Quick answer: Configuring a forecasting product, fitting a standard model to clean data, or building a screen that ranks options is generally unlikely to be a core R&D activity on those facts, subject to the activity's own facts and the statutory tests — on those facts the outcome could ordinarily be determined in advance on the basis of current knowledge, information or experience. A core R&D activity may sit somewhere narrower: establishing whether a required predictive performance was achievable at all on your data, where that could only be determined by applying a systematic progression of work, and the activity was conducted for the purpose of generating new knowledge. Two exclusions in s 355-25(2) of the ITAA 1997 must be considered first where relevant: market research or market testing, and developing, modifying or customising computer software for the dominant purpose of the claimant's internal administration. Neither necessarily applies — characterisation turns on the facts of the particular activity, and you self-assess.
10 August 2026 — this article describes the current rules. The 2026–27 Federal Budget announced reforms to the R&DTI that will apply to income years starting on or after 1 July 2028. Until then, the program continues to be administered under the current legislation.
Decision-support projects are the most commonly mis-framed AI claims we see. The system forecasts demand, scores a risk, or ranks the options and hands a recommendation to a human, who decides. The build is expensive and the business impact is real, and neither is what the R&D Tax Incentive asks about.
A core R&D activity has to satisfy both limbs: the 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 the activity is conducted for the purpose of generating new knowledge (business.gov.au). The established-science limb does real work in this field: a disciplined cycle of tuning a model and measuring the result does not satisfy it on its own — the progression has to rest on principles of established science, such as the statistical or machine-learning theory that explains why the chosen design could reach the operating point on data like yours. Forecasting projects struggle against the first limb not because forecasting is easy, but because nobody seriously doubted a usable forecast could be produced. The doubt was about accuracy being good enough to be worth it, and about effort and integration.
This article stays inside forecasting and decision-recommendation systems; pricing engines and language-model reliability have their own treatment in our Insights.
Start with the Exclusions, Not the Core Test
Several categories in s 355-25(2) — activities that cannot be core R&D activities at all, however experimental they look — sit directly across the typical decision-support roadmap:
Market research, market testing and sales promotion
Subsection 355-25(2) excludes market research, market testing and market development, and sales promotion including consumer surveys. Read that against a decision-support roadmap and some activities may fall within the exclusion where they are directed at consumer interest, preferences, market acceptance or promotion — including uplift modelling, propensity-to-buy scoring, A/B tests on conversion, elasticity or promotional-response work when used for those purposes.
The line worth drawing is between research into customer behaviour and operational forecasting. Estimating how a segment responds to a promotion, or what a new region would buy if entered, is research into the market. Forecasting how many units of a stock line will move over a lead time so a reorder point can be set is an operations-planning quantity, even though customer demand is what ultimately drives it. Neither characterisation follows from the subject of the forecast: the same demand number can serve either purpose, so characterisation has to be established from the facts of the activity — why it is being run and what it feeds — not inferred from the fact that customers appear in it.
These are frequently the most statistically sophisticated activities in the business. Sophistication is not the test. Where the activity is properly characterised as market research, market testing or sales promotion, the exclusion applies on its face.
Software for the dominant purpose of internal administration
The same subsection excludes developing, modifying or customising computer software where the dominant purpose of doing so is the internal administration of business functions of the developer, an entity connected with the developer, an affiliate of the developer, or an entity of which the developer is an affiliate, and decision-support tools run the claimant's own business functions more often than not: a rostering recommender, a stock-replenishment engine, an internal collections prioritiser. Two points get lost: the exclusion reaches software development, modification or customisation carried on for that dominant purpose, not everything a company does internally, and its group scope is defined ("connected with" or "affiliate"), not the broader "associate" concept used elsewhere in Div 355.
Management studies and efficiency surveys
Also excluded, and a decision tool built to find where time, money or capacity is being lost sits close enough to that description to be tested against it rather than relabelled. See what does not qualify.
