AI in R&D: when AI projects may qualify for R&D tax relief

That does not mean every AI project qualifies for R&D tax relief. HMRC looks at whether the project seeks an advance in science or technology, and whether the work directly contributes to resolving scientific or technological uncertainty. A project can use advanced technology and still fall outside the rules if the work is routine, commercially focused or readily solvable by a competent professional.

For AI startups, software teams and innovation-led businesses, the opportunity can be real. The key is knowing where the qualifying R&D sits, what costs can be reviewed and what evidence you need before you make a claim.

Can AI projects qualify for R&D tax relief?

AI projects may qualify for R&D tax relief where they seek an advance in science or technology and involve genuine scientific or technological uncertainty.

That could include developing a new machine learning model, improving model performance where standard methods are not enough, solving complex system uncertainty, or creating AI systems that need to work reliably in difficult real-world environments.

Simply using AI tools will not usually be enough. Integrating an existing AI API, using ChatGPT for routine tasks, or building a commercially new product with known methods is less likely to qualify on its own. The focus should be on the technical problem, the uncertainty faced and the work done to resolve it.

If you are unsure whether your AI work meets HMRC’s definition, start with our guide to what counts as R&D, or use our free eligibility test to get an early view of which R&D incentives may be relevant.

Why AI creates real R&D questions

AI projects often sit in a difficult middle ground.

On one hand, AI can involve genuine technical uncertainty. Teams may need to solve problems around model accuracy, explainability, data quality, bias, performance, security, scalability or integration with complex systems. The outcome may not be readily deducible from public knowledge or standard practice.

On the other hand, AI is increasingly accessible. Many teams now use pre-trained models, third-party APIs, open-source libraries and commercial AI tools. These can be powerful, but using them does not automatically make the work qualifying R&D.

HMRC’s software guidance recognises artificial intelligence as one of the established branches of software and information technology that continues to evolve. That is useful context, but it does not create a separate R&D test for AI. The same core definition still applies.

The question is not “Are we using AI?”. The better question is: “What advance were we seeking, what uncertainty stopped us getting there, and why could that uncertainty not be readily resolved?”.

The R&D definition that matters for AI

For tax purposes, R&D is narrower than many people expect. It is not the same as product development, commercial innovation or general experimentation.

The project must seek an advance in science or technology

HMRC’s guidance explains that an advance in science or technology means an advance in overall knowledge or capability in a field of science or technology. It is not enough for the work to be new to your company. It must go beyond what is publicly available or readily deducible by a competent professional in the field.

For AI, that means a new feature, user experience or workflow improvement will not automatically qualify. A project may be commercially valuable and still not meet the R&D definition.

A stronger case might involve building an AI system that performs in a way existing methods cannot reliably achieve, adapting knowledge from another technological field in a non-obvious way, or making an appreciable technological improvement to an existing process, product or service.

The uncertainty must be scientific or technological

Technical uncertainty is different from commercial uncertainty.

A commercial uncertainty might be whether customers will adopt the product, whether investors will fund it, or whether a market will pay for it. Those questions matter to the business, but they are not enough for R&D tax relief.

A technological uncertainty might be whether a model can reach the required accuracy under real-world data conditions, whether a system architecture can process sensitive data at the required scale, or whether an AI-enabled decision system can remain reliable, auditable and secure under operational constraints.

HMRC guidance also recognises system uncertainty. This can arise from the complexity of a system and the way its components interact, even where the individual components are already known. That can be relevant for AI systems built from models, data pipelines, cloud infrastructure, APIs and operational workflows.

A competent professional must not be able to readily solve it

The uncertainty needs to be one that a competent professional could not readily resolve using available knowledge.

That does not mean the project must be groundbreaking in a public-facing way. It also does not mean the work has to succeed. HMRC guidance confirms that R&D can still take place even if the advance sought is not achieved or only partly achieved.

For AI claims, this is where evidence matters. You need to show why standard tools, libraries, models or techniques were not enough, and what technical work your team carried out to move beyond them.

This is where the view of a competent professional matters. They help explain why the uncertainty was genuinely scientific or technological, and why the answer was not readily available. The competent professional must be able to demonstrate relevant expertise and a successful track record in AI; this is especially important where a company is claiming an advance in AI but operates primarily in another sector, such as manufacturing or life sciences.

