Quick answer: Building the pipeline, the dashboard and the KPI set ordinarily fails the core R&D activity test because the outcome could be known in advance on the basis of current knowledge, information or experience. Separately, particular reporting and internal-administration activities may fall within the exclusions in s 355-25(2) of the ITAA 1997 — a different reason, tested on its own facts. What may be core R&D is narrower: work where a new inferential capability's achievable performance was genuinely unknown — estimating a state nobody measures from sparse or biased sensing, detecting rare events where ground truth is scarce, or transferring a model to sites with different sensor populations. You self-assess.
7 August 2026 — this article describes the current rules. The 2026-27 Federal Budget announced R&DTI changes scheduled to start 1 July 2028 (ATO — Tax Reform: better targeting the R&D Tax Incentive); those are not yet law.
Where a telemetry pipeline uses established ingestion, storage and integration methods, a competent professional may be able to determine the relevant technical outcome in advance. In that case, uncertainty about effort, schedule, vendors or data quality does not by itself establish core R&D. The uncertainty was about effort, schedule, vendor and data quality, not about whether the technical outcome was achievable.
That is what the definition asks: 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, conducted for the purpose of generating new knowledge (business.gov.au). The established-science limb is where a lot of analytics work comes apart: a sweep over model architectures, features and hyperparameters steered only by a validation score is a disciplined search, but it is not a progression grounded in traffic-flow theory, network hydraulics, structural mechanics or statistical estimation theory, and the hypothesis has to say why the approach should recover the quantity at all.
The exclusion list in s 355-25(2) is a second, independent filter, and two categories sit across the analytics roadmap: management studies or efficiency surveys — which may include benchmarking or analysis directed at operational efficiency, depending on the substance and purpose of the activity — and software activities where the dominant purpose is the internal administration of business functions of the developer, an entity connected with it, or an entity affiliated with it, a test that is software-specific with a defined group scope and which we treat properly in a separate article in our Insights.
Routine telemetry work is not specifically excluded merely because it involves telemetry; rather, it may fail the core R&D test where its outcome can be determined in advance using existing knowledge or established methods. Which route an activity fails on matters downstream: activities of an excluded kind face the dominant-purpose limb if they are put forward as supporting activities. Our what does not qualify page sets out the exclusion list in full.
Where a Core R&D Activity May Arise
Across traffic, transit, water, power and structural-condition work, what may remain is the question of whether a new inferential capability could reach a required level of performance at all, where that could only be established by systematic experiment.
Having a metric does not establish that question. Useful evidence may include a contemporaneous assessment of existing knowledge, a baseline or comparison capability where relevant, the unresolved technical constraint, and an experimental evaluation method appropriate to the hypothesis being tested.
Usually implementation (not core R&D)
May contain a core R&D activity
Ingesting loop-detector, AVL, SCADA, smart-meter or IoT telemetry into a warehouse or lakehouse
Estimating an unobserved network state (turning movements, leak location, structural condition) from sparse or spatially biased sensing, where achievable accuracy was unknown
Data-quality rules, deduplication, gap-filling with standard interpolation, schema and semantic layers
Establishing whether a state-estimation approach remains valid under the specific sparsity and bias of your sensor population, against a baseline and a stated measure
Dashboards, operational reporting, KPI definition, benchmarking depots, districts or corridors
Rare-event detection where labels are scarce and the achievable precision/recall trade-off could not be predicted from existing knowledge
Applying an off-the-shelf forecasting or anomaly package and tuning thresholds
Determining whether a capability transfers across sites with materially different sensor populations, or whether performance is irreducibly site-specific
Retraining an existing model on new data for a new site, expecting it to work
Novel sensor fusion where the achievable combined accuracy across modalities was an open question
Cost, energy or process-efficiency studies over the resulting data
May raise the management-study or efficiency-survey exclusion, depending on their purpose and substance
Volume, velocity, messiness and integration count are engineering difficulty; they can coexist with a genuine unknown but do not demonstrate one. Nor does the public-good objective: reducing congestion, cutting non-revenue water or extending asset life is legally irrelevant to whether an activity meets s 355-25.
