The Department of Industry, Science and Resources has flagged “risk areas” where AI projects are unlikely to qualify as Core R&D, even as the government pushes for Australia to become a “builder” of technology rather than just an “adopter.”
Companies integrating AI into their operations are being warned that routine adoption of existing tools and platforms won’t clear the bar for the R&D Tax Incentive (R&DTI).
The Department of Industry, Science and Resources (DISR) has issued guidance setting out where AI-related work is likely to fail the test for Core R&D, because the problems involved can be solved by experts using existing knowledge rather than genuine experimentation.
The R&DTI offsets up to $150 million in eligible expenditure, designed to support research that wouldn’t otherwise happen.
What counts as Core R&D
Under the legislation, Core R&D activities are those whose outcome can not be determined in advance by an expert using existing knowledge.
Instead, the outcome has to be worked out through a systematic progression of work testing a hypothesis through planned experiments.
Supporting R&D activities are those directly related to a Core activity, such as a review of industry publications carried out specifically to refine a hypothesis ahead of an experiment.
The rules are technology-neutral. It doesn’t matter whether a business is working with AI, quantum computing or a new type of tractor; the same test applies.
Where AI projects are falling down
DISR’s guidance lists a series of activities it considers routine software engineering rather than eligible R&D. They include:
- Implementing logging, alerts or dashboards using established tools to confirm a model is behaving as expected
- Cleaning, formatting and aligning data to meet a model’s requirements using standard techniques
- Running regression or acceptance tests where the expected outcomes are already known
- Adjusting parameters such as learning rates when the effect of the change is already understood
- Plugging a known model output, such as a score or recommendation, into a dashboard using pre-defined logic
In other words, if an expert team can solve the problem with knowledge they already have, it’s unlikely to pass the test.
|
Activity Category |
Routine Activity (Generally Ineligible as Core) |
Experimental Activity (Potentially Eligible as Core) |
|---|---|---|
|
System Monitoring |
Implementing logging, alerts, and dashboards to confirm expected model behaviour. |
Investigating novel fine-tuning strategies to resolve scalability issues where established methods have failed. |
|
Data Handling |
Cleaning, formatting, and aligning data to meet documented input requirements using standard techniques. |
Experimentally testing a new AI architecture to determine if it can cope with a specific, “noisy” data regime. |
|
Testing |
Running regression or acceptance tests to confirm a system works as intended when outcomes are known. |
Conducting a series of experimental runs to test a hypothesis when outcomes cannot be determined in advance. |
|
Parameter Tuning |
Adjusting parameters (e.g., learning rate) using established methods where the effect is well understood. |
Testing unproven chunking methods on real-world data where experts cannot predict the response. |
|
Integration |
Integrating known model outputs into apps using pre-defined logic and interfaces. |
Developing and testing a proposed solution to resolve domain-specific performance limitations |
A tale of two chatbots
The guidance uses two chatbot examples to illustrate the difference.
In the first example, a company researches available platforms, frameworks and Retrieval-Augmented Generation pipelines, reads reviews and picks a technology stack it’s confident it can implement.
That’s not Core R&D. The company relied on existing knowledge to determine the outcome, with no technical hurdle requiring experimentation.
The second example is about a company chatbot that gives inconsistent answers to questions about complex documents. This paper provides a way to compare the performance of the implementation of a novel retrieval strategy called dynamic chunking with a given baseline in the same QA system.
The results do not match the company’s initial assumptions, forcing it to revise its hypothesis and run further tests.
That activity can qualify as Core R&D, because experts couldn’t have predicted the outcome in advance.
Why the government is drawing the line
Prime Minister Anthony Albanese used a recent speech at the University of Sydney to argue Australia should become a “builder” of technology rather than just an “adopter,” pointing to the country’s universities and skilled workforce as the basis for sovereign capability in areas such as cybersecurity and biotechnology.
The government is also establishing an Office of AI and a single national framework of Australian Standards for AI.
Albanese told the audience the issues raised by AI are not purely technical but economic, legal and social, and pointed to applications such as cancer-screening tools, cutting paperwork for small business and lifting agricultural productivity as the kind of work the government wants to see funded.
Businesses claiming the R&DTI are assessed activity by activity, and are required to keep contemporaneous records including documented hypotheses, experiments and results to support a claim if audited.
Also read: R&D Tax Threshold Changes in 2028: From 2% to 1.5% Will Save You?
Where advisers say the guidance falls short
Not everyone advising on R&D claims agrees the line between “core” and “routine” is easy to apply in practice.
Even many of our clients struggle to tell where routine software engineering ends and genuine experimentation begins, particularly when AI development involves iterative testing that can look similar either way.
This uncertainty can discourage businesses from documenting or pursuing work that might otherwise qualify.
Some people in the sector also argue the criteria are becoming harder to satisfy. If the “technical hurdle” threshold is seen as too high or too complex to demonstrate, businesses may decide against lodging a claim at all, even where their AI work involves genuine uncertainty.
There are also concerns the guidance favours businesses with more resources. Larger firms are better placed to run structured experiments, track baseline metrics and produce the contemporaneous records DISR expects, while smaller companies may lack the time or headcount to document their work in the same way.
That could work against the policy’s own goals, given smaller firms are often the ones pushing the boundaries of what existing AI tools can do.