AI work that may qualify for R&D tax relief

AI work may qualify where it forms part of a project seeking an advance in science or technology. The examples below are not automatic qualifying categories. They are areas worth reviewing where the technical uncertainty is clear.

Developing new AI models or algorithms

A project may have a stronger R&D case where the team is developing a model, algorithm or technical method that goes beyond routine implementation.

This could include work to improve performance, reduce error rates, handle unusual data conditions, increase efficiency, or create a new approach where existing methods do not achieve the required outcome.

The claim should explain what was technically difficult, what standard approaches were considered, and why those approaches were not enough.

Improving machine learning performance where standard methods are not enough

Machine learning projects often involve iteration, but iteration alone is not the test.

A stronger case may exist where the uncertainty sits in achieving a required level of accuracy, reliability, sensitivity, specificity or robustness under difficult conditions. For example, the team may be dealing with noisy data, limited training data, unusual edge cases, model drift or conflicting performance requirements.

The R&D narrative should not simply say “we trained a model”. It should explain the uncertainty, the experiments carried out, the failed approaches and the technical learning gained.

Building AI systems for complex real-world environments

AI can behave differently in controlled development environments compared with real operational settings.

A project may involve qualifying R&D where the team is trying to make an AI system work reliably in conditions that are technically complex, variable or constrained. Examples could include robotics, autonomous systems, manufacturing inspection, clinical workflows, financial risk engines, energy forecasting or logistics optimisation.

The important point is the system challenge. If the uncertainty comes from how multiple technical components interact, the claim should explain that system uncertainty clearly.

Developing explainable, auditable or secure AI

AI systems in regulated or high-stakes settings often need to be explainable, auditable, secure and reproducible.

Work in these areas may be relevant where the team is solving a technological uncertainty, rather than only preparing documentation or meeting a compliance checklist. For example, a company may need to create a technical method to trace model outputs, reduce unacceptable bias, protect sensitive data, or monitor performance degradation in a live environment.

The R&D case will be stronger where the work required technical development, testing and iteration, not just policy decisions.

Creating data pipelines that solve technical uncertainty

Data work is common in AI, but not all data preparation is R&D.

Routine cleaning, labelling or formatting may not qualify on its own. The position may be different where the data pipeline is central to resolving technological uncertainty. For example, the team may need to build a method for handling inconsistent, incomplete or high-volume data in a way that directly supports the AI system’s technical advance.

The claim should connect the data work to the R&D project. It should explain why the work was needed to resolve the uncertainty, rather than treating all data activity as qualifying by default.

Using AI to support wider R&D

A company does not need to be an AI company for AI to be relevant to R&D.

AI may support R&D in materials science, drug discovery, engineering, clean energy, manufacturing, climate modelling, health technology or software development. In these cases, the AI may be a tool used to resolve uncertainty in a wider scientific or technological project.

The same rules apply. The question is still whether the project seeks an advance in science or technology, and whether the AI-related work directly contributes to resolving the relevant uncertainty.

AI work that is less likely to qualify

It is important to be clear about what is less likely to meet the test. This helps set realistic expectations and supports a stronger claim.

Using AI tools for routine tasks

Using generative AI to draft content, summarise documents, produce internal notes, write basic code or support everyday administration is unlikely to qualify on its own.

Those uses may improve productivity, but they are usually not seeking an advance in science or technology.

Integrating an existing AI API without technical uncertainty

Many businesses use existing AI APIs to add functionality to a product. That can be commercially useful, but routine integration is less likely to qualify.

A stronger case may exist where the integration creates genuine technological uncertainty around architecture, performance, scale, security, reliability or system behaviour. The claim should focus on that technical uncertainty, not on the fact that an AI API was used.

Building a commercially new AI product using known methods

A product can be new in the market and still not qualify.

HMRC guidance is clear that a product, process or service does not become an advance in science or technology simply because science or technology is used in creating it. Routine analysis, copying or adaptation of existing products, processes or services is not enough.

For AI businesses, this means the claim should not rely on product novelty. It should explain the technological advance and uncertainty.

