Quick answer: Moving a working model onto a microcontroller, camera, gateway or vehicle is usually implementation, not R&D. Applying a documented vendor quantisation or pruning toolchain to a supported device is generally implementation where a competent professional can determine the relevant technical outcome in advance from existing knowledge, information or experience. What may be a core R&D activity under s 355-25 of the ITAA 1997 is narrower: where the accuracy achievable at a hard power, memory, latency or thermal budget could not be known or determined in advance on the basis of current knowledge, information or experience, could only be determined by applying a systematic progression of work based on principles of established science, and that work 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 be administered under the current legislation.
The model works on a workstation, with as much memory and time per inference as it wanted. Now it has to run on a battery-powered node with kilobytes rather than gigabytes of RAM, inside a sealed enclosure that cannot shed heat, returning an answer before the next frame arrives. Teams reasonably describe that second phase as the hard part; difficulty is not the statutory test. Embedded work also produces measured, iterative artefacts that look like experimental evidence whether or not an experiment took place.
This article covers one question: whether the accuracy attainable at a fixed hardware budget had to be established experimentally on the physical device. It does not restate general AI and software eligibility — see R&D for software and AI — the work that falls outside the incentive (what does not qualify), or how expenditure and R&D assets are treated.
Routine Porting Is Generally Implementation
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; that 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 that is conducted for the purpose of generating new knowledge (business.gov.au). The established-science limb does real work at the edge: a compress, measure and adjust loop can be disciplined and still fail the test where the hypothesis rests on nothing more than what the last run returned, rather than on established science such as quantisation error and numerical precision, information loss under pruning, or the thermal behaviour of the enclosure. Applied to an edge deployment, many routine implementation activities may fall outside the core R&D definition:
The toolchain is published and documented: Post-training quantisation, quantisation-aware training, pruning, distillation, operator fusion and graph compilation are established techniques. Applying them using established methods where the relevant technical outcome can be determined in advance is generally implementation rather than core R&D.
The target is often already published: Published vendor, framework or model-zoo results may be relevant evidence of what a competent professional could determine in advance, particularly where the architecture, hardware, data characteristics and operating constraints are materially comparable.
Iteration is not experiment: Compressing, measuring, adjusting and re-measuring until the numbers land is an optimisation loop. A systematic progression starts from a hypothesis about something that could not otherwise be determined, not from a target already known to be reachable.
AusIndustry makes the same point in its software development sector guide and the companion guidance on AI-related activities: applying existing techniques to a new dataset or a new deployment target, however skilfully, is not by itself experimental.
Where the Uncertainty at the Edge Can Be Genuine
The question to put to each activity: at the outset, could a competent engineer with the published literature and the vendor documentation have determined what accuracy was attainable inside this budget? If yes, it is implementation.
Often closer to a core R&D activity
Usually not a core R&D activity
Establishing whether a required accuracy is attainable at a fixed memory, power and latency budget where no published result covers that combination of architecture, data and silicon
Reaching a published benchmark on a supported chip with the vendor's quantisation or pruning toolchain
A novel or substantially modified architecture whose degradation under aggressive compression cannot be predicted from published results
Compressing a standard architecture with documented quantisation behaviour
An unusual sensor modality — non-standard spectra, acoustic or vibration signatures, irregular fusion — where the informative signal may not survive the precision loss
Retraining a standard vision model on your own images with an established pipeline
A thermal or power envelope with no predictive basis: sustained duty cycles under passive cooling, where throttling and accuracy interact in ways the datasheet does not answer
Measuring power draw to confirm what the datasheet already states
"No one has done it on our device" is not automatically an unknown
Novelty of combination is common and is not the test on its own. Where existing knowledge — the literature, the vendor's characterisation of the silicon, your own prior builds — reasonably indicates what will happen, the outcome was determinable in advance even though nobody had assembled that configuration before.
