The rapid adoption of Artificial Intelligence (AI) has led to a surge in R&D Tax Incentive claims across Australia. However, the regulator has significantly tightened the criteria for what constitutes eligible AI development. Many companies mistakenly believe that any work involving a Large Language Model (LLM) or a neural network is inherently experimental. In reality, the Australian Taxation Office (ATO) and AusIndustry now apply a much stricter “Competent Professional” test to AI projects. We help you navigate these new guidelines to ensure your AI innovation is recognised as a core R&D activity rather than routine software engineering.
The AI Technical Uncertainty Test
To qualify for the R&DTI, your AI project must address a technical uncertainty that cannot be resolved by a competent professional using existing knowledge. In the context of AI, this means you cannot simply be applying a well-known model to a new dataset. If a senior Data Scientist can predict the outcome of a training run based on established industry benchmarks, the activity is unlikely to meet the “Core” requirement.
Eligible activities typically involve the development of novel neural network architectures, the creation of custom loss functions to solve specific edge cases, or significant optimisations to reduce latency in real-time inference. We work with your technical team to isolate these genuine uncertainties from the standard implementation of off-the-shelf tools like OpenAI’s API or open-source libraries.
Distinguishing Data Labelling from Experimentation
A major red flag for regulators in 2026 is the inclusion of manual data labelling and cleaning as a core R&D activity. While high-quality data is essential for any Machine Learning (ML) project, the act of labelling data is generally considered a supporting activity or business-as-usual data management.
To withstand an audit, your documentation must show that the experimentation happened within the model itself. This might include hyperparameter tuning, testing various data augmentation techniques, or experimenting with different feature engineering strategies. At Incentur, we ensure that your claim correctly categorises these tasks, protecting you from the “routine data processing” exclusions that lead to 100 per cent clawbacks.
Hypotheses in the Machine Learning Pipeline
The regulator expects your AI development to follow a systematic progression of work, which begins with a formal hypothesis. In an AI context, a hypothesis should be more specific than “we hope this model is accurate.” It should target a specific technical variable, such as “using a transformer-based architecture with a custom attention mask will reduce training time by 20 per cent without compromising F1 scores on our proprietary dataset.”
We provide the frameworks to help your engineers document their “failure sequences.” In the eyes of an AusIndustry examiner, a project that worked perfectly on the first try is rarely considered R&D. By recording the failed experiments, the biased outputs, and the model collapses, we build a “regulator-grade” evidence file that proves your work was a genuine pursuit of new knowledge.
AI Infrastructure and Compute Costs
The cost of training large-scale models can be astronomical, and these cloud compute costs (such as AWS or Azure instances) are often eligible for the incentive. However, the ATO is increasingly scrutinising cloud bills to ensure they relate only to the experimental phase of the project. Once a model is deployed into a production environment for commercial use, the compute costs typically become ineligible.
Our team provides the precision required to apportion these costs. We help you map your cloud expenditure directly to your experimental logs, ensuring that you claim the maximum amount possible for your training and validation phases while excluding the costs of running the AI for your end-users. This level of technical and financial separation is the hallmark of a boutique, expert-led claim.