Prompt engineering without deeper technical development

Prompt engineering can be skilled and commercially valuable, especially when building AI-assisted workflows. But prompt design alone will not usually be enough unless it sits within a wider qualifying R&D project.

If the team is only testing prompts to get better outputs from an existing model, the work may be closer to routine optimisation. If the prompt work is part of a broader technical project involving system uncertainty, model behaviour, evaluation methods or reliability constraints, it may be worth reviewing in context.

AI examples by sector

AI R&D can appear in different sectors. These examples are intended to help teams think about where the uncertainty may sit.

Software and SaaS

In software and SaaS, AI R&D may involve model architecture, performance optimisation, secure deployment, data processing, monitoring, explainability or model behaviour at scale.

A routine chatbot integration may not qualify. A project to build an AI-enabled system that remains reliable across high-volume, sensitive or technically complex data environments may be more relevant.

Health tech and life sciences

In health tech, AI projects may involve diagnostic support, imaging analysis, triage systems, patient-risk modelling, clinical workflow integration or drug discovery support.

The R&D question is not whether the technology is used in healthcare. It is whether the team is resolving scientific or technological uncertainty. Evidence around validation, performance, bias, data quality and clinical constraints may help explain the technical context.

Manufacturing and robotics

In manufacturing and robotics, AI may support defect detection, predictive maintenance, machine vision, autonomous control, production optimisation or robotic perception.

The strongest cases often involve real-world variability: changing materials, environmental conditions, sensor noise, machine interaction or reliability under production constraints.

Clean energy and climate tech

AI may support energy demand forecasting, grid optimisation, battery performance modelling, renewable asset management, climate modelling or materials discovery.

Where the work involves solving technical uncertainty around modelling accuracy, system constraints or performance under changing conditions, it may be relevant to review.

Financial services and professional services

AI in financial services and professional services may involve fraud detection, risk modelling, document intelligence, decision support or secure workflow automation.

Commercial use alone is not enough. The claim needs to show the technological uncertainty, particularly around model performance, auditability, security, explainability or integration with complex systems.

What AI costs can you review for R&D tax relief?

Once the qualifying activity is identified, the next step is cost review. HMRC guidance says companies need to check whether the R&D activity qualifies, check which costs are allowed, and only include costs from the start to the end of the project. R&D starts when work begins to resolve the scientific or technological uncertainty and ends when that uncertainty is resolved, or when work to resolve it stops.

For AI projects, relevant costs to review may include:

• Staff costs for developers, data scientists, engineers, technical leads and other employees working on qualifying R&D
• Employer National Insurance and pension contributions linked to qualifying staff time
• Externally provided workers
• Contractor costs, subject to the relevant rules
• Software licence fees used for R&D
• Data licence costs
• Cloud computing and GPU costs
• Consumables, where relevant
• Clinical trial or testing costs, where relevant to the project

The cost review should be careful. It is rarely appropriate to treat all AI product development spend as qualifying. A developer might spend part of their time resolving technical uncertainty and part of their time on routine feature development. A cloud environment might be used partly for model training and partly for production hosting.

The claim should explain the basis for the apportionment and link the costs back to qualifying R&D activity.

Cloud, compute and data costs in AI R&D

Cloud and data costs are particularly important for AI businesses.

For accounting periods beginning on or after 1 April 2023, HMRC guidance says qualifying expenditure includes data licence costs and cloud computing costs. Cloud computing includes data storage, hardware facilities, operating systems and software platforms.

HMRC also states that data licences and cloud computing services can be qualifying expenditure under the new RDEC and ERIS when they are employed in activities that directly contribute to resolving scientific or technological uncertainty. Costs linked only to qualifying indirect activities do not qualify.

For AI claims, this distinction matters.

Model training, experimentation, testing environments and performance benchmarking may be relevant where they support qualifying R&D. Production hosting, customer delivery, routine storage or general business infrastructure should usually be reviewed separately.

A strong claim should separate:

• Model training from live service delivery
• Experimentation from routine operation
• R&D data from commercial-use data
• Technical testing from ordinary hosting
• One-off R&D infrastructure from ongoing platform costs

Good cost mapping makes the claim easier to explain and easier to defend.

