AI and robotics companies do some of the clearest qualifying R&D in the UK, and file some of the most scrutinised claims. Genuine development work on models, perception, control systems and physical integration qualifies readily, and is worth between 14.7p and 26.97p per £1 of qualifying spend depending on the scheme. But HMRC examines software and AI claims closely, so the work has to be framed as an advance in the field of science or technology, and evidenced like one.
| Claim element | AI and robotics |
|---|---|
| Typical qualifying activities | Model architectures or training methods designed because existing approaches cannot meet the accuracy, latency or safety constraints; Making models work under scarce labels, heavy class imbalance, or distribution shift between training and deployment; Search, planning and constraint solvers where whether any method meets the problem's scale or guarantees is uncertain; Perception in conditions the published methods were never characterised for; Transfer from simulation to hardware, where a policy that converges in the simulator degrades on the machine; Sub-systems a competent professional cannot readily deduce how to combine to give the intended function |
| Costs that usually qualify | Staff time apportioned to the qualifying work, including the engineers on a robotics build; Data licence and cloud computing costs for training runs, simulation and evaluation environments; Materials used up in prototypes and test builds, as consumables; Agency workers under your direction as externally provided workers, at 65% of payments to an unconnected provider |
| Costs that usually do not | The production cluster serving live customers; The test rig itself, and any other capital expenditure; Data or cloud costs attributable to a qualifying indirect activity; Work bought in from an offshore ML contractor |
| Where claims go wrong | Framing the advance in your company's capability rather than in computer science or software engineering; Naming artificial intelligence or computer vision as the field without saying what moved in it; A dataset licence carrying a right to sell, publish or share the data; Assuming the claim is yours when the customer contemplated the R&D in the contract; A small payroll meeting the PAYE cap with no exemption checked |
| Relief available | Most companies claim the merged R&D expenditure credit; a loss-making SME that meets the R&D intensity condition claims ERIS instead. Current and earlier rates are set out in R&D tax relief rates by year. |
Where does the advance have to sit?
In the field of computer science or software engineering, not in your company’s capability. HMRC’s software guidance places software development inside computer science and information technology, and the Guidelines require an advance in overall knowledge or capability in the field, not a company’s own state of knowledge alone. Applying an existing model or framework to a business problem it has not met before is a step for the business: the field already knew the technique worked.
HMRC’s compliance guidance puts it in a worked example. Software written to analyse market research data is not R&D, because market research is not a qualifying field, unless the work seeks an advance of computer science or software engineering itself. An algorithm, the same example says, does not qualify as such unless it extends overall knowledge or capability of algorithms themselves. Nor is naming a field enough: where the advance is claimed in a sub-area such as artificial intelligence or computer vision, HMRC expects the competent professional to explain how the project is new in, or an appreciable improvement to, that sub-field relative to what is publicly available or readily deducible from it. Since April 2023 the Guidelines also treat a mathematical advance as science in its own right.
What AI development qualifies as R&D?
Development that pushes past what a competent professional could achieve with published knowledge and established techniques. In practice that covers more than model work:
- Novel model architectures or training methods, developed because existing approaches cannot meet the accuracy, latency or safety constraints of the problem.
- Making models work under hostile data regimes: scarce labels, heavy class imbalance, distribution shift between training and deployment.
- AI beyond machine learning: planning and scheduling systems, search and constraint solvers, and hand-built vision pipelines, where the uncertainty is whether any method can meet the problem’s scale or guarantees — no learned model required.
- Deployment engineering that is genuinely unresolved, such as running inference within hard power or memory budgets on edge hardware, where the work shades into the device development on our semiconductors and photonics page.