What Actually Separates a Trial from a Tuning Exercise Here
Analytics teams generate a great deal of documentation — data dictionaries, lineage diagrams, model cards, sprint notes — that records what was built rather than what was unknown; AusIndustry sets out what records should show at business.gov.au. The problems below are specific to sensor-network data.
Design the validation split around the technical question: Where the hypothesis concerns generalisation across locations, assets or time periods, random row-level splitting may create leakage or overstate performance. Holding out appropriate geographic, asset or temporal groups may provide a stronger evaluation, depending on the data and hypothesis.
Incident labels may be biased toward the incumbent detection process: Where operator-confirmed labels are generated principally from alerts produced by an existing system, an evaluation may partly measure agreement with that system rather than the underlying event population.
Sensor drift and missing-not-at-random gaps: Calibration drift across a study window looks like a change in model performance; a reference check separates the instrument story from the model story. Telemetry gaps may be correlated with the conditions of interest — for example congestion, transients or heavy loading — so the missing-data mechanism should be considered when designing and interpreting the experiment.
Production runs and experimental datasets are not the same data: The nightly pipeline rewrites the tables a trial read. Versioned or otherwise reproducible data snapshots, together with records of the code/configuration and evaluation population used for each trial, can help make the experimental sequence reproducible and auditable. Where the same engineers work on the platform and the trials in one fortnight, the boundary is only recoverable if the two streams are recorded separately as the work happens — a record-keeping and allocation matter for the company and its registered tax adviser.
Supporting Activities
Activities that are not core may qualify as supporting R&D activities where they are directly related to core R&D activities. Under s 355-30 of the ITAA 1997, where the activity is of a kind excluded from being core by s 355-25(2), or produces goods or services, or is directly related to producing goods or services, it qualifies only where it is conducted for the dominant purpose of supporting a core activity (business.gov.au). The production-related limb may be relevant to analytics work. Where a particular pipeline, deployment, installation or delivery activity produces, or is directly related to producing, goods or services, the additional dominant-purpose test applies if that activity is being assessed as supporting R&D. Each activity must be assessed on its own facts.
A general-purpose data platform serving ordinary reporting and operations may present a different purpose analysis from a labelled evaluation dataset created specifically for an experiment. Each activity still needs to satisfy the direct-relationship requirement and, where applicable, the dominant-purpose test. If there is no core activity, there is nothing for supporting activities to attach to.
A Hypothetical Worked Example
Illustrative only. The parameters below are invented to show where an activity boundary falls; nothing here states or implies that such activities would be eligible or that a claim would be accepted.
A South Australian analytics company is engaged on an 8 km signalised arterial with 12 intersections. Sensing is stop-line and advance inductive loops, four to eight per approach, plus signal controller logs; turning movements are not measured. The baseline is a manual count programme — two four-hour peak surveys at three of the twelve sites once every three years, with those fixed proportions applied to every site and every time of day. The loop vendor's documentation and the published work reviewed address volume and occupancy estimation, not turning-movement recovery from loops alone at this spacing. The unresolved constraint is observability: whether turning proportions can be recovered from loop and phase data at all when loops sit on approaches and not on turn pockets at eight of the twelve sites.
Threshold and ground truth: The threshold comes from the decision the numbers feed, signal retiming — mean absolute error of 8 percentage points or better per movement per 15-minute interval, on intersections the method has not seen, at all times of day rather than peaks only. Ground truth is three weeks of temporary video at four intersections. Two approaches are compared against the baseline: a constrained flow-conservation reconciliation across the corridor, and a supervised model on loop occupancy, gap and phase-timing features.
Supervised model results: The supervised model reaches roughly 4 pp under a random split of 15-minute rows and 19 pp when trained on three video sites and tested on the fourth. That comparison is the informative result: it rules out a single cross-site supervised model on this feature set, because what it had learned was site-specific loop geometry rather than a transferable relationship. Intersection-level hold-out plus week-blocked splits are fixed for everything after it.
Addressing gap structure: The next iteration hits the gap structure — two advance loops fail intermittently in week two, and the failures correlate with sustained high occupancy, the congested conditions the retiming is for. Dropping incomplete intervals had removed those conditions from the evaluation set; restoring them with an occupancy-imputation variant moves the flow-conservation result from about 7 pp to about 11 pp.
