Quick answer: Connecting a commercial large language model to a workflow is generally unlikely to be a core R&D activity on those facts — prompt engineering, API integration and standard retrieval plumbing follow known patterns — though every activity turns on its own facts. What may be a core R&D activity is work establishing whether a required reliability property — a rate of unsupported assertions on a specialised corpus, extraction fidelity on degraded documents, accuracy inside a fixed latency budget — was achievable at all, where the outcome could not be known or determined in advance on the basis of current knowledge, information or experience and could only be determined by applying a systematic progression of work based on principles of established science, and where the activity is conducted for the purpose of generating new knowledge. You self-assess.
7 August 2026 — this article describes the current rules. The 2026–27 Federal Budget announced proposed R&DTI reforms for income years starting on or after 1 July 2028. Until any amendments take effect, the R&DTI continues to operate under the current rules.
The recurring fact pattern: a business connects a commercial large language model to a workflow — support triage, document extraction, internal knowledge search — spends six figures of engineering time on it, and asks whether that is R&D.
Much of the work may be routine implementation. Where a distinct experimental activity exists within the project, a common evidentiary risk is that the reliability target or evaluation measure is documented only after results have been observed.
This article stays on the reliability question in LLM and retrieval-augmented systems. Whether AI and machine-learning development is eligible at all, and when an AI project moves beyond implementation, each have their own article in our Insights.
Where the Boundary Falls
Australian eligibility does not turn on how novel the technology is. A core R&D activity is an experimental activity whose outcome could not be known or determined in advance on the basis of current knowledge, information or experience, and could only be determined by applying a systematic progression of work that is based on principles of established science and proceeds from hypothesis to experiment, observation and evaluation and leads to logical conclusions — and which is conducted for the purpose of generating new knowledge (business.gov.au).
The established-science requirement is particularly relevant to LLM work: a disciplined tweak-and-rescore loop is not a systematic progression unless the hypothesis and the measure rest on principles of established science — here, what is known about retrieval, language modelling and statistical measurement — so an orderly sequence of trials does not, by itself, demonstrate the required systematic progression if the hypothesis and experimental approach are not grounded in established scientific principles.
Applied to a typical LLM project, that splits the work:
Standard API invocation is generally implementation where documented functionality is used as intended: Calling a commercial model through its published interface does not, by itself, establish a technical uncertainty. Any relevant R&D analysis depends on the specific technical problem being investigated.
Integration using established components and patterns is generally implementation where a competent professional can determine the outcome in advance: The use of orchestration, chunking, embeddings, vector indexes, re-rankers or guardrails does not by itself determine eligibility; the question is whether the work involves a genuine technical uncertainty.
Prompt iteration using established techniques is generally optimisation rather than core R&D where the expected technical outcome can be determined using existing knowledge: Repetition alone does not establish a systematic progression of work.
AusIndustry publishes both a software development sector guide and an AI-related activities sub-guide: the sector's instinct is to describe the technology, and the rules ask about the activity.
Where the Reliability Question Becomes a Real Unknown
Three families of question recur, and they share a shape: a required property, on your data, with no reliable predictive basis in published knowledge or your own experience. Grounding on a specialised corpus — benchmarks report performance on general text, not the achievable rate of unsupported assertions across a decade of internal engineering standards. Extraction fidelity on scanned faxes, photographed forms and handwritten annotations, where clean PDFs are solved and this document population is not. And accuracy inside a hard envelope — a call-handling response window, fixed on-premises hardware, a cost ceiling per transaction — which is a different question from whether the accuracy is attainable at all.
More likely to contain a core R&D activity
More likely implementation
Establishing whether a defined grounding or unsupported-assertion rate is achievable on a specialised corpus, where published results give no reliable basis
Reducing hallucinations by adding retrieval and a citation check using documented techniques
Determining achievable extraction fidelity on a degraded or adversarial document population, tested against a measure defined in advance
Extracting fields from clean, structured documents with an established pipeline
Determining whether required accuracy is attainable inside a hard latency, hardware or cost envelope with no reliable predictive basis
Optimising cost by caching, batching or switching to a smaller published model
Developing and testing a novel evaluation or verification method where no adequate measure of correctness existed
Prompt iteration, few-shot curation or chunk-size tuning to a known target; benchmarking vendor models to choose one
Two limits on the left column. Vendor selection by itself is generally not a core R&D activity where the exercise simply compares available products against known criteria. And difficulty is not the test: deadline pressure, data volume and integration complexity are business risks that can sit alongside technical uncertainty without evidencing it (what does not qualify).
