Quick answer: Building the engine — encoding markup rules, integrating the ERP, fine-tuning an off-the-shelf model — is normally implementation that a competent professional could specify in advance. A core R&D activity may sit in the systematic investigation of whether features recoverable from your historical job records can predict labour hours to a defined technical measure, where that was unknown in advance and the work generates new knowledge. Exclusions in s 355-25(2) — market testing, efficiency surveys, and software for the dominant purpose of internal administration — may apply. 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 be administered under the current legislation.
Historical quoting records are rarely the asset they look like. Four properties recur across fabrication, trades and installation businesses:
Sparse: A few hundred genuinely comparable jobs, not a few hundred thousand.
Noisy: Recorded hours reflect who did the job, what the weather did, and which variations were absorbed rather than billed.
Inconsistent: Job types, units of measure and coding conventions changed across an ERP migration that lost the detail.
Selected: You hold actual hours only for the quotes that became jobs, at the price you quoted, which is not a random sample of the work you priced.
Whether a usable predictive relationship survives those properties, and at what accuracy, is often not answerable from the desk. That is where an AI quoting project may split unevenly: much of the work may be implementation, while a narrower part may involve an experimental investigation into achievable predictive performance.
The Candidate Activity Is a Measurement Question
Core R&D activities are experimental activities whose outcome cannot be known or determined in advance on the basis of current knowledge, information or experience, but can 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 are conducted for the purpose of generating new knowledge (business.gov.au).
The established-science limb does real work on a quoting project. Retraining and re-scoring until a metric improves does not, by itself, demonstrate a systematic progression of work. The experimental approach should be grounded in established scientific or statistical principles and supported by a stated hypothesis, an appropriate evaluation design and measures capable of testing that hypothesis.
Read against a quoting project, that puts most of the work outside the definition:
The build is implementation
The estimator interface, the ERP connection, the encoding of existing markup and contingency rules, the inference endpoint, the approval workflow — none of these has an unknown outcome. They have unknown effort. A competent professional in the field can determine in advance that they will work; the open question is how long they take.
Standard fine-tuning is normally implementation too
Fine-tuning an off-the-shelf model on your own job history follows a documented procedure. That alone does not settle the statutory question — a documented method does not make its performance outcome knowable in advance. What matters is whether the requirement being aimed at is an ordinary one that established practice is known to reach, in which case a competent professional would ordinarily determine the outcome in advance rather than by experiment. What such a professional may not be able to determine in advance is whether an unfamiliar accuracy requirement is reachable on this dataset — and where that is the question, and it is pursued to generate new knowledge, experimental work may still be required. AusIndustry's AI sub-guide to the software sector guidance makes the general point: applying existing techniques and tools in a well-understood way is not, by itself, an eligible activity.
The candidate activity is narrower than "the data"
Holding, consolidating or deploying a dataset is not an activity of this kind — data is a resource, not an investigation. The candidate core activity is the systematic investigation of whether specified features recoverable from the historical record can predict a defined quantity, such as labour hours, to a defined standard on data with the properties above.
Three Exclusions That Sit in the Path
Subsection 355-25(2) of the ITAA 1997 lists categories of activity that cannot be core R&D activities at all, however uncertain they felt. Three sit directly in the path of a quoting project:
1. Market research, market testing or market development: The subsection excludes market research, market testing or market development, and sales promotion including consumer surveys. What the activity measures is what tends to characterise it:
What is being measured
How it tends to be characterised
Whether the model predicts internal labour hours within a defined error threshold on held-out jobs
A technical measure — capable of supporting a core activity where the achievable accuracy was unknown in advance
Whether the model's quoted price wins the job
Customer response to price — on its face market testing
A/B testing two price points across real customers to see which converts
Market testing or market research on its face
Surveying customers on willingness to pay
Expressly caught — consumer surveys
Teams describe the second and third rows as "experiments", and in the ordinary sense they are. But an activity of an excluded kind cannot be a core R&D activity, and a win-rate uplift is a commercial result rather than new technical knowledge. The research protocol and its records are separable from commercial pricing trials: one measures error against recorded internal cost, the other measures customer behaviour. Where both run in the same quarter, the company and its registered tax agent determine claim treatment and expenditure allocation.