The outcome: The outcome is mixed: flow-conservation reconciliation reaches roughly 9 pp on the held-out intersection inter-peak and 14 pp in the PM peak, better than the baseline in both windows and short of the 8 pp threshold in both. Whether the threshold is reachable from loop and phase data alone at this spacing is unresolved at the end of the sequence.
Where the boundary falls: For this illustrative example, the candidate experimental activity is documented from the point at which the hypothesis and evaluation approach are established through to the experimental iterations, evaluation and recorded conclusion. Over 14 weeks the programme runs roughly 1,900 engineer-hours, of which around 260 sit inside that sequence — trial design, feature construction for the trials, running and evaluating them, and the abandoned supervised branch. The rest is ingestion, the dashboard, the KPI pack, the map UI and works-order integration; the video-derived labelled evaluation dataset, around 120 hours, was built only for the trials. Whether any of this constitutes eligible core or supporting activities is for the company's own self-assessment, with advice and lodgement from its registered tax agent.
Where an RSP Fits, and the $20,000 Point
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 (business.gov.au); its software development sector guide and AI-related activities sub-guide cover this territory. An RSP may assist with experimental design, technical R&D work and contemporaneous supporting records within the research fields for which it is registered. Earlier involvement can help ensure that the technical question, hypothesis and evaluation approach are documented as the work progresses.
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 — where total notional deductions fall below that figure, the offset base is generally limited to the substituted base in the table in s 355-100(2) of the ITAA 1997, being qualifying expenditure to a non-associate RSP for services in a field for which it is registered, together with eligible CRC Program contributions (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/non-refundable split are set out in refundable vs non-refundable offset.
Talk to Ignition Research before you register or lodge — as a Registered Research Service Provider based at Lot Fourteen in Adelaide, we work with transport, utility and asset-analytics teams to separate the platform build from the experimental question, set the baseline, measure and hold-out protocol in advance, and record the trial sequence as it happens. Your company self-assesses and remains responsible for its own claim; we are not a registered tax agent. See also R&D Tax Incentive in Adelaide and R&D for software and AI. Get in touch.
Frequently Asked Questions
Q: Is building a data pipeline or dashboard eligible for the R&D Tax Incentive?
A: Ordinarily not as a core R&D activity, because the outcome could be known or determined in advance on the basis of current knowledge, information or experience. Separately, particular reporting and internal-administration activities may fall within the s 355-25(2) exclusions; those are two different reasons and each is tested on its own facts. Parts of the work may qualify as supporting R&D activities where they are directly related to a core R&D activity and, where the additional test in s 355-30(2) applies, are conducted for the dominant purpose of supporting that core activity. You self-assess.
Q: Is anomaly detection on sensor data R&D?
A: It may be, subject to self-assessment against all the statutory requirements and exclusions, where the events are rare, ground truth is scarce, and the achievable precision/recall trade-off on your signal could not be known in advance and was determined by a systematic progression of work against a stated baseline and measure. Tuning an off-the-shelf detector until the alert volume looks right generally will not qualify.
Q: Are management studies or efficiency surveys excluded from core R&D?
A: Yes. Subsection 355-25(2) of the ITAA 1997 expressly excludes management studies or efficiency surveys from being core R&D activities, regardless of how technically demanding they are, and benchmarking depots, districts or corridors may fall within this exclusion where the activity is substantively a management study or efficiency survey. Such work may still qualify as a supporting activity only where it is directly related to a genuine core activity and conducted for the dominant purpose of supporting it.
Q: Is retraining a model for a new site an R&D activity?
A: Routine retraining on new data, where it is expected to work, is deployment. It can become a core activity where whether a capability transfers at all across sites with materially different sensor populations was genuinely unknown, and that question was investigated through the required systematic progression of work using an appropriate documented evaluation method — including the configurations that failed.
Sources & Further Reading
ATO — Eligibility for the R&D tax incentive — the $20,000 lower bound (s 355-100(1)) and the substituted RSP/CRC base (s 355-100(2))
legislation.gov.au — Income Tax Assessment Act 1997 — Div 355, incl. ss 355-25 and 355-30
Related: R&D for software and AI · what does not qualify · what an RSP is · refundable vs non-refundable offset · claiming R&D under $20,000
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.