The Artefacts This Kind of Research Produces
Records should be made as the activities happen and show they were conducted as registered (business.gov.au). In LLM and retrieval work, an evaluation harness can be useful evidence of the experimental progression. Depending on the activity, it may include representative test inputs, defined evaluation measures and repeatable scoring procedures, alongside the other technical records generated as the work progresses. Structured that way, the work generates:
A start and an end date: The commit fixing the target, measure and baseline; the entry recording the final trial scored and the conclusion drawn. Recording when the experimental activity commenced and when conclusions were reached can help establish its boundary from surrounding implementation work.
A named baseline: The score of the system as first built, on the same held-out set, before any trial — with the prior search that established the answer was not already available.
Variants mapped to the technical record: Keep sufficient versioned records to identify what changed between trials, the result observed and the conclusion reached. Git branches, tagged commits or experiment logs may be useful ways to do this.
Abandoned lines, kept: Records of unsuccessful or abandoned approaches, including the results observed and the conclusions reached, can provide useful evidence of the experimental progression.
A written apportionment method: Which time codes cover trial work and which cover ticketed production work, recorded while the hours are worked.
The common failure mode is a project with a rich git history of prompt and pipeline edits and no scored baseline: it describes what was built, not a measured unknown being resolved.
Supporting activities and the internal-administration exclusion: Activities that are not core may qualify as supporting R&D activities where they are directly related to core R&D activities — and where an activity is of a kind excluded from being core, produces goods or services, or is directly related to producing goods or services, only where it is conducted for the dominant purpose of supporting a core activity (business.gov.au). Where there is no core activity, there is nothing to support.
Where the system runs your own business functions, s 355-25(2) of the ITAA 1997 excludes from core R&D activities software development whose dominant purpose is the internal administration of business functions of the developer, an entity connected with it, or an entity affiliated with it. Internal knowledge search and back-office document processing sit close to that line. One point still circulates that should not: there is no current "multiple sale" exception to the exclusion — that wording came from the repealed s 73B of the ITAA 1936.
A Worked Example — Hypothetical
Illustrative and hypothetical. Not a ruling, not based on any client, and nothing here indicates a claim would be accepted; eligibility is self-assessed on the facts.
An Adelaide engineering services business builds an assistant answering staff questions from twelve years of internal technical standards, variation notices and site reports — 11,400 documents, roughly a third of them scans of paper originals.
The build (March–July): A commercial LLM through its API, a chunking and embedding pipeline, a vector index, a web front end and a reviewer queue. In use, answers are confidently wrong often enough that staff verify all of them.
The baseline (early August): A held-out set of 240 questions taken from a year of help-desk tickets, each with an answer confirmed by two senior engineers, and a rubric counting an answer defective if it contains any assertion not supported by a retrieved passage. The system as built scores 22% defective, answering 91% of questions rather than declining. Target recorded before any trial: 5% or below, coverage 85% or above. The team records why published benchmarks gave no basis for predicting whether that was reachable on a corpus holding superseded and superseding versions of the same standards.
The trials (August–December): Model version, question set and rubric held fixed; retrieval, document representation and verification varied:
Section-boundary segmentation instead of fixed windows: 22% → 17%, coverage 90%.
Hybrid dense and keyword retrieval with a re-ranking pass: 17% → 12%, coverage 88%.
Auto-generated per-document summaries added to the index: 12%, unchanged. Abandoned — ruling out index density as the binding constraint.
A verification pass suppressing any sentence not entailed by a retrieved passage: 4% defective, but coverage collapsed to 61%. Abandoned at that granularity; the log records the lost answers clustering on questions whose correct answer spanned a superseded and a current standard.
Retrieval restricted to the standard in force at the date referenced in the question, verification applied only to numeric values and clause references: 6% defective, coverage 84%.
The result: The target was not met: 6% and 84% against 5% and 85%. The recorded conclusion is that on this corpus the binding constraint is document supersession, not retrieval quality — a conclusion that did not exist at the outset.