Commercial pressure is not technical uncertainty
Bill-of-materials cost, launch dates, customer specifications and competitor products are business risks. They can sit alongside technical uncertainty but are not evidence of it — a distinction treated separately in Insights.
A Hypothetical: Mapping the Accuracy–Resource Frontier
Illustrative only. The figures are invented, not drawn from any client or published result, and nothing here says the activities would be eligible — that is a matter for the company's own self-assessment.
A manufacturer builds an inspection camera for a sealed, unventilated industrial housing. Its defect-classification model runs at 97.4% recall on a workstation in float32, 12 MB. The enclosure and battery fix four constraints: 512 KB working set, 40 ms per inference at the operating clock, 180 mW sustained draw, ambient inside the housing reaching 45 °C with the module throttling above an 85 °C junction. The product requires ≥94% recall at ≤2% false calls. One sensor channel is near-infrared, which the vendor's published quantisation characterisations do not cover.
Existing knowledge answers the visible-light case and is silent on the infrared channel at that footprint. The stated unknown was whether ≥94% recall was attainable inside all four constraints simultaneously. Hypothesis: it was, using channel pruning plus 8-bit post-training quantisation.
Trial
Configuration
Measured
Conclusion
1
60% structured channel pruning, uniform across both channels, INT8 post-training quantisation
480 KB, 27 ms, 165 mW; recall 96.2% visible / 88.1% infrared
Fails. Loss is not uniform across channels; it concentrates in low-contrast infrared defects. Uniform pruning ruled out
2
Same footprint, quantisation-aware training with per-channel scales
495 KB, 29 ms, 171 mW; recall 91.4% infrared
Fails. Improves but not to target; residual error traces to the first two convolution stages, where 8-bit activation scaling compresses infrared dynamic range
3
INT16 activations retained in the first two stages, INT8 elsewhere
508 KB, 41 ms, 205 mW; recall 94.6% infrared — but the housing reached throttle at 20 min of continuous duty and latency rose to 63 ms
Fails. Accuracy met, latency and sustained power not. Accuracy and sustained power are coupled through throttling, which single-shot bench measurement does not reveal
4
INT16 first stage only; inference gated to 8 fps by a low-cost motion trigger
501 KB, 34 ms, 148 mW sustained; no throttle over a 4-hour soak; recall 94.1% infrared
An operating point inside all four constraints was located; the attainable frontier for this architecture, sensor and silicon was measured
For this illustrative example, the candidate experimental activity is documented from the point at which the technical hypothesis and evaluation approach are established through to the trials, evaluation and recorded conclusion. Hardening that configuration into shipping firmware, board layout, driver work, OTA update and the app sits outside it. Field trials conducted to assess customer interest, preferences or market acceptance may fall within the market research or market-testing exclusion in s 355-25(2). A technical trial conducted in a customer environment should instead be assessed according to its actual purpose and the technical question being investigated.
The Harness Makes the Progression Measurable
The evidentiary package is the prior unknown, the hypothesis, the controlled method, the observations and the conclusions (business.gov.au on records). The harness is not itself that package; it is what makes measurements comparable across trials, so a progression can be shown as systematic rather than as unrelated adjustments. A long log of build sizes and latency numbers documents an optimisation.
Four things distinguish this discipline in practice:
Define the experimental work below the overall project level: Distinct technical questions should be separately identifiable where appropriate, with records sufficient to show the hypothesis, experimental work and conclusions relating to each.
Keep the evaluation environment sufficiently controlled and versioned: Depending on the experiment, useful records may include dataset splits, target hardware and firmware versions, measurement methods and relevant tolerances, together with records of any material changes during the work. Without appropriate records of the evaluation conditions, comparing results across trials may be unreliable.
Mixed time is separated as it is worked: On embedded projects the same two or three engineers move between the experiment and the shipping firmware inside the same day, in the same repository, on the same board. What works is a commit or branch convention tied to the activity, timesheet codes that match it, and a written apportionment method for shared work such as board bring-up. Retrospective reconstruction weakens contemporaneous support even where the underlying work was experimental.