How the merged R&D scheme affects AI companies

For accounting periods beginning on or after 1 April 2024, the merged R&D expenditure credit scheme and enhanced R&D intensive support replaced the old RDEC and SME schemes. HMRC guidance says the merged RDEC rate is 20%, and ERIS is available for loss-making R&D intensive SMEs that meet the relevant intensity condition.

For AI companies, this matters for several reasons:

First, the scheme you claim under depends on your accounting period and company position. Secondly, the rules around contracted-out R&D and overseas activity may be important if you use international developers, data teams or specialist AI contractors. HMRC’s manual confirms that overseas restrictions apply to EPW (externally provided workers) payments and contractor payments, with specific exceptions where R&D necessarily has to be carried on abroad.

Thirdly, process requirements have become more formal. HMRC requires an Additional Information Form to support new R&D claims, and it must be submitted before or on the same day as the Company Tax Return. If the Company Tax Return is submitted first, HMRC says the claim will be rejected.

Some companies also need to submit a Claim Notification Form before claiming. This can apply to first-time claimants and companies whose last claim was made more than three years before the end of the claim notification period.

The practical takeaway is simple: do not leave the R&D review until the claim deadline is close. AI projects can involve complex activity, fast-moving evidence and mixed costs. The earlier you map the position, the better.

Evidence HMRC may expect for AI R&D claims

The strongest AI R&D claims are usually built from evidence created during the project, not reconstructed at year end.

HMRC’s guidance on qualifying activities says the R&D must be part of a project conducted to a method or plan, and that the project should seek to resolve specific uncertainties to achieve an advance in science or technology. It is not enough to discover an advance during other activities.

For AI projects, useful evidence can include technical, financial and narrative records.

Technical evidence

Keep records that show what the team was trying to achieve and how the technical work developed. That might include:

• Architecture documents
• Model design notes
• Experiment logs
• Training run records
• Test results
• Performance benchmarks
• Error analysis
• Data-quality assessments
• Security and reliability testing
• Version histories
• Technical decision logs
• Records of failed approaches

Git commits and tickets can help, but they rarely explain the full R&D story on their own. A short technical note written during the project can be more useful than a long reconstruction months later.

Financial evidence

The cost evidence should connect the numbers to the R&D activity. Useful records include:

• Payroll data
• Time apportionments
• Contractor invoices
• Statements of work
• Cloud billing exports
• Data licence invoices
• Software subscriptions
• Project cost codes
• Ledger reconciliations

The finance record and the technical record should tell the same story. If the claim says the R&D ended in March, but the cost schedule includes large production hosting costs from June, that mismatch may need explaining.

Narrative evidence

The R&D narrative should answer the core questions clearly:

• What advance in science or technology were you seeking?
• What scientific or technological uncertainty did you face?
• Why could a competent professional not readily resolve it?
• What did your team try?
• What worked, what failed and what changed as a result?
• Which costs link to that work?

This is especially important for AI, where generic language can weaken the claim. “We developed an AI platform” is not enough. “We tested three model architectures because standard approaches could not maintain accuracy across incomplete data at the required processing speed” is much clearer.

Common mistakes when claiming R&D tax relief for AI

Assuming all AI work qualifies

AI is not a shortcut to eligibility. The claim still needs to meet HMRC’s definition of R&D.

Focusing on product novelty instead of technical uncertainty

A new AI product is not automatically qualifying. The claim should focus on the advance in science or technology, and the uncertainty involved in seeking it.

Treating all software development as R&D

AI projects often include a mix of qualifying and non-qualifying work. Routine frontend development, admin dashboards, standard integrations and commercial deployment may need to be separated from the R&D activity.

Not separating training from production

Cloud and compute costs can be significant in AI projects. The claim should distinguish model training, experimentation and testing from production hosting and customer delivery.

Relying on generic AI language

Phrases like “machine learning”, “AI-powered” or “advanced algorithm” do not explain the uncertainty. HMRC needs to understand what was technically difficult and what work was done to resolve it.

Missing overseas contractor restrictions

Many AI teams use international contractors or development partners. For accounting periods beginning on or after 1 April 2024, overseas contractor and EPW costs need careful review under the reformed rules. The restrictions do not mean