artificial intelligence related activities and the R&D Tax Incentive", sitting under its broader Software development sector guidance. The throughline of both is consistent: the eligibility question is not "did we use AI?" but "was the technical outcome genuinely uncertain, and did we run a systematic experiment to resolve it?"
What usually points to eligible R&D
A technical unknown that existing ML approaches are known to be inadequate for — and that you can't resolve just by reading the literature or trying the documented method.
Developing novel architectures or algorithms where it's genuinely uncertain whether they'll work or perform.
A stated hypothesis and a planned experimental progression (e.g. designing experiments, varying approaches, measuring against defined criteria, evaluating results).
A purpose of generating new knowledge — not just getting your product working.
What is usually routine (and not core R&D)
Integrating an existing AI API or off-the-shelf model into your product to achieve a known outcome.
Training a standard model on your own data using established, documented methods where the result is expected to work.
Fine-tuning an existing LLM with standard techniques to adapt it to your domain — generally routine, unless you're genuinely resolving a technical unknown along the way.
Prompt engineering and configuration of existing models.
To make the contrast concrete:
Typically not core R&D: integrating an existing AI API or off-the-shelf model to achieve a known product outcome, where no technical uncertainty exists and established implementation patterns are used.
Possible R&D: experimentally developing a novel retrieval, chunking or evaluation architecture because existing retrieval-augmented-generation (RAG) methods fail under a documented technical constraint that can only be resolved through experimentation.
Training, fine-tuning and APIs: where the line falls
These are the four scenarios we're asked about most, so it's worth being specific.
Using an AI API or off-the-shelf model is, on its own, applying an existing tool to a known outcome — generally not eligible core R&D. The clever product you build around it may be valuable, but value isn't the test; technical uncertainty is.
Training a model on your data with standard methods is usually routine. If, however, standard approaches demonstrably fail and you have to experiment to find one that works, the experimental portion may be core R&D — and you'd need records showing why the outcome was uncertain.
Fine-tuning an existing LLM is often routine adaptation, but it may form part of core R&D where it is used to resolve a demonstrable technical uncertainty that cannot be addressed using established training or optimisation techniques, and where systematic experimentation is required.
Data labelling, collection and preparation is typically supporting activity at best (directly related to, and undertaken for the dominant purpose of supporting, an eligible core activity) — not core R&D in its own right.
One more trap that bites AI/software in particular: the dominant-purpose internal-administration exclusion. Software (including AI tools) developed for the dominant purpose of your own internal business administration is excluded from being a core R&D activity (s 355-25(2) of the Income Tax Assessment Act 1997). An internal ML tool built mainly to run your own operations can fall outside the incentive even if it was technically hard to build.
For the broader picture across software and AI products, see our R&D Tax Incentive for software and AI page; for the common knock-out reasons, see what does not qualify.
Records and scrutiny: why AI claims get a hard look
AusIndustry and the ATO scrutinise software and AI claims heavily, and have done for years — the ATO's Taxpayer Alert TA 2017/5 flagged concerns about software-development claims that don't reflect genuine eligible activities. AI hasn't changed that posture; if anything, the ease of applying modern ML makes contemporaneous evidence more important, not less.
Practically, that means keeping records made at the time, that show:
the technical unknown and why it couldn't be resolved with existing knowledge or available tools;
the hypothesis you set out to test;
the experiments you ran — what you varied, what you measured, what you observed — and the conclusions you drew;
a clear separation between the experimental work and the routine engineering around it.
If your evidence only shows that you built a working AI feature, that tends to read as routine development. If it shows uncertainty, hypothesis and systematic experimentation, you're describing R&D. Getting that framing right — without overclaiming — is exactly the structuring work an RSP supports.
Come back to the one principle that decides most of these calls: using AI is not the test — the test is whether there is a technical unknown resolvable only through systematic experimentation.
Talk to Ignition Research before you register an AI project — we help separate genuine experimental development from routine build work, and structure the evidence so an eligible claim stands up to scrutiny.
Frequently asked questions
Q: Can I claim the R&D Tax Incentive on AI model development? A: You may be able to, but only where the work is genuinely experimental — a technical outcome that couldn't be known in advance, resolved through a systematic, hypothesis-led progression of work. Simply applying existing tools, models or APIs to a known outcome is generally not eligible. business.gov.au's AI sub-guide sets out how the test applies. Self-assess against your facts.
Q: Is training a machine-learning model eligible R&D? A: Often not by itself. Training a standard model on your own data using established methods is usually routine. The experimental portion may be eligible if standard approaches are demonstrably inadequate and you have to experiment to find one that works — and you can evidence that uncertainty.
Q: Is fine-tuning an existing LLM core R&D or routine? A: Usually routine adaptation. Fine-tuning an existing LLM with standard techniques to suit your domain generally isn't core R&D. It can shade into eligible work where you're resolving a genuine technical unknown that established techniques can't solve — which you'd need records to support.
Q: Is using an AI API or off-the-shelf model claimable? A: Generally not as core R&D. Integrating an existing API or pre-trained model to achieve a known outcome is applying an existing tool, not experimenting to generate new knowledge. The product may still be commercially valuable, but commercial value isn't the eligibility test.
Sources & further reading
business.gov.au — Artificial intelligence related activities and the R&D Tax Incentive: business.gov.au
business.gov.au — Software development sector guide: business.gov.au
ATO — Taxpayer Alert TA 2017/5 (software development R&D claims): ato.gov.au
ITAA 1997 s 355-25 (core R&D activities; internal-administration exclusion): legislation.gov.au
Related: R&D Tax Incentive for software and AI · What does not qualify · Registered Research Service Provider
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
