Navigating R&D Tax Incentives for AI and Machine Learning Projects

Published Date : 30/05/2025 

Tech start-ups often wonder if their machine learning and artificial intelligence activities qualify for the R&D Tax Incentive. This article breaks down the eligibility criteria and common pitfalls. 

When advising tech start-ups, the most frequent question we receive is ‘will this development be eligible under the R&D Tax Incentive?’ While it might seem straightforward for programming a new service or application, the legislative definition of ‘eligible R&D’ for the R&D Tax Incentive (R&DTI) differs from the usual business understanding of R&D. In our experience, R&D claimants who self-assess their eligibility often incorrectly classify complex but ‘routine’ software development as ‘experimental’.

With the recent surge in artificial intelligence (AI) and machine learning (ML), companies must be cautious about assuming that activities around AI are automatically eligible under the R&DTI. To assist companies, the Department of Industry, Science and Resources (DISR) has released a ‘Hypothetical Machine Learning case study’ which outlines how R&D claimants using ML techniques can assess their eligibility for the R&DTI.

The case study focuses on a company developing an irrigation decision support system (IDSS) to predict crop water needs based on weather and soil conditions. When customer reports indicated the system was inaccurate, the company decided to enhance it by incorporating satellite images. Their development journey involved key R&D hallmarks such as:

Identifying a gap: Before incorporating satellite imagery, the company conducted online research to learn more about its use for predicting soil moisture. This research found that weather variables were often considered unimportant in previous studies, and that ML models frequently produced false negatives because the model couldn’t identify the importance of the weather variables.

Formulating a hypothesis: Based on this research, the company set its hypothesis as: ‘Applying a variable relevance framework to the dataset will create a machine learning model which accurately identifies which satellite imagery data and weather variables correlate to predict soil moisture’.

Experimentation: The company developed a new framework to identify key weather variables for soil moisture prediction, avoiding predefined assumptions. They trained and tested their ML model using this framework along with a dataset of weather variables and satellite imagery. After training, the company tuned the hyperparameters for the machine learning algorithm used to develop the IDSS by employing a basic dataset to establish benchmarks for both the experiment and the trained model.

They conducted experiments with the framework by training the IDSS algorithm on various datasets with different variables and subsequently testing the resulting model for accuracy. These results were captured and evaluated to understand how certain variables were relevant to predicting soil moisture.

Activity eligibility: Based on the case study, the core R&D activity demonstrated the following key eligibility criteria:

- The outcome could only be determined by applying a systematic progression of work. This work was based on principles of computer science and proceeded from hypothesis to experiment, observation, and evaluation, leading to logical conclusions.

- Crucially, the outcomes of this core R&D activity could not have been known or determined in advance by a competent professional based on current knowledge. This was evidenced by the company’s thorough background research undertaken before starting experimentation.

- The core R&D activity was also conducted for the purpose of generating new knowledge. This was in the form of a new or improved product (i.e., the IDSS system) and a deeper understanding of the variables involved in ranking weather and soil moisture levels.

Common Errors: Often, we see companies self-assessing all of their ML/AI software development as R&D, when some aspects may be routine in nature. If companies apply known processes, such as known software development, AI, and ML processes, to their parameters, they are unlikely to qualify for the R&DTI. In this scenario, the company is not generating new knowledge in the form of a new product or process, rather, they are applying existing knowledge to their business operations.

Some examples of routine processes that would not qualify as a core R&D activity may be:

- Bug, beta, and user acceptance testing

- Data mapping and data migration

- Testing the efficiency of different algorithms that are already known to work

- Routine computer and software maintenance

Contemporaneous documentation: In the event of an ATO or DISR audit, companies are required to provide documentation that substantiates their claims regarding both activities and related expenditure. There is no prescribed set of documentation that a claimant must keep, as it all depends on the nature of the company, the industry it operates in, the R&D activities undertaken, and what records make sense for the company to keep.

However, specific types of documentation/evidence that could be kept include:

- Original experimental hypotheses and design conceptions/project plans

- Correspondence with professionals or experts (emails, reports, meeting notes)

- Payslips, timesheets, and invoices

- Internet/Google searches and screenshots of information found

- Review of scientific, technical, or professional literature

- Technology reviews and patent searches

All documentation should be dated, clearly demonstrate the current industry knowledge, and how the knowledge you intend to generate is considered ‘new’. Significant additional requirements exist to be able to substantiate the R&D tax incentive expenditure claim for the ATO. Guidance on the ATO’s expectations can be found here.

What does this mean for you? This case study strengthens the importance of assessing each element of an ML/AI project to ensure that the claimant is correctly identifying core and supporting activities.

In addition, DISR’s guidance highlights its expectations that you:

- Compile contemporaneous documentation before commencing any experimentation

- Work to and can document a specific hypothesis or hypotheses

- Can describe how you intend to validate a technical or scientific challenge

If you’d like help with identifying eligible ML/AI R&D activities and improving your current record-keeping practices, contact your local William Buck Advisor. 

Frequently Asked Questions (FAQS): 

Q: What is the R&D Tax Incentive?

A: The R&D Tax Incentive is a government program designed to encourage businesses to engage in research and development activities by providing tax benefits.

Q: How can I determine if my ML/AI project is eligible for the R&D Tax Incentive?

A: To determine eligibility, your project must involve a systematic progression of work, generate new knowledge, and have outcomes that cannot be determined in advance by a competent professional.

Q: What are common mistakes in claiming R&D activities?

A: Common mistakes include assuming all ML/AI development is eligible, applying known processes to routine tasks, and failing to document the experimental process thoroughly.

Q: What documentation should I keep for an R&D Tax Incentive claim?

A: You should keep original experimental hypotheses, project plans, correspondence with experts, payslips, timesheets, invoices, and internet search records to support your claim.

Q: How can I get help with my R&D Tax Incentive claim?

A: You can contact a local William Buck Advisor for assistance with identifying eligible activities and improving your record-keeping practices. 

More Related Topics :