R&D Tax Credits for AI Systems: What Qualifies and What Doesn't

The UK's R&D tax credit scheme - now operating under the merged RDEC/SME scheme introduced in April 2024 - provides tax relief on qualifying research and development expenditure. The headline rates are significant: the merged scheme provides a 20% above-the-line tax credit for most companies, with enhanced rates for R&D-intensive SMEs in some circumstances.

For organisations developing AI systems for CNI applications, the first question is whether the activity qualifies as R&D in HMRC's sense. The statutory definition requires that the activity seeks to achieve an advance in science or technology through the resolution of scientific or technological uncertainty. This is a specific and sometimes misunderstood test.

The key term is 'technological uncertainty'. This does not mean that the outcome of the project is uncertain - many perfectly ordinary development projects have uncertain outcomes. It means that the answer to the technological question being addressed is not known or readily deducible by a competent professional in the field. In AI development, technological uncertainty is common and often significant: uncertainty about whether a particular architecture will achieve the required performance on a specific task, uncertainty about how to adapt a technique from one domain to another, uncertainty about the computational feasibility of a proposed approach at the required scale.

What typically qualifies in AI development for CNI: development of novel AI architectures for specific CNI applications; work to adapt AI techniques to domain-specific constraints (safety requirements, real-time performance requirements, classified data environments); development of AI systems that must operate in conditions that existing techniques do not handle well; research into AI explainability and interpretability for regulated environments; and development of human oversight mechanisms for AI systems in high-stakes settings.

What typically does not qualify: routine application of established AI techniques to new datasets; development that uses AI as a tool but where the AI itself is not being advanced; implementation of AI systems developed by others without modification; and commercial activities adjacent to R&D (marketing, sales, business development) even where they are related to AI products.

The claim process requires technical documentation - a project description that explains the technological uncertainty being addressed and the work done to resolve it - and financial analysis that allocates qualifying expenditure to the R&D activity. Many AI development organisations systematically under-claim because they apply too conservative a definition of qualifying activity. A rigorous R&D tax credit assessment frequently reveals qualifying expenditure that has not previously been claimed.