AI projects can qualify. The label does not do it.

Yes, when the project seeks an advance in the field and the team faced technological uncertainty that a competent professional could not readily resolve, for example whether a model could reach a required accuracy on difficult data. Using a standard model for a standard task is routine. The question is what was unknown, not whether the word AI appears in the project name.

Training can be part of R&D, when it is the way the team investigates an unresolved problem, such as how to make a model work with scarce or noisy data. Re-running a known training recipe on new data to get a standard result is routine. The records should show the experiments, the results that failed and why the approach changed.

Most cloud costs for R&D can be, for accounting periods beginning on or after 1 April 2023, when cloud computing and data licences were added to the qualifying cost categories. They qualify where the capacity is used for R&D, not for the qualifying indirect activities. Capacity that serves live customers does not. Your billing needs to be traceable to the development and testing of the uncertain work.

Data licence costs are a qualifying category for accounting periods beginning on or after 1 April 2023, where the data is used for R&D. Buying a dataset to train and test a model in an uncertain project fits. Data bought for ordinary operations, or to resell, does not. We tie each licence to the project that used it and keep the contract.

Using one, on its own, is not. The Government's guidelines say using science or technology is not an advance in itself. Work built around a model can qualify where you had to resolve uncertainty to get it to perform a task it could not do as supplied, and the answer was not readily available to a competent professional.

Usually not. Routine collection, cleaning and labelling of data is preparation. It can fall within a qualifying project where the data problem is itself technically uncertain, for example finding a way to extract a usable signal when standard methods fail. We look at whether the cleaning was solving a technical problem or doing a job.

It is hard to make a case. Writing and adjusting prompts to get better outputs from a model you have bought is generally routine experimentation. It could form part of a qualifying project where the team was trying to overcome a real technical limitation with no known solution, and kept detailed records of what they tried and what it showed.

With a written statement of the problem and the target, experiment logs and tracked runs, evaluation results including those that missed, and notes on why you changed approach. Version history and model cards help. We interview the data scientists and engineers, because what they say about what was unknown at the start is the core of the technical case.

GPU capacity rented in the cloud is a form of cloud computing, which can qualify where it is used for R&D. Rented capacity used to run a product for customers does not. If you buy hardware, it is capital expenditure and is excluded from R&D relief, though a capital allowance may apply. Keep invoices that show which project used which capacity.

HMRC's overseas restriction is described in terms of contracted-out R&D and externally provided workers, and cloud computing is listed as a separate cost category. We do not assume the answer, and check the particular arrangement against HMRC's guidance before including the costs. If an overseas contractor does development for you, that is a different matter, and the restriction applies.

They can. A competent professional is someone with knowledge and experience in the relevant field of science or technology who could judge whether something is uncertain. For an AI project that is usually a senior data scientist or machine learning engineer. They must be able to explain the uncertainty to HMRC, which is why we interview them.

Yes. The Government's guidelines say a project can be R&D even if it fails. A model that did not reach its target, with records showing what was tried, demonstrates the uncertainty was real. The costs of the work are claimable. What matters is that the work was a systematic attempt to resolve a real technical question.

Generally not. Policies, risk assessments, ethics reviews and compliance documentation are not scientific or technological work. Technical work to build a measurement or control method, for example to detect bias in a way no existing tool can, could qualify if it meets the R&D test. We would look at the technical project, not the governance wrapper.

Licence fees for software used in R&D can qualify, in a reasonable share where use is partly for R&D. Usage-based fees for an API may be treated as software, data or cloud costs, depending on what they buy. Fees incurred for a live service do not qualify. We classify each one against HMRC's cost categories.

HMRC applies the same R&D guidelines to AI as to any technology, and information and communication make up 27% of claims on its provisional 2024-25 figures. We have not seen AI-specific guidance. A claim describing a project simply as using AI will not persuade. A claim that explains what was uncertain, and how the team resolved it, will.