Is training ML models eligible for R&D?

Is training a machine learning model eligible R&D?

Sometimes eligible

Sometimes. Training a model qualifies when an experienced ML engineer could not predict whether an architecture or training method would meet a measurable target under your data and deployment constraints, and you resolve it by planned training runs. Training a standard model type on clean data to a predictable result is routine development.

Reviewed by George WalchFacts last verified 24 October 2026

A weed detector for a sprayer's onboard processor

An illustrative agtech company needs to detect three weed species at 90% recall at 30 frames per second on a 4 TOPS edge processor mounted on a boom sprayer, trained on only 2,500 labelled images spread across seasons and light. Published detectors report results on different crops and far larger datasets, and its ML lead cannot say whether a pruned architecture with synthetic augmentation will hold recall at that frame rate. It plans training runs varying pruning level and augmentation, holding a field-captured test set fixed, and records each evaluation.

Retraining a credit model with an AutoML service

A lender retrains its gradient-boosted credit-limit model each quarter using a hosted AutoML service, which picks features and hyperparameters and reports a slightly better AUC. The method is documented and its result is the kind it always produces. Nothing was unknown to a competent data scientist, so there is no core activity, however much compute it used.

The evidence that supports it

A dated record of the published benchmarks and papers checked for your task, data size and hardware, and why none answered the question.

The hypothesis with its numeric target and deployment constraint, written before training began.

Training run logs from an experiment tracker, tied to the variant each run tested.

A fixed, versioned test set captured from the real operating conditions.

Labelling instructions and batch records showing which experiment each labelled set served.

An evaluation note per iteration explaining what the result changed in the next design.

Training a model looks like experimentation. There are runs, metrics, iterations and failures. But a training loop is a development process, and AusIndustry and the ATO both warn that a systematic-looking process is not the same as an experiment. The question is whether the result of training was knowable before you started.

Where the unknown usually sits

For a model trained from scratch or built on a custom architecture, a genuine knowledge gap tends to come from one of three directions:

The data regime. Very few examples, heavy label noise, extreme class imbalance, or a distribution that published datasets do not resemble.
The deployment constraint. A latency, memory or power budget, for example an edge processor or a phone, where published accuracy figures assume data-centre hardware.
The architecture itself. AusIndustry's AI guidance gives, as an example that may be core, testing an architecture against domain-specific performance limits when it is not known whether it can cope "with the available data regime".

The law looks for an outcome no competent professional could determine from worldwide knowledge, pursued through a systematic progression from hypothesis to conclusion. "We have never trained one of these" is not that.

A new data regime is a context, not a gap

Moreton Resources [2019] FCAFC 120 held that applying existing technology in a new context can be a core activity. In ML terms, a standard architecture applied to data unlike any published benchmark can qualify. But Moreton is not a free pass: you still need evidence that the published results could not predict behaviour in your context, and a planned way to find out. Record why your data differs and why that difference matters to the outcome.

What stays routine

AusIndustry lists several AI tasks as unlikely to be core: monitoring and dashboards using established tools, cleaning data to a model's documented input requirements, regression and acceptance tests with known expected outcomes, and parameter changes whose effect is well understood. The ATO's software alert TA 2017/5 adds testing the efficiency of algorithms already known to work. Feature engineering with standard transforms usually belongs on the same list, unless how a transform behaves on your data is itself the open question.

Edge cases specific to training

Labelling. Building a labelled set for an experiment can be supporting work; labelling to keep a production model fresh is not. The wider data question is on the data pipelines page.
AutoML and architecture search. The tool searches; it does not hypothesise.
After it works. Once the uncertainty is resolved, training the final model at full scale for release, and retraining it on schedule, is usually productionising. The core activity ended when the uncertainty did. If scaling up raises its own unknown, such as whether a result found on a subset holds on the full dataset, that question is assessed on its own.
Licensed weights. Expenditure on acquiring core technology, meaning technology the R&D sets out to build new knowledge or products on, cannot be notionally deducted (s 355-225(2)). Licensed pretrained weights you build on may be core technology; ordinary compute and API spend used in the experiments is a separate question. Your accountant decides.
Runs that went nowhere. Keep them with the variant each tested. They support the case that the outcome was unknown when the plan came first, but do not prove it alone; see failed experiments.

Adapting a pretrained language model is a narrower question, answered on the LLM fine-tuning page. For how AI companies scope activities overall, see R&D for AI companies.

Frequently asked questions

Is data labelling claimable?

Labelling done to build the training and test sets for a genuine experiment can be a supporting activity. Labelling that feeds routine retraining of a model already in production supports the product, not an experiment, and does not qualify.

Does using AutoML or neural architecture search make it R&D?

Not by itself. Those tools search a space someone defined. The question is whether the result was something published knowledge could not answer. A search can be the method of an experiment; it is never the reason the work is one.

Can we claim the licence fee for pretrained weights?

Possibly not. Expenditure on acquiring core technology (technology the R&D sets out to build new knowledge or products on), or the right to use it, cannot be a notional deduction under s 355-225(2), and licensed weights you build on can fall within that. Raise it with your accountant, who decides what goes in the ATO schedule.

How many training runs do we need to show?

There is no number. Show the runs that tested the hypothesis, including the ones that failed, with what changed between them and why. A single successful run with no plan behind it shows very little.

The R&D Tax Incentive is a self-assessment program. This page is general information, not tax, legal or financial advice; eligibility depends on your circumstances and you should seek independent advice for them.