2. Management studies or efficiency surveys: Also excluded. A quoting project scoped internally as "reduce estimator time per quote by 60%" or "standardise how the branches price work" describes a management or efficiency outcome. The statutory tests are applied to what the activity is on its facts rather than to the label in the business case — but a project whose only stated objective is an efficiency outcome has no technical measure to assess against.
3. The internal-administration exclusion: The same subsection excludes developing, modifying or customising computer software for the dominant purpose of use by the developer — or an entity connected with it, or an affiliate of it — in the internal administration of its business functions.
There is no four-part "internal use software" test in Australia: That is a United States concept with no application here. The Australian rule is the dominant-purpose test above.
"Internal" is not the trigger; dominant purpose is: A quoting engine used only by your own estimators is not automatically caught. The question is whether the dominant purpose of developing it was internal administration of business functions — a factual question about why the software was built. Where the exclusion does apply, it applies to that software development activity; it does not automatically dispose of a distinct investigation into predictability, which is assessed on its own facts.
What the Exclusions Do Not Do
Excluded categories are excluded from being core activities. That is not the same as being outside the claim altogether.
An activity may qualify as a supporting R&D activity under s 355-30 where it is directly related to a core R&D activity — and where the activity is of a kind excluded from being core under s 355-25(2), or produces goods or services, or is directly related to producing goods or services, only where it is conducted for the dominant purpose of supporting that core activity (business.gov.au).
Sequence matters. Where a core activity exists, data consolidation, labelling and the evaluation harness around it may attach as supporting work. Where no core activity exists, there is nothing to support.
A Worked Example: A Structural Steel Fabricator
Hypothetical and illustrative only. Not a client, not a ruling, and not a statement that these facts would be eligible.
A South Australian structural steel fabricator has eleven years of records: roughly 2,600 quotes, of which 430 became jobs with usable time records. It wants to quote non-standard assemblies — one-off fabrications with no close historical analogue — faster and more consistently.
Target variable: Actual labour hours adjusted for approved variations, not invoiced hours, because absorbed variations distort the invoiced figure.
Baseline: The existing rules-based estimate — tonnage multiplied by a per-tonne hour rate, plus a complexity loading applied by the estimator. On the holdout it produced a mean absolute error of about 31% of actual hours.
Threshold and holdout, fixed before any modelling: MAE at or below 15% on non-standard assemblies, measured on a temporal holdout of the most recent 90 jobs. Temporal rather than random, because coding conventions drifted twice in eleven years and a random split lets later conventions leak into training.
Trial 1: Twenty-four tabular features from the ERP job header — tonnage, member count, connection count, coating type, shop versus site. MAE 27%. Better than baseline, short of the threshold.
Trial 2, abandoned: Adding text features from the drawing register descriptions produced MAE 12% — apparently a pass. A leakage audit found three of the contributing fields (final invoice line count, variation register entries, actual delivery date) are only populated after the job completes and would be empty at quoting time. Removed, the same configuration returned MAE 26%. The trial failed, and it ruled out the proposition that description text alone carries the missing signal — the apparent gain was leakage, not information available at quote.
Trial 3: Decomposing the estimate by connection type — predicting hours per connection detail and summing — gave MAE 18% across the holdout, and 14% on the 140 jobs with complete connection schedules.
Result reached: The threshold was not met across non-standard assemblies generally; it was met on a bounded subpopulation. The knowledge generated was that the binding constraint is the completeness of connection schedules in the historical record, not the choice of model family.
Where the boundary falls: For this example, the candidate experimental activity is documented from the point at which the hypothesis, evaluation protocol and threshold are established through to the evaluation and recording of the experimental results. The rules engine, the estimator interface, the ERP integration and the deployment sit outside it. Data consolidation and the evaluation harness may be candidates for supporting R&D where they are directly related to the core activity and, where s 355-30(2) applies, are conducted for the dominant purpose of supporting it.
These facts would identify a candidate activity for self-assessment, not an entitlement. The activity boundary would be substantially narrower than the software project; expenditure treatment requires separate assessment.
The Artefacts This Kind of Work Produces
AusIndustry's record-keeping guidance expects records created as the work happens. In practice, a data-and-prediction investigation of this kind produces a specific and fairly short list of artefacts:
A protocol fixed before the holdout is sealed: The protocol should record the question, target variable, baseline, evaluation threshold and split rule before the relevant experiment. Any subsequent changes should be contemporaneously documented and technically justified.