Fine-tuning an available model with standard tooling to ship a product feature is commercial development, however valuable: new to your company, not to the field. HMRC’s own examples cut the same way. Adapting a natural language processing technique to a website by readily available methods published in open-source communities is routine use of existing knowledge, even for a company that has never used it. Independently reproducing a data-processing method a competitor holds as a trade secret does qualify, because the original is neither publicly available nor readily deducible. Where each of those lines falls — architectures and training methods, difficult data regimes, deployment constraints — is worked through with the evidence each needs in when machine learning development qualifies as R&D.
What robotics development qualifies as R&D?
Robotics claims usually rest on integration and environment uncertainty. Components that each work to specification can still produce a system whose behaviour cannot be predicted: perception, planning and actuation interact, and the real world refuses to match the simulator. Qualifying work typically includes control systems for unstructured environments, closing the gap between simulation and physical performance, and prototype iterations where each build tests a hypothesis the last one raised.
Saying where the uncertainty sits is what turns that into a claim. Perception in conditions the published methods were never characterised for. Real-time control of a machine whose dynamics nobody can predict from the datasheets. Transfer from simulation to hardware, where a policy that converges in the simulator degrades on the machine for reasons still to be isolated. Safety cases that have to hold across operating conditions the established verification methods were never built to cover.
Integration meets a standard objection — every component came from a catalogue — and the Guidelines answer it: the uncertainty is whether a competent professional could readily deduce how those sub-systems combine to give the intended function, not whether the parts were new. Assembly to an established pattern is not R&D. The iteration record is the evidence, covered in robotics prototyping and technological uncertainty.
| Activity | Usually inside the claim | Usually outside the claim |
|---|---|---|
| Models and training | Architectures, loss functions or training regimes designed because published approaches cannot meet the accuracy, latency or safety constraints | Fine-tuning a documented model with standard tooling; running larger versions of an experiment whose outcome the field can already predict |
| Perception, control and integration | Behaviour the field’s published methods were not characterised for; sub-systems a competent professional cannot readily deduce how to combine | Tuning inside an operating window the field understands; assembly to an established pattern |
| Prototypes | Design, construction and testing while the uncertainty is live | Work after the test findings are reflected in the design and further testing is satisfactorily completed |
How closely does HMRC look at AI and robotics claims?
Closely, and it would be wrong to pretend otherwise. HMRC checked around one in six R&D claims in 2023-24, its latest published figure, and software-based claims have attracted particular attention through the compliance push of recent years. The recurring failure is framing: claims written around the product and the market rather than the technological uncertainty.
HMRC caseworkers are advised by the department’s own computer specialists, who do not hold themselves out as competent professionals on any particular project — the caseworker still makes the decision. Where a specialist and your competent professional disagree, HMRC’s guidance says the competent professional may be asked to clarify, by reference to public-domain information, what prevented the advance and the uncertainties from being readily deduced. That is the baseline argument again, made after the fact. Our page on HMRC R&D enquiries explains what a check involves, including claims filed by other advisers.
What does the Get Onbord decision establish?
Less than it is usually made to carry. In 2024 the First-tier Tribunal allowed the appeal of a company that had built an automated know-your-client verification process using AI and machine learning. HMRC objected on one ground: that the work advanced nothing in the field. The tribunal held that each component did not have to be new in itself, and that what mattered was whether the technology the company set out to build was already publicly available or readily deducible. It accepted a former director with no formal qualification in the field as the competent professional. But the decision binds only the parties, and HMRC had conceded everything except the advance, so costs, records and apportionment were never tested. Our register entry gives the paragraph references and the part people over-read.
Which costs go into an AI or robotics claim?
Staff costs apportioned to the qualifying work: research engineers, the people running and evaluating experiments, and the mechanical and electronics engineers on a robotics build. Software licence fees for the R&D, at a reasonable share where a licence is only partly used for it.
