R&D Tax Incentive
The R&D Tax Incentive for AI and machine learning companies
AI development is a rich source of genuine R&D and the fastest-growing source of claims, which is exactly why the regulators look at it closely. There is no special AI category in Division 355: the same four-part test applies, and the line between real experimentation and applying powerful off-the-shelf tools has never been easier to blur.
Where the R&D actually lives
Eligible AI work lives where model behaviour is genuinely undocumented: a novel architecture, training under a constraint no published method handles, a capability the literature has not demonstrated for your data and conditions. Calling a frontier model's API, fine-tuning to an expected result, or tuning prompts is applying known tools, not resolving unknowns.
What qualifies, and what does not
| Activity | Eligible? | Why |
|---|---|---|
| Developing a novel model architecture where performance is genuinely unknown | Likely core | Outcome cannot be known in advance |
| Devising a training technique for a hard constraint (tiny dataset, on-device latency) no known method handles | Likely core | Genuine technical uncertainty |
| Building evaluation harnesses and dataset pipelines solely to run those experiments | Supporting | Directly enables the core experiment |
| Fine-tuning a documented model on your data to an expected result | No | Predictable application of known methods (see the Camalic decision) |
| Prompt engineering and model swapping to improve outputs | No | Benchmarking known options, not hypothesis-driven experimentation |
| Building product features around a third-party LLM API | No | Routine integration per documentation |
Classifications are indicative: eligibility always turns on the specific facts, the four criteria in Division 355, and the records behind the work.
The evidence you already produce
ML work produces exceptional evidence almost for free: experiment trackers, training runs, evaluation metrics, model cards, and the dated discussion around approaches that failed. The failed runs matter most, because they are direct proof the outcome could not be known in advance. Keep them linked to the hypothesis they tested.
Watch-outs for AI and machine learning companies
The AAT's Camalic decision established that training an existing algorithm on new data is not R&D, and that claims without contemporaneous hypothesis documentation fail. 'AI-powered' is marketing language, not an eligibility statement, and reviewers know the difference. Frame the claim around the specific technical unknowns, not the sophistication of the tools.
Frequently asked questions
- Is AI development automatically eligible for the R&D Tax Incentive?
- No. AI work is assessed against the same four core-activity criteria as all software. It qualifies when a genuine technical uncertainty is resolved through systematic experimentation, and fails when established techniques are applied to a predictable result.
- Does fine-tuning an LLM qualify?
- Fine-tuning a well-documented model on clean data to a result the literature supports is predictable and does not qualify. Fine-tuning in a genuinely undocumented regime, evidenced by a systematic experiment with a stated hypothesis, may.
- What evidence do AI claims need?
- The experimental chain: the documented hypothesis, the runs that tested it, the metrics that evaluated it, and the conclusions, all created at the time. Experiment trackers and evaluation logs are ideal contemporaneous records; keep the failed approaches, not just the winner.
- Why do AI claims attract extra scrutiny?
- Volume and blur. AI is where new spend concentrates, marketing language often overstates the technical reality, and off-the-shelf models mean a competent professional can achieve a great deal without experimenting. Honest framing and good records answer all three.
Guides for AI and machine learning companies
Is AI and machine learning development eligible for the R&DTI?
AI and machine learning work is a rich source of genuine R&D, and a fast-growing target for ATO scrutiny. Here is how the eligibility test actually applies to AI development, with examples of what qualifies and what does not.
Read articleR&D Tax Incentive for software development: 2026 guide
Software is one of the most-claimed and most-scrutinised areas of the R&D Tax Incentive. This guide maps the four eligibility criteria to real software work, with examples of what qualifies, what does not, and the traps that sink claims.
Read articleContemporaneous evidence for a defensible R&D claim
Eligibility gets you in the door. Evidence keeps you there. This guide covers what contemporaneous records the ATO and AusIndustry expect, the evidence hierarchy that holds up in a review, and the documentation habits that make a claim defensible.
Read articleThe R&D Tax Incentive is a self-assessment program. This page is general information, not tax, legal, or financial advice; eligibility depends on your specific circumstances and you should seek independent advice for them.