When the Achievable Performance Is Genuinely Unknown
Clear the exclusions and the core test still has to be met. In our experience the condition it describes comes from the problem rather than the technology. Three cases recur:
A rare, delayed or weakly observed target: A failure that occurs a few times a year, a default that materialises eighteen months after the decision, a "good outcome" recorded only as a proxy. Where ground truth is that scarce, censored, or contaminated by the decisions already made from it, whether any signal of useful strength exists in the records is a genuine unknown.
A distribution that shifts: Systems that recommend actions change the data they are later trained on. Where performance must hold across a regime change — a new site, a new customer mix, a post-disruption demand pattern — whether it can hold, and by what mechanism, may be an experimental question. Retraining on a schedule is not.
A threshold rather than an average: These systems run at an operating point: escalate above this score, hold stock above that forecast. A model with a respectable headline metric can be useless there.
In each, the technical uncertainty needs to be expressed through a testable hypothesis and investigated through the required systematic progression of work, with a negative result remaining a possible outcome.
Reproducing a Human Judgement Faster
The most common decision-support pitch is that an experienced person currently makes this call and the model will make it faster and at scale. Where a competent human already performs the task reliably, that fact alone does not establish that a software implementation has a determinable outcome. Where current knowledge and available methods already show how the task can be reproduced from the information available to the system, and the remaining uncertainty concerns engineering effort, data access, latency or adoption, the work is generally unlikely to be a core R&D activity on those facts.
On those facts the work is generally unlikely to be a core R&D activity, subject to the activity's own facts and the statutory tests — and where the purpose is to reduce the cost of an internal business function, the management-studies and internal-administration exclusions are live as well. The position can differ where the person relies on information the system cannot observe, so whether the judgement can be recovered from the recorded data at all is open; that has to be framed before the work, not afterwards.
A Worked Hypothetical: Intermittent-Demand Replenishment
Illustrative and hypothetical: the figures are invented to show the shape of an experiment, not to describe any client, and nothing here indicates eligibility.
A distributor carries 14,000 lines, about 5,200 of them long-tail — demand is intermittent and most weeks are zero.
Baseline: The incumbent ERP reorder-point rule uses a 12-week moving average. Replayed over the development and validation windows it delivers a 91.4% line fill rate on those lines at 9.6 weeks average on-hand cover.
Requirement, recorded before any modelling: On 4 August the project charter states the target: 96% line fill on the long tail at no more than 10.0 weeks average cover — +4.6 points of service for no more than +0.4 weeks of cover. It also records what happens if the target is missed: the reorder rule stays as it is.
The search, recorded the same week: Examined the published intermittent-demand forecasting literature (Croston family, pooled/hierarchical count models, quantile-based inventory policies), demand collections, and licensed method documentation. None answered whether 96% fill at ≤ 10.0 weeks cover is reachable on this catalogue's sparsity and weekly review cycle. Hypothesis: pooling across lines, with the reorder point derived from the lead-time demand distribution, reaches that operating point on this data.
Protocol & controls: Three years of history split three ways: development window, validation window, and final 26 weeks sealed (opened once only after clearing the target on validation). Held constant: extract, lines, lead-time distribution, review cycle, fill definition. Varied: demand model and reorder-point derivation.
Trials:
Trial 1 — Intermittent-demand method of the Croston family (per line): 93.1% fill at 9.8 weeks on the validation window: better than baseline, short of target.
Trial 2 — Gradient-boosted regression per line on calendar/lead-time features (Failed): 92.4% fill at 11.7 weeks. Per-line fitting had too few positive observations to estimate from, ruling out per-line supervised learning.
Trial 3 — Pooled model (lines grouped by family/movement class, shrunk toward group): 95.2% fill at 9.9 weeks. Closer, still short.
Trial 4 — Pooled model with reorder point from simulated lead-time demand quantile: 96.4% fill at 9.8 weeks on validation. On the sealed 26-week holdout: 96.3% fill at 9.7 weeks (vs 91.1% for incumbent). Requirement met on protocol.