Data and compute. Data licence and cloud computing costs are qualifying categories for accounting periods beginning on or after 1 April 2023, which covers every claim under the current schemes. Cloud computing services are described in the statute as including remote data storage, hardware facilities, operating systems and software platforms; a data licence is a licence to access and use a collection of digital data. Two restrictions apply to these categories alone. The cost drops out where the licence gives you a right to sell the data, or to publish or share it beyond what the R&D reasonably needs — read the dataset agreement before signing. And it drops out so far as it is attributable to a qualifying indirect activity. Which software and cloud costs qualify works through the apportionment.
Data acquisition and labelling have no category of their own, so each cost follows the route by which you obtained it. A dataset bought in is a data licence cost; labelling by employees is a staff cost; labelling by agency workers under your direction enters as externally provided workers, at 65% of payments to an unconnected provider; and labelling bought in as a finished service is a contractor payment, but only where the work bought in is itself part of the R&D contracted out. The activity still has to sit inside the R&D project: labelling by known methods to build a general dataset is not, by itself, work to resolve an uncertainty.
Robotics hardware. Materials used up in prototypes and test builds qualify as consumables; the rig itself does not. Capital spending on the R&D can attract R&D allowances instead. Where the first article was always going to be delivered to a customer, the boundary tightens across every cost category: can I claim R&D tax relief on a prototype that is later sold?
Contracted-out model development. Where one company pays another to carry out R&D, only one of them claims it. The customer claims where it intended or contemplated, when the contract was made, that R&D of that sort would be done; where it did not, the contractor claims in its own right. AI companies meet this from both sides: commissioning model development, and building models to a client’s specification. Read contracted-out R&D before assuming the claim is yours.
Location. The overseas restriction is written for externally provided workers and contractors: the work must be done in the UK, or the workers within UK PAYE and Class 1 NIC, with one narrow exception, for conditions the R&D needs that are absent from the UK and would be wholly unreasonable to replicate here — never for cost or the availability of workers. It does not reach software, data licence or cloud computing costs, so compute bought from a provider whose region sits outside the UK is not caught on that ground. An offshore ML contractor is. Overseas R&D costs sets out the exception.
The PAYE cap can bite small payrolls. Payable credits are capped at £20,000 plus 300% of relevant PAYE and NIC. An exemption applies where the company is taking steps towards creating relevant intellectual property, creating it, or doing a significant amount of management work on relevant IP it holds — the activity wholly or mainly undertaken by its own employees — and where its connected-party contractor and externally provided worker spend does not exceed 15% of its qualifying expenditure. Many genuine AI and robotics developers meet it, including teams whose IP is still being worked towards rather than already created, but it needs checking rather than assuming. The PAYE cap page has the conditions.
Grants no longer reduce the claim. The old subsidised-expenditure restriction is abolished for current-scheme periods, so an Innovate UK funded project claims in full: grant funding and R&D tax relief has the detail.
Which scheme applies, and what is it worth?
Pre-revenue AI and robotics companies are exactly who Enhanced R&D Intensive Support (ERIS) was designed for. A loss-making SME whose relevant R&D expenditure is at least 30% of its total relevant expenditure receives up to 26.97p per £1: a £26,970 payable credit on £100,000 of qualifying spend. That denominator is broadly the trading costs in its accounts for the period, not just the R&D ones, so deep in the development phase the test is often passed comfortably. If intensity falls away later, one year of grace may still be available, where the company met the condition in its most recent prior 12-month accounting period and obtained relief for it.
Profitable companies claim under the merged scheme: a 20% credit that nets to £15,000 on £100,000 of spend at the 25% corporation tax rate, £14,700 at the 26.5% marginal rate where augmented profits fall between £50,000 and £250,000, or £16,200 where the 19% rate applies.
The gap matters at scale. On £400,000 of qualifying spend in a period beginning on or after 1 April 2024, a loss-making company with intensity above 30% takes £107,880 in cash under ERIS; the same spend under the merged scheme gives an £80,000 credit, worth £64,800 net to a loss-maker. Both sit before the PAYE cap. Every rate, and the date test that fixes which applies, is on rates by year; the ERIS intensity calculator works the 30% ratio including connected companies, and the claim value calculator estimates both.
