This article is for businesses that are building artificial intelligence into their products or operations, including software companies, manufacturers and agricultural technology businesses, and want to know whether that work qualifies for the R&D Tax Incentive. It explains where the line sits on AI and the R&D Tax Incentive, the exclusion that catches internal software, and the records that let a claim stand up to review.
Using AI tools is a business activity
Paying for an AI subscription and using it to draft documents, summarise contracts or analyse sales data does not create R&D, however much it improves the business. The tool exists, it does what its developer designed it to do, and using it does not generate new knowledge. The cost sits with ordinary business expenses.
The same is true of configuring an AI product inside an existing platform to a result the vendor already supports. Where the outcome is known before the work starts, the work is implementation.
What core R&D looks like in AI work
Core R&D activities are experimental activities conducted to generate new knowledge, where the outcome cannot be known or determined in advance on the basis of current knowledge, information or experience. The outcome can only be determined by a systematic progression of work: a hypothesis, an experiment, observation, evaluation and logical conclusions.
In AI development, that can describe work such as finding out whether a model can reach the accuracy a product needs on a type of data that existing approaches have not handled, or whether a model can run within the processing and power limits of a piece of equipment in the field. In each case a competent developer could not tell you the answer at the start, the team forms a view about what might work, tests it and records why runs succeeded or failed.
The work around those experiments can qualify as supporting R&D where it is directly related to the core activity and done for the dominant purpose of supporting it. Preparing and labelling a dataset for a particular experiment is one example.
The Department of Industry, Science and Resources publishes a software development sector guide that sets out how the rules apply to software, and much of it applies directly to AI projects.
The exclusion that catches internal software
Developing, modifying or customising software for the dominant purpose of your own internal administration cannot be a core R&D activity. The same applies to software for a connected or affiliated entity. The department's examples of internal administration include payroll and accounting, invoicing, ordering, quality control reports and information management.
So an AI tool built to code supplier invoices for your own accounts is excluded, even if building it involved genuine technical uncertainty. The same tool built as a product for sale to other businesses is assessed on its merits as potential core R&D. Where internal software is directly related to a core R&D activity and built mainly to support it, it may still qualify as supporting R&D, and the records need to show that purpose.
Integration and fine-tuning sit on both sides of the line
Connecting your system to an existing model through its interface, so that it performs a task the model is known to handle, is generally development work with a predictable outcome. It may be difficult and it may take months, but difficulty on its own is not the test.
Writing and refining prompts so an existing model produces a result it is already capable of is part of using the tool. The same applies to adjusting settings or connecting data sources the product was designed to accept.
Fine-tuning or adapting a model can be different. Where the team does not know whether a technique will achieve the required result on your data, designs experiments to find out and evaluates the results against a measure set in advance, the work can meet the definition of core R&D. The question in every case is whether the answer could have been known at the start by someone with the relevant expertise.
The records an AI claim needs
AI projects move quickly, and the record keeping has to keep pace with them. A claim that holds up has the hypothesis written down before the experiment starts, the measure that will decide success, the model versions and datasets used for each run, the results including the runs that failed, and the conclusions the team drew. Time records that link each person's hours to the experiment they supported complete the picture.
Much of this sits in experiment tracking tools, code repositories and project boards as a matter of course. The work is making sure it is captured at the time, linked to the hypothesis it tested and kept. Our article on R&D record keeping covers the records that apply to every claim, and our guide to software R&D covers software projects more broadly.
What AI and the R&D Tax Incentive can be worth, and when to register
For a company with aggregated turnover below $20 million, the R&D tax offset is refundable and equals the company's tax rate plus an 18.5% premium. A company paying tax at 25% receives an offset of 43.5% of eligible expenditure. Companies with turnover of $20 million or more receive a non-refundable offset of their tax rate plus 8.5% on R&D expenditure up to 2% of total expenditure, and plus 16.5% above that.
Activities must be registered with the department within ten months of the end of the income year. For a company with a 30 June year end, R&D carried out in the 2025-26 year must be registered by 30 April 2027.
Questions we are asked
Does using ChatGPT or other AI tools count as R&D?
No. Using an existing AI tool for business tasks does not generate new knowledge, so it is not an R&D activity. The subscription is an ordinary business expense.
Can developing AI software qualify for the R&D Tax Incentive?
It can, where the outcome could not be known in advance and the team follows a systematic process of hypothesis, experiment, observation and evaluation to find it out. Software built mainly for your own internal administration is excluded from core R&D.
Is an AI tool we build for our own use eligible?
Not as core R&D if its dominant purpose is internal administration, such as accounting, invoicing, ordering or quality control reporting. It may qualify as supporting R&D where it is directly related to a core R&D activity and built mainly to support it.
What records do we need for an AI R&D claim?
The hypothesis and success measure recorded before each experiment, model versions and datasets, results including failed runs, conclusions, and time records linking staff hours to each experiment.
When is the R&D registration deadline?
Ten months after the end of the income year. For a 30 June year end, 2025-26 activities must be registered by 30 April 2027.
Assessing your AI work
If your team is building with AI and you would like to know which parts of the work qualify under AI and the R&D Tax Incentive rules, our R&D Tax Incentive advisory team can review the project with you before the year ends. Book a time with us.
