Accountants for AI startups

AI: enormous compute bills and a claim worth getting right.

Compute is your second-largest cost and it is claimable. Whether the model work itself qualifies is a harder question, and the answer is not always yes.

Talk to us See pricing

The short answer

We act for UK AI and machine learning startups: putting cloud compute and data licence costs into R&D claims where they belong, assessing honestly whether the model work itself qualifies, and modelling cash against training costs that do not behave like ordinary overheads.

Two things make AI companies different to account for, and they pull in opposite directions.

The first is helpful: cloud computing and data licence costs qualify for R&D relief, and for a company whose largest non-payroll line is compute, that is a significant part of a claim that used to be excluded. The second is not: fine-tuning an existing model, calling an API and building a product around it is, in HMRC's terms, usually a business achievement rather than a technological advance. A great many AI companies are doing exactly that, and being told by somebody that it is a claim.

Where the money actually goes

The shape of an AI company's costs

Compute

Training and inference are a large, lumpy and genuinely claimable cost where they sit behind qualifying work. Production inference for a running product does not qualify, so the split has to be tracked rather than estimated.

Data and licences

Datasets and data licences bought for development have qualified since April 2023 and are frequently left out of a first claim entirely.

Research payroll

The largest line and the one that carries the technical argument. Which people spent what share of their time on genuinely uncertain work is the substance of the claim.

Cash that moves in steps

A training run is not a monthly overhead. Cash modelling has to reflect the actual shape of spend or the runway number is meaningless.

R&D relief in this sector

The honest test for model work

The question is whether it was genuinely unknown at the outset that the approach would function at all, and whether a competent professional in the field could have deduced the answer. Research on architectures, training methods or techniques where that uncertainty was real is a strong claim, and often a very good one.

Fine-tuning a published model on your own data, prompt engineering, building an interface over an API, and integrating a vendor's model into a product are ordinary development. They may be commercially brilliant. They are not an advance in science or technology, and a claim built on them is a claim that fails an enquiry — expensively, because the money will usually have been spent by then.

We will give you a straight answer on which side your work sits before anyone starts writing, which is the point of not charging a percentage of the claim.

How we handle R&D claims SEIS and EIS

A worked example

What leaving compute out costs you

A loss-making UK AI company. Total expenditure £700,000. Research payroll on genuinely uncertain model work is £180,000; the compute and data licences behind that work are £240,000. Qualifying R&D is therefore £420,000.

Intensity: £420,000 ÷ £700,000 = 60%, so the intensive route applies.
Claim including compute: £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.

Leaving the compute out costs this company £64,728. Cloud and data licence costs have qualified since April 2023, and they are still the single most commonly omitted category in a first AI claim — because whoever prepared it was working from a list written before the rule changed.

Illustrative figures, chosen to show how the arithmetic behaves. Your numbers will differ, which is exactly why the calculation gets done before anything is filed.

What we do for you

The job, end to end

Compute and data licence costs split between qualifying development and production, and evidenced
A straight answer on whether the model work itself qualifies, before a claim is built
The intensity calculation, since heavy compute spend often puts an AI company on the more generous route
The PAYE and NIC cap modelled before the credit goes into a forecast
EMI options for researchers you are competing for against much larger budgets
Cash modelled around training runs rather than a smooth monthly average

Questions

Straight answers

Does our cloud compute qualify for R&D relief?

Cloud computing and data licence costs have qualified since April 2023, so compute behind genuinely qualifying development work goes into the claim. Production inference serving a live product does not. For a company whose biggest non-payroll cost is compute this is often the difference between a small claim and a substantial one, so the split is worth tracking through the year.

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

Usually not. Fine-tuning a published model on your own data, prompt engineering and building a product over an API are ordinary development, however valuable commercially. A claim needs uncertainty that a competent professional in the field could not readily resolve — research on architectures or training methods where it was genuinely unknown whether the approach would work.

Are datasets we buy claimable?

Data licence costs for development purposes have qualified since April 2023. They are one of the most commonly omitted categories in a first AI claim, usually because the person preparing it is working from a pre-2023 list of qualifying costs.

How should we forecast cash with big training runs?

Not on a monthly average, which is how most templates do it and why so many AI runway numbers are wrong. Model the actual commitments — reserved capacity, planned runs, data purchases — as the steps they are, and keep the R&D credit out of the forecast until the PAYE cap has been checked, because that is what decides how much of it you actually receive.