A dataset manifest: Row counts, date range, inclusion and exclusion rules, and a checksum, so a result can be reproduced against the same data a year later.
An experiment log, one row per trial: Configuration, features, result against the threshold, and disposition — carried forward, or abandoned and why. The abandoned rows are the part of the record that shows what was not known.
Versioned feature definitions and scoring code: Feature engineering, split definitions and the evaluation script sit in version control alongside the model configuration. A result whose feature definitions cannot be reproduced is not a result.
A boundary in time: Engineers routinely work on the investigation and the build in the same week; the research record separates the two by activity, while the company and its tax adviser determine how expenditure is allocated.
The failure mode that recurs is leakage — a field populated only after job completion sits in the training data, the holdout score looks excellent, and the model performs at baseline on live quotes. The second is a threshold established only after the result, which weakens the evidence that the experiment was designed in advance to test a defined hypothesis.
Where an RSP Fits
business.gov.au describes Research Service Providers as scientific or technical service providers registered in specific fields that a company can engage to conduct R&D activities on its behalf (business.gov.au). On a quoting project, an RSP may be particularly useful when engaged early enough to assist with experimental design, technical R&D work and contemporaneous supporting records within its registered research fields.
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 — where total notional deductions fall below $20,000, the offset base 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, under s 355-100(2) of the ITAA 1997 (ATO). Using an RSP does not guarantee eligibility — you still self-assess, and an RSP supplies research capability, not tax advice. See claiming R&D under $20,000. Offset rates and the refundable and non-refundable tiers are covered separately in refundable vs non-refundable offset.
Frequently Asked Questions
Q: Is building an AI quoting engine eligible for the R&D Tax Incentive?
A: Building an AI quoting engine is not automatically a core R&D activity. Routine implementation components — such as encoding established pricing rules, ERP integration and interface development — are unlikely to qualify where their technical outcomes can be determined in advance. A distinct core R&D activity may exist where achievable predictive performance on the relevant historical data could not be known or determined in advance and could only be resolved through the required systematic progression of work. You self-assess against all statutory tests and exclusions.
Q: Is fine-tuning a model on our own job data core R&D?
A: Standard fine-tuning is normally implementation, because the requirement being aimed at is usually an ordinary one that established practice is known to reach, so the outcome can be determined in advance rather than by experiment. That the procedure is documented is not by itself the reason. It may qualify as a supporting R&D activity where it is directly related to a core activity and, where required, conducted for the dominant purpose of supporting it. Where an unfamiliar accuracy requirement on your particular dataset genuinely cannot be determined in advance, experimental work may still be involved, assessed on its own facts.
Q: Are pricing experiments on customers claimable as R&D?
A: Not as a core R&D activity. Section 355-25(2) of the ITAA 1997 excludes market research, market testing and market development, including consumer surveys, from being core R&D activities. Testing price points on real customers to see which converts is market testing on its face. Testing whether a model predicts internal labour hours within a defined error threshold is a different question and is assessed on its own facts.
Q: Does the internal-use software exclusion apply to an in-house estimating tool?
A: Australia has no four-part internal-use software test — that is a United States concept. The Australian rule in s 355-25(2) excludes software developed for the dominant purpose of use by the developer, an entity connected with it or an affiliate of it in the internal administration of business functions. Whether an in-house estimating tool is caught turns on the dominant purpose of developing it, which is a factual question about your circumstances.
Sources & Further Reading
ATO — Eligibility for the R&D tax incentive — the $20,000 lower bound and the substituted RSP/CRC base in s 355-100(2)
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. s 355-25 and s 355-30
Related: R&D for software and AI · what does not qualify · R&D for property and construction · what an RSP is · claiming R&D under $20,000 · refundable vs non-refundable offset · R&D Tax Incentive in Adelaide
Talk to Ignition Research if you are planning an AI or data-driven estimating project and need technical R&D support. As a Registered Research Service Provider at Lot Fourteen in Adelaide, we assist fabrication, trades and industrial software teams with experimental design, technical investigations and contemporaneous supporting records within our registered RSP scope. We do not determine R&DTI eligibility or provide tax advice: your company self-assesses and remains responsible for its own claim, with tax 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.