AI and machine learning work is increasingly common in software R&D, and it is easy to describe loosely. That makes it worth being precise about when AI work is genuinely eligible and when it is not.
The short answer: building with AI is not automatically R&D, and neither is training a model. Eligibility turns on the same test as any other software, whether you were resolving a genuine technical uncertainty through systematic experimentation. What changes with AI is how easy it is to blur that line.
The test has not changed, the scrutiny has
There is no special AI category in Division 355. AI development is assessed against the same four core-activity criteria as all software. If you can name the technical unknown and show the experiment, you may have a claim. If you used established techniques to a predictable result, you do not, no matter how sophisticated the tooling.
The eligibility test, applied to AI
For a refresher on the four core-activity criteria, see the software eligibility guide. Applied to AI and machine learning, the questions become concrete:
- Was the outcome genuinely unknown? Could a competent professional, with access to worldwide knowledge, have predicted whether your approach would work? Fine-tuning a well-documented model on a clean dataset to a result the literature already supports is predictable. Pushing a model into a regime where its behaviour is not documented may not be.
- Did you run a systematic experiment? Hypothesis, experiment, observation, evaluation, conclusion. In ML this maps naturally onto experiment tracking: a hypothesis about an architecture or training regime, runs that test it, metrics that evaluate it, and a conclusion that changes the next step.
- Were you generating new knowledge? Generating new knowledge has to be a real purpose of the work (it need not be the only one), not just shipping a feature that happens to use a model.
Which AI activities qualify?
The work that qualifies is the work where you genuinely did not know if something would work and had to experiment to find out: novel architectures, undocumented training regimes, capabilities beyond what the literature demonstrates. Applying known techniques, however sophisticated, is unlikely to on its own.
| AI/ML activity | Likely core R&D? | Why |
|---|---|---|
| Developing a novel model architecture where performance is genuinely unknown | Yes | Outcome cannot be known in advance; requires experimentation |
| Devising a new technique to train under a hard constraint (latency, memory, tiny dataset) no known method handles | Yes | Genuine technical uncertainty |
| Researching whether a model can achieve a capability not demonstrated in the literature for your conditions | Often | Depends on whether the uncertainty is real and documented as such |
| Calling a third-party LLM API per its documentation | No | Routine integration, outcome knowable |
| Fine-tuning a standard model on your data to an expected result | No | Predictable application of known methods |
| Prompt engineering to improve outputs | Usually no | Iterative tuning, not hypothesis-driven experimentation |
| Building a RAG pipeline from documented components | Usually no | Integration of known techniques |
The recurring trap with AI is mistaking sophistication for uncertainty. Using a frontier model is not R&D. Building a product around one is not R&D. The R&D, if there is any, lives in the specific places where you genuinely did not know if something would work and had to experiment to find out.
Tuning is not experimentation
Sweeping hyperparameters, trying prompts, or swapping models to see which scores best is benchmarking known options, not resolving a technical uncertainty. To be core R&D, the question has to be one the field could not already answer, and the work has to be a structured experiment, not a search over settings.
Why AI claims need extra care
Three reasons. First, volume: AI is where a lot of new spend is going, so it is where claims are most often loosely framed. Second, the gap between marketing language and technical reality, "AI-powered" is a sales phrase, not an eligibility statement, and reviewers know it. Third, the outcome-unknown test is genuinely harder to satisfy now that powerful models are off-the-shelf, because a competent professional can do far more without experimenting than they could five years ago.
None of this means AI work is hard to claim. It means the claim has to be framed honestly and evidenced well.
How do you document an AI claim?
Keep the records your experiments already produce. AI development generates excellent contemporaneous evidence almost for free: experiment trackers, model cards, evaluation runs, and the dated discussion around failed approaches are the records that can show the outcome was uncertain. The failures matter most: a tracker full of approaches that did not work is strong evidence of genuine uncertainty, alongside a search showing the answer was not already published.
The discipline is the same as any R&D claim, capture the experimental story as the work happens. See how to document a defensible claim for the evidence hierarchy and what fails in a review.
The 2028 angle
The 2026-27 Budget proposed removing supporting activities from 1 July 2028, with higher rates on core activities; it is exposure draft legislation, not law. If enacted, it would reward a sharp account of the experimental core and penalise claims padded with integration and data-preparation work framed as supporting. Read the 2026-27 Budget changes for the full picture, and start tightening how you describe your AI experiments now.
The R&D Tax Incentive is a self-assessment program. Rand helps you prepare and document a defensible claim, but your company and its directors remain responsible for its accuracy, and you should seek advice specific to your circumstances.
