Writing an algorithm of your own is the classic picture of software R&D. The legal test is less about whether you wrote it than about what the world already knew when you started. An algorithm is novel for the R&D Tax Incentive only if a competent professional with access to worldwide published knowledge could not have told you whether your approach would work.
The prior-art question for algorithms
Algorithms are unusually well documented. Before starting, search:
- Academic literature and preprints, where most algorithmic results appear first.
- Patent databases, because many algorithmic methods are patented rather than published as papers.
- Public code repositories and library documentation, since a working open-source implementation is published knowledge. The open source page covers the other side of this.
- Expert advice, recorded in writing.
Keep the search dated. In Absolute Vision Technologies [2022] AATA 2319, a case about software scheduling activities, the Tribunal noted there was no evidence of a literature review confirming a knowledge gap existed. Hypotheses that read as project goals and documents prepared after the fact also counted against the company.
One point works in your favour: knowledge that is not publicly available or reasonably accessible does not count. A rival's private solution does not close the gap.
The reverse also holds. Finding no paper or patent that solves your exact problem is not enough on its own. The test is whether the outcome could be determined from current knowledge, information or experience, and an experienced practitioner can often infer how a variation on a known method will behave without anyone having published it. The search shows what was known; the reasoning about why that knowledge could not predict your result is what the claim rests on.
Calculable versus uncertain
Some properties of an algorithm can be worked out on paper: worst-case complexity, correctness for a defined input class, memory bounds. If a competent professional could calculate the result, the outcome was knowable, and the work is analysis rather than experiment. The stronger core activities sit where analysis gives out: how a heuristic behaves on real data distributions, whether an approximation holds its accuracy at scale, whether a method converges under conditions nobody has characterised.
The ATO's software alert TA 2017/5 lists testing the efficiency of algorithms already known to work as ineligible unless it is an experiment resolving a hypothesis. Choosing between known algorithms by benchmark is selection.
Timing matters too. The gap is judged on what was known when the work was done. If a paper solving your problem appears halfway through the year, work before it can still be an experiment, but carrying on after it means asking again whether anything is left to find out, and recording the answer.
Edge cases peculiar to algorithm work
- Business rules with a technical name. A pricing, ranking or matching formula whose weights express commercial judgement is ordinary logic. Tuning it against sales or clicks falls under the market testing exclusion explained on the A/B tests page.
- Implementing a paper. Reproducing a published method is applying known knowledge unless you have evidence its results do not carry to your setting.
- Patents. Patenting is a commercial and legal step, excluded from core R&D, and has no bearing on whether the experiment qualifies.
- When speed is the point. If the algorithm is fixed and the question is making it fast enough on your infrastructure, see performance and scale engineering.
If the stakes are high and the position unclear, an advance finding gives a binding view on eligibility before the claim.