Accountants for AI and machine learning startups.

Computing and data costs behind qualifying work can go into an R&D claim. Whether the model work itself qualifies is a separate question.

Talk to us Try the R&D estimator

We act for UK AI and machine learning startups. We include cloud computing and data licence costs in R&D claims where they qualify, assess whether the model work itself qualifies, and forecast cash around training costs that arrive in large, irregular amounts.

Two rules affect how much an AI company can claim for R&D.

The first is that cloud computing and data licence costs qualify for R&D relief. They were excluded before April 2023, and for a company whose largest cost after payroll is computing, they can be a large part of the claim. The second is that fine-tuning an existing model, using another company's model through its interface and building a product around it is usually ordinary development in HMRC's terms, and does not qualify.

A worked example of computing costs in an R&D claim

A loss-making UK AI company, and a small or medium-sized enterprise (SME) for R&D purposes: fewer than 500 staff, and turnover of no more than €100 million or a balance sheet of no more than €86 million, counting linked and partner businesses. Total expenditure £700,000. Research payroll on technically uncertain model work is £180,000. The computing and data licence costs behind that work are £240,000. Qualifying R&D is therefore £420,000.

Intensity: £420,000 ÷ £700,000 = 60%, so the intensive route is available.
Claim including computing costs: £420,000 × 186% = £781,200 surrendered, at 14.5% = £113,274.
Claim on payroll alone: £180,000 × 186% = £334,800, at 14.5% = £48,546. The company still meets the 30% test, because the test counts every cost that could be claimed, whether or not it is.
Loss check: the £781,200 surrendered cannot be more than the trading loss. With total expenditure of £700,000 plus the extra deduction of £361,200 (86% of £420,000), the trading loss is £1,061,200 minus the company's income, so income for the year has to be £280,000 or less.
Cap check: the £113,274 is paid in full only if PAYE and National Insurance for the year are at least £31,092, because £20,000 + (300% × £31,092) = £113,276. Below that, the credit is limited to the cap unless the company qualifies for the exemption from it.

Leaving the computing costs out would cost this company £64,728. Cloud computing and data licence costs have qualified since April 2023, so check that a first claim includes them.

Illustrative figures. We run the calculation on your own numbers before anything is filed.

Run your own numbers

Where an AI company's money goes

Computing costs: Training models and running them can cost large, irregular amounts, and the cost qualifies where it supports qualifying R&D. Running the model behind a live product does not qualify, so track the split as you go.
Data and licences: Datasets and data licences bought for development have qualified since April 2023, and they are easy to leave out of a first claim.
Research payroll: Usually the largest cost, and the part of the claim that carries the technical case. The claim rests on which people spent what share of their time on technically uncertain work.
Cash that moves in steps: A training run is a large one-off cost. Put each run into the cash forecast in the month it is paid for, so the runway figure is accurate.

When model work counts as R&D

The test is whether it was unknown at the outset whether the approach would work, and whether a competent professional in the field could have worked out the answer. Research on architectures, training methods or techniques where that uncertainty was real can support a strong claim.

Fine-tuning a published model on your own data, writing prompts, building an interface over another company's model and integrating a supplier's model into a product are ordinary development, however valuable they are commercially. They do not seek an advance in science or technology, so a claim built on them is likely to fail an HMRC enquiry, and the credit may have to be repaid after it has been spent.

We assess whether your work qualifies before a claim is prepared. Our fee for an R&D claim is a fixed amount set by your qualifying spend.

How we handle R&D claims SEIS and EIS

What we do for AI companies

Computing and data licence costs split between qualifying development and the live product, with evidence for the split
An assessment of whether the model work itself qualifies, before a claim is prepared
The intensity calculation, because high computing costs can take an AI company over the 30% threshold for the intensive route
The PAYE and National Insurance cap modelled before the credit goes into a forecast
Enterprise Management Incentive (EMI) share options for the researchers you want to hire
Cash forecasts built around the timing of training runs

Common questions

Do our cloud computing costs qualify for R&D relief?

Cloud computing and data licence costs have qualified since April 2023, so computing costs behind qualifying development work go into the claim. Running the model behind a live product does not qualify. For a company whose largest cost after payroll is computing, this can make a large difference to the claim, so track the split through the year.

We fine-tune an existing model. Is that R&D?

Usually not. Fine-tuning a published model on your own data, writing prompts and building a product over another company's model are ordinary development, however valuable commercially. A claim needs uncertainty that a competent professional in the field could not readily resolve, such as research on architectures or training methods where it was unknown whether the approach would work.

Are datasets we buy claimable?

Data licence costs for development have qualified since April 2023, so they can go into the claim alongside the payroll and computing costs.

How should we forecast cash with big training runs?

Enter reserved computing capacity, planned training runs and data purchases in the months they are paid, instead of spreading them as a monthly average. Keep the R&D credit out of the forecast until the PAYE cap and the size of your trading loss have been checked, because together they decide how much of it you receive.