R&D Tax Incentive for AI work in Australia.

The numbers

The R&D Tax Incentive (RDTI) is administered jointly by AusIndustry and the ATO. For the 2025-26 income year, eligible companies can claim:

• 43.5% refundable tax offset for companies with aggregated turnover under $20M. If your tax liability is lower than the offset, the difference comes back to you as cash.
• 38.5% non-refundable tax offset for companies above $20M turnover. Used to reduce tax payable; unused portion can be carried forward.

For a $200,000 AI engagement, that's $87,000 back at the small-company tier. The numbers move quickly.

What qualifies as R&D under the scheme

The RDTI is not a generic innovation grant. It targets experimental work that generates new knowledge by applying established scientific principles, where the technical outcome cannot reasonably be known or determined in advance. Two activity classes:

• Core R&D activities, experimental work involving systematic progression from hypothesis to experiment to observation to evaluation. Must address a technical uncertainty that can't be resolved by a competent professional with publicly available information.
• Supporting R&D activities, work that directly supports core R&D. Includes data preparation, infrastructure setup, evaluation harnesses, and deployment work that enables the experimental core.

The bar is "an outcome that can't be known in advance." If a competent engineer could read documentation and ship the result, it's probably not R&D.

Why agentic AI work qualifies

Agentic AI is, by construction, R&D under the scheme. A few reasons:

• Novel architectures. Agent design, tool-use planning, reasoning loops, memory, multi-agent coordination, is current research territory. The optimal architecture for a given business problem genuinely isn't known in advance.
• Empirical iteration is mandatory. You cannot predict ahead of time how an agent will perform on real customer data, what failure modes will emerge, or which evaluation metrics will move with which prompt or tool design. Iteration on real data is not optional.
• The integration layer carries technical uncertainty. Real production agents read across enterprise stacks (ERP, data warehouse, customer systems, audit trails) under safety constraints. Designing reliable behaviour under that complexity is non-trivial and typically requires experimentation.

This doesn't mean every line of code is R&D. Building a UI is not R&D. Configuring a database is not R&D. Wrapping an API in a wrapper is not R&D. But the core agent-design work, the evaluation harnesses, the safety scaffolding, and the integration novelty almost always are.

What you actually need to claim

Two artefacts. Get both right and the claim is straightforward; miss either and the claim is fragile.

1. Activity records, kept contemporaneously

Notes from sprint planning, design discussions, evaluation runs. The hypothesis you tested, the experiment you ran, the result you observed, the next step you took. These do not have to be elaborate. A weekly engineering note that records "we hypothesised X, ran experiment Y on dataset Z, observed result R, decided to pursue path P" satisfies the requirement.

The documentation should be created during the work, not retrofitted at the end of the financial year. AusIndustry can ask for evidence and frequently does.

2. Cost records, allocated cleanly

You're claiming a percentage of eligible expenditure. Eligible expenditure includes:

• Salaries of staff doing the R&D (apportioned by time)
• Contractor and consultant costs for R&D work
• Cloud and tooling costs directly attributable to R&D
• Overhead apportioned at a defensible rate

Your delivery partner's invoices should distinguish R&D-eligible work from non-R&D work. Ask for it.

Three mistakes that kill claims

1. Treating the claim as a year-end retrofit. Records made nine months after the work happened are weak evidence. AusIndustry knows this and reads it accordingly.
2. Claiming work that has no technical uncertainty. A standard CRUD application built by a competent developer is not R&D, no matter how much AI vocabulary surrounds it. Be honest about which parts of the engagement are experimental and which are construction.
3. Not registering on time. Registration with AusIndustry is required before claiming with the ATO, and must happen within 10 months of the end of the income year. Miss the window and the claim is gone.

What we hand over

Bedstone isn't a tax specialist and we don't lodge claims. What we do hand over to your tax advisor or in-house finance team is the readiness pack: structured documentation of the technical uncertainties addressed during the engagement, the experimental approach taken, the evaluation results, and a clean cost breakdown that distinguishes R&D-eligible work from build work. Your advisor takes that pack and prepares the AusIndustry registration and the ATO claim.

This is a deliberate division of labour. Tax registration and lodgement are work for tax specialists. Documentation of the technical R&D is work for the engineering team that did the R&D. Asking your tax advisor to invent the technical narrative after the fact is how claims get rejected on review.

Quick checklist before signing an AI contract

• Will the partner deliver structured engagement documentation suitable for an RDTI claim?
• Are sprints framed as hypothesis-experiment-observation cycles, with notes kept contemporaneously?
• Are invoices structured to distinguish R&D-eligible from non-eligible work?
• Will the partner write up technical uncertainties and experimental outcomes in a form your tax advisor can use?
• Is the partner happy to be on a call with your tax advisor if AusIndustry asks for clarification later?

If the answer to any of these is "we don't really do that", you're going to lose the offset. The claim itself is straightforward; it's the underlying documentation that makes or breaks it.

The takeaway

Most AU companies running serious AI work qualify for somewhere between $30K and $300K back through the RDTI on a typical mid-six-figure engagement. The work is genuinely R&D under the scheme. The reason claims don't go in is that the engineering team didn't document the experimental nature of the work in real time, or the contracts and invoices didn't separate R&D from non-R&D effort cleanly enough to defend.

If you're scoping AI work and your delivery partner has never heard of the RDTI, that's a signal. The competent ones know the scheme, document accordingly, and hand over the readiness pack as a deliverable. None of it is hard. It just needs to be done from day one.