What qualifies for R&D Tax Credits in AI
HMRC assesses AI R&D claims on an activity-by-activity basis against the BIS Guidelines. The central question is whether a competent professional in the relevant field could have readily deduced the solution using existing knowledge — if yes, the activity is routine and does not qualify. Qualifying activities in AI typically include:
- Developing novel model architectures or training techniques not available in the published literature at the time work began
- Resolving training instability (loss divergence, gradient issues, convergence failures) where the cause is genuinely uncertain
- Building bespoke data pipelines to handle domain-specific edge cases that existing frameworks cannot address
- Designing novel evaluation frameworks for domain-specific AI outputs where no established benchmark exists
- Research into interpretability, safety, or robustness properties of models where the theoretical basis is unsettled
What does not qualify
Applying an existing model to a new dataset using standard fine-tuning techniques is routine. Wrapping an API (OpenAI, Anthropic, Google) in a product interface is not R&D. Prompt engineering, even at significant scale, does not meet the technological uncertainty test. HMRC has increased scrutiny of AI claims since 2023, and any claim in this sector should be supported by contemporaneous project-level documentation.
GPU and cloud computing costs
From April 2023, cloud computing costs used directly in qualifying R&D are eligible. Training runs, evaluation runs, and development environments on AWS, GCP, or Azure that are part of a qualifying project can include the GPU instance cost. Production inference serving live users is not claimable.
ERIS for R&D-intensive AI companies
Loss-making AI companies that spend at least 30% of total expenditure on qualifying R&D can access Enhanced R&D Intensive Support (ERIS) at a 27% net effective rate, compared to 15% under the standard merged scheme. This is particularly relevant for pre-revenue AI research companies and early-stage model developers.
Frequently asked questions
Fine-tuning an existing foundation model using standard techniques (LoRA, RLHF on a commodity base model) is unlikely to qualify on its own, because the competent professional test requires the work to go beyond what a skilled AI engineer could readily deduce. However, if the fine-tuning process involves resolving genuine uncertainty about training stability, novel data pipeline design, or domain-specific alignment methods that are not well documented, those specific activities may qualify on an activity-by-activity assessment.
Yes, provided the qualifying work is contracted R&D that the AI consultancy itself is bearing the cost and risk of. Under the merged scheme from April 2024, where a consultancy is engaged on a fixed-price contract to solve a technically uncertain problem and absorbs the cost if the approach fails, the R&D expenditure is claimable by the consultancy, not the client.
Yes. From April 2023, cloud computing costs (including GPU instances on AWS, Azure, or GCP) used directly in qualifying R&D are eligible under the merged scheme. Training runs that are part of a qualifying project can include the GPU cost. Inference in a live commercial product is not claimable.
AI and machine learning companies typically fall within the Information and Communication sector for HMRC purposes, where the 2024 average claim is around £57,000. R&D-intensive AI businesses that cross the 30% expenditure intensity threshold can access ERIS at 27% net, which can materially increase the recovery on qualifying GPU and payroll costs.
Check your AI company's R&D eligibility
Use our 2-minute eligibility checker for an indicative recovery range, or request a free assessment introduction to an HMRC-registered specialist.