Document the prior-knowledge assessment contemporaneously where practicable: Useful records may identify the vendor material, framework documentation, published results, papers or expert input considered, and what technical question remained unresolved.
Failed branches — the pruning strategy that collapsed recall, the pre-processing that did not survive quantisation — can corroborate that the outcome was not known in advance, where they were hypothesis-driven and recorded at the time rather than reconstructed later.
The Rig You Built to Run the Experiment
Edge experiments usually need hardware that is not the product: an instrumented bench, a thermal chamber or enclosure mock-up, a power-measurement harness, a bank of target boards.
Activities that are not core may qualify as supporting R&D activities where they are directly related to core R&D activities — and where the activity is of a kind excluded from being a core activity under s 355-25(2), 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). A bench created solely to conduct the experiment may provide strong evidence that its dominant purpose is to support the core R&D activity. Existing production equipment that is also used for experimental runs requires a separate assessment of the particular supporting activity, including its direct relationship to the core activity and, where applicable, its dominant purpose.
Where an RSP Fits, and the $20,000 Point
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). For edge AI the useful moment is upstream: fixing the constraints and the measure, freezing the harness, and separating the experimental question from the firmware sprint before work starts. See what an RSP is.
There is also a threshold point for smaller and earlier-stage hardware companies. R&D expenditure for the income year must generally be at least $20,000 — the lower bound in s 355-100(1) of the ITAA 1997 — 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, provided the services are within a research field for which the RSP is registered. Below $20,000 of total notional deductions the offset base is generally limited to the substituted base in the table in s 355-100(2): qualifying expenditure to a non-associate RSP for services in a field for which it is registered, plus eligible monetary contributions under the CRC Program. Other in-house amounts do not automatically join that base. Using an RSP does not guarantee eligibility — you still self-assess. See claiming R&D under $20,000. Offset rates and the refundable/non-refundable distinction are covered in refundable vs non-refundable offset.
Ignition Research can be engaged before experimental work begins to help define the technical unknown, design the test method and separate experimental work from product engineering. We are a Registered Research Service Provider at Lot Fourteen in Adelaide. We are not a registered tax agent: registration and claim decisions rest with your company and its tax adviser. Get in touch.
Frequently Asked Questions
Q: Is porting a machine learning model to embedded hardware eligible for the R&D Tax Incentive?
A: Usually not on its own. Porting a working model to a supported device with a documented runtime and toolchain has an outcome determinable in advance from current knowledge, which is the test in s 355-25 of the ITAA 1997. Eligibility may survive where the achievable accuracy inside a hard power, memory, latency or thermal budget was genuinely unknown, had to be established by systematic experiment, and the work was conducted to generate new knowledge. You self-assess.
Q: Is model quantisation or pruning an eligible R&D activity in Australia?
A: Applying an established quantisation or pruning toolchain to reach a published benchmark is implementation. It may form part of a core R&D activity where the achievable performance under the relevant architecture, data, hardware or operating constraints could not be known or determined in advance and could only be resolved through the required systematic progression of work.
Q: Is TinyML development claimable under the R&DTI?
A: There is no separate TinyML rule. The same tests apply: was the outcome knowable in advance, was the work a systematic progression from hypothesis to experiment, observation, evaluation and conclusion, and was it conducted to generate new knowledge? Constrained-hardware projects can meet that test; the label does not, and routine TinyML implementation may be engineering with a determinable result.
Q: Is a test rig built for edge AI experiments a supporting R&D activity?
A: It may be. Supporting activities must be directly related to core R&D activities, and where the activity is of a kind excluded from being a core activity, produces goods or services, or is directly related to producing goods or services, it must also be conducted for the dominant purpose of supporting the core activity under s 355-30 of the ITAA 1997. A thermal bench created solely to conduct the experiment may provide strong evidence that the activity was undertaken for the dominant purpose of supporting the core R&D activity. With no core activity, there is nothing to support.
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 · 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.