Where the boundary sits: The routine system build may be implementation on these illustrative facts. The technical investigation preceding the experiments should be assessed separately: preparatory research may potentially qualify as supporting R&D where it is directly related to the planned core activity and satisfies the applicable supporting-activity requirements. For this example, the candidate experimental activity begins where the technical hypothesis and evaluation approach are established and continues through the experiments, evaluation and recorded conclusion. Assembling the held-out set may potentially qualify as supporting R&D where it is directly related to the core activity and, where s 355-30(2) applies, is conducted for the dominant purpose of supporting that activity. The web front end and reviewer queue were built to run the service: their status would depend on a separately evidenced connection to a core activity and, where applicable, on dominant purpose — production use alone does not establish that connection. On the illustrative timesheet, 310 of 1,260 engineering hours sit on trial branches and scoring runs, the rest on production tickets.
Where an RSP Fits
AusIndustry describes Research Service Providers as scientific or technical service providers you can engage to conduct R&D activities on your behalf, registered in specific fields of research (business.gov.au). For an LLM project, an RSP is useful before the sprint containing the experiment: framing the reliability question as a testable hypothesis, defining the measure, designing the harness, and fixing the activity boundary while the work is happening. Ignition Research does this from Lot Fourteen in Adelaide; we supply research capability, not tax advice.
R&D expenditure for an income year must generally be at least $20,000, qualifying expenditure incurred to a non-associate RSP may still form part of the offset where total notional deductions are below the usual $20,000 threshold — the substituted base in the table in s 355-100(2) of the ITAA 1997 is generally limited to qualifying expenditure incurred to a non-associate RSP for services in a field for which it is registered, together with eligible CRC Program contributions. Using an RSP does not guarantee eligibility — you still self-assess. See claiming R&D under $20,000. Offset rates, the refundable and non-refundable tiers and the intensity premium are set out in refundable vs non-refundable offset.
Frequently Asked Questions
Q: Is using ChatGPT or an LLM API in our product eligible for the R&D Tax Incentive?
A: Invoking a commercial LLM through its published API exercises documented functionality, so on those facts it is generally implementation. What may be eligible is a distinct activity establishing whether a required reliability property was achievable at all on your data, where the outcome could not be determined in advance and the work was conducted to generate new knowledge. You self-assess.
Q: Is prompt engineering R&D?
A: Prompt engineering is not automatically a core R&D activity. Routine prompt iteration using established techniques is unlikely to qualify where the outcome can be determined using existing knowledge. Prompt-related work may form part of a broader experimental activity where it is used to test a hypothesis addressing a technical outcome that could not be known or determined in advance.
Q: Is building a RAG system claimable R&D in Australia?
A: Building a retrieval-augmented system from documented components — chunking, embedding, a vector index, a re-ranker — is normally implementation. Eligibility may arise where an activity was directed at resolving whether a defined performance property was attainable on a specialised corpus, with the hypothesis and measure recorded before the trials. Supporting activities must meet s 355-30, including dominant purpose where the activity is of an excluded kind, produces goods or services, or is directly related to producing goods or services.
Q: What evidence do we need for an LLM R&D claim?
A: You need contemporaneous records sufficient to demonstrate the relevant statutory requirements. Depending on the experimental design, useful records may include the technical unknown and prior-knowledge assessment, the hypothesis, baseline or comparison method, evaluation measures, trial configurations and results, conclusions reached, unsuccessful approaches, and a reasonable method for distinguishing R&D work from other engineering activities.
Sources & Further Reading
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. ss 355-25, 355-30 and 355-100
Related: R&D for software and AI · what does not qualify · what an RSP is · R&D Tax Incentive in Adelaide · claiming R&D under $20,000 · refundable vs non-refundable
Talk to Ignition Research before the next quarter's sprint. As a Registered Research Service Provider at Lot Fourteen in Adelaide, we work with teams building LLM and retrieval systems to state the reliability question as a hypothesis in advance, design the harness that measures it, and mark the boundary between experiment and build while the work happens. We are not a registered tax agent: your company self-assesses and remains responsible for its own claim, with advice and lodgement handled by your tax adviser. Get in touch.
This article is general information from a Registered Research Service Provider about the R&D Tax Incentive. It is not tax, legal or financial advice; eligibility depends on your circumstances and you should self-assess and seek your own advice.
Thinking about a project like this?
If you're weighing up an AI, software or technical improvement project and can't tell yet whether it's implementation or research, start with a quick read on where it sits.