What should the claim record?
The competent professional’s account, written while the work was still uncertain, with the baseline set out clearly enough that a reader can see what the field could already do. Experiment trackers, training logs and evaluations against that baseline supply the AI half; build logs, test reports and the failures that changed the next iteration supply the robotics half. What both usually lack is the connective tissue — a note of what question each run or build was asking, and what the answer changed. HMRC has published what it expects of the documentation in its Guidelines for Compliance; the person is covered in who counts as a competent professional. Two compliance deadlines sit around all of this. The Additional Information Form has been mandatory for claims made on or after 1 August 2023, in practice 8 August 2023, and a first-time claimant, or one that has not claimed in the three years ending with the notification deadline, must notify HMRC within six months of the end of the period of account or the claim is invalid.
What our AI and robotics clients say
“Working with LimestoneGrey has been a genuinely positive experience. Their communication is clear and proactive and they are always on hand to answer queries and provide reassurance throughout the process. We have complete trust in their expertise and feel confident that our claims are being handled professionally.”
Dr Varghese, Laennec AI Limited
“We have been very happy with the service provided by LimestoneGrey. In particular, their attention to detail has been exceptional, making the entire process smooth and reassuring for us.”
Marcus, Neurabotics Limited
Talk it through with a chartered adviser
LimestoneGrey is a firm of Chartered Tax Advisers and Chartered Accountants, regulated by ICAEW, specialising in R&D tax relief. We focus on R&D-intensive companies, with advanced AI and robotics among our core sectors. Every claim is prepared by our specialist team and signed off by a chartered adviser, enquiry support is included as standard, and the fee is agreed before any work starts.
If you want a straight answer on whether your development qualifies, how the intensity test falls on your numbers, or how to evidence a claim HMRC may check, get in touch.
Sources
- Guidelines on the meaning of R&D for tax purposes — paragraphs 6 and 8 on the advance lying in the field rather than the company, 15B on mathematical advances treated as science from April 2023, 20 on readily deducible knowledge, 29 and 30 on system uncertainty, 31 and 32 on qualifying indirect activities, and 39 on where prototype work ends.
- CIRD81960: the Guidelines applied to software — software development as part of computer science and information technology; why naming a sub-area such as artificial intelligence or computer vision is not specific enough; HMRC’s computer specialists, and the request to clarify by reference to the public domain.
- GfC3: how to identify qualifying R&D activities (part 4) — example 4.4 on reproducing a trade secret independently, 4.5 on adapting a natural language processing technique by published open-source methods, and 4.6 on market research software and on algorithms.
- CTA 2009 s1125 and s1126ZA — data licences and cloud computing services, and the restrictions where a right to sell, publish or share the data is obtained or the cost is attributable to a qualifying indirect activity; both take effect for accounting periods beginning on or after 1 April 2023.
- CTA 2009 s1138A — the overseas restriction, headed “externally provided workers and contractors”, and an exception turning on conditions present outside the UK but expressly not on cost or the availability of workers.
- Check what R&D costs you can claim — the cost categories, the 65% rule for unconnected staff providers and contractors, and the exclusion of capital expenditure.
- Get Onbord Ltd (in liquidation) v HMRC — [2024] UKFTT 617 (TC): the competent professional without formal qualifications, and components that need not each be new.
- HMRC’s approach to R&D tax reliefs 2023 to 2024 — compliance checks covering 17% of claims in 2023 to 2024, the source of the one-in-six figure.
- CTA 2009 s1112E — the PAYE cap exemption: condition A on relevant intellectual property created or managed wholly or mainly by the company’s own employees, and condition B’s 15% limit on connected-party contractor and externally provided worker spend.
- R&D tax relief: the merged scheme and enhanced R&D intensive support — the 20% credit and the ERIS conditions.