Where the work divides: The experimental work spans the trial sequence under the frozen protocol (requirement recording to sealed holdout read). The ancillary work includes the harness, extract pipeline, and evaluation rig. The deployment work (ERP rollout, planner interface, retraining) sits outside as internal administration. Promotional-uplift modelling built alongside is market research. Supporting activities must satisfy s 355-30, including dominant purpose where production or exclusions are engaged.
What This Discipline Actually Produces
AusIndustry sets out its expectations for record keeping, and publishes both a software development sector guide and an AI-related activities sub-guide. Beyond that, structuring this kind of work produces four artefacts:
A requirement with a measure and an operating point: Dated before the first trial. Without the operating point it is a preference, and a preference cannot fail.
A dated search: What was looked at, when, what it did not answer, and why existing results were not transferable to your data, service definition, and operating point.
A frozen holdout protocol: Written before Trial 1 specifying rows, weeks, access, and trigger. Iterate on validation; leave the holdout sealed until a candidate clears the target.
Versioned assets and logs: Extract, feature definitions, configuration, evaluation script, and contemporaneous time tracking separating trial work from platform work.
Where an RSP Fits
A Research Service Provider is a scientific or technical service provider, registered in specific fields, that a company can engage to conduct R&D activities on its behalf (business.gov.au). From Lot Fourteen in Adelaide, our work on projects like this one is upstream of any claim: separating the market-facing and internal-administration work from the technical question, stating the unknown and its measure before the sprint, and designing the trial sequence so the record exists as a by-product of running it.
There is also a threshold point for smaller claimants: R&D expenditure for the income year must generally be at least $20,000, and 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. Using an RSP does not guarantee eligibility — you still self-assess. See claiming R&D under $20,000. Which offset applies, and on what conditions, is covered separately in refundable vs non-refundable offset.
Talk to Ignition Research before you register a forecasting or decision-support project — as a Registered Research Service Provider at Lot Fourteen in Adelaide, we help teams separate the market-facing and internal-administration work from the technical question, state the unknown and its measure in advance, and generate the record 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.
Frequently Asked Questions
Q: Is building a demand forecasting model eligible for the R&D Tax Incentive?
A: Usually not on its own. Configuring a forecasting tool or fitting a standard model to clean data is generally unlikely to be a core R&D activity, subject to the activity's own facts and the statutory tests, because the outcome could be known in advance on current knowledge. A core R&D activity may exist where the achievable performance on the actual data was genuinely unknown and could only be determined by a systematic progression of work carried out for the purpose of generating new knowledge. You self-assess.
Q: Is demand forecasting excluded as market research?
A: It can be. Section 355-25(2) of the ITAA 1997 excludes market research, market testing, market development and sales promotion (including consumer surveys) from being core R&D activities. Work directed at what customers will buy, at what price, in response to what offer generally falls inside that description however advanced the method. Operational forecasting — how much of a stock line to hold at a site over a lead time — may not. It does not turn on whether customers feature in the forecast; it turns on what the activity is actually for, and you self-assess that characterisation on the facts.
Q: Does the internal-administration exclusion apply to an in-house decision tool?
A: It may. The exclusion covers developing, modifying or customising computer software where the dominant purpose of doing so is the internal administration of business functions of the developer, an entity connected with it, or an entity affiliated with it — and many decision-support tools run the claimant's own planning, rostering or credit functions. It is a dominant-purpose test with a defined group scope, so it turns on the facts rather than on the tool being in-house.
Q: Is automating a human judgement an eligible R&D activity?
A: Generally not. If an experienced person already performs the task reliably, a level of achievable performance is demonstrated, and building a faster version is an engineering exercise. The position can differ where whether the judgement is recoverable from the recorded data is itself unknown — but that has to be framed with a measure before the work, not argued afterwards.
Sources & Further Reading
ATO — Tax Reform: better targeting the R&D Tax Incentive — the announced measures and the statement that the measure is not yet law
industry.gov.au — Research and Development Tax Incentive — the proposed changes would apply to income years starting on or after 1 July 2028 if enacted
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 · claiming R&D under $20,000 · refundable vs non-refundable offset · R&D Tax Incentive in Adelaide
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.