HMRC already answered the shape of this
The framework predates the current wave, and it holds. In HMRC’s machine-learning case study, choosing between learning paradigms did not qualify. The novel neural-network implementation did, because available approaches could not meet the requirement. The computer-vision study repeats the split: selecting an AI platform did not qualify, and the feasibility research on the network approach did. Swap in today’s nouns and the manual reads as if it were written for 2026. Model selection, prompt architecture and vendor evaluation sit on one side. Genuinely extending what the technology can do sits on the other.
The guidelines supply the reason. An advance means the field’s overall knowledge or capability. Uncertainty exists only where competent professionals could not readily resolve the question. A model’s documented capabilities, published techniques and vendor cookbooks are exactly what “readily available” means. The claim starts where they stop.
The build-versus-use divide, applied
- Rarely qualifying: integrating models through published APIs; prompt and context engineering to documented patterns; RAG built from standard components; fine-tuning by the book; and evaluation or vendor selection, which is expressly the non-qualifying shape in both HMRC studies.
- Potentially qualifying: novel architectures or training methods. Making established approaches work under constraints the literature says defeat them (latency, hardware, data regimes), with evidence of the failures. Genuinely new methods for evaluation, safety or reliability where published approaches demonstrably fall short. And AI-adjacent systems engineering past documented limits, the pattern that qualified the checksumming and certificate-authority work in HMRC’s other studies.
- The constant: the claim is a sub-project with a boundary, not a product with a theme. The software sector page carries the full method.
The evidence problem AI makes harder
In a field moving this fast, the baseline question decides claims: what was readily available to a competent professional at the time? Memory will not hold it. Two habits make AI claims defensible. Date-stamp the baseline: the model versions, papers and documented capabilities you consulted, and why they failed your requirement. And record the negative results, because the failed attempts are the uncertainty made visible. A claim that says “the models could not do this in March”, with March’s evidence attached, is very different from one asserting the same thing in retrospect, eighteen months of releases later.
The claims to resist
Four claims dominate the current market, and all four should be resisted. “We use AI so we qualify” is the category error HMRC’s studies exist to kill. Whole-product claims wrap a theme around one uncertain component. Compute bills get claimed wholesale across training, staging and production. And narratives assert novelty without a dated baseline, in a field where last quarter’s frontier is this quarter’s tutorial. The pressure to claim these is commercial; the enquiry that follows is not. Where the work is real (and plenty of it now is), the method on this page claims it safely. Where it is not, the honest page says what to do instead.
- HMRC, CIRD81960 and CIRD81980 (software guidance and case studies, including machine learning), Corporate Intangibles Research and Development Manual, gov.uk
- DSIT, Meaning of research and development for tax purposes: guidelines (2023), gov.uk
- HMRC, Check what Research and Development (R&D) costs you can claim (data licences and cloud computing), gov.uk
Frequently asked questions
We build products on top of large language models. R&D?
The test is where your work sits. Prompting, integrating and orchestrating existing models through their published interfaces is sophisticated use of available technology, HMRC's platform-selection examples fail exactly that shape. Work that pushes past documented capability, where the field's knowledge genuinely ran out and yours had to extend it, can qualify, evidenced like any other uncertainty.
Does fine-tuning or RAG count as advancing the field?
Not by the label. Applying established fine-tuning or retrieval patterns from the literature to your domain is adoption. The qualifying shape is the one in HMRC's fraud-detection study: a genuine capability gap the standard approaches could not close, and systematic work on the approach itself.
Can we claim our compute and data costs?
Training and experimentation compute within the uncertainty window can qualify under the data-and-cloud category (periods beginning on or after 1 April 2023), and data licences likewise where they directly contribute. Production inference and hosting cannot, and the apportionment discipline is the same as everywhere else.
Using AI tools to do our R&D faster, does that affect the claim?
Tools do not disqualify work, and never have. The uncertainty must remain scientific or technological rather than dissolved by the tool; where an assistant makes the answer readily available to a competent professional, that specific question stops being uncertain. Document what the tools could not do, which is where your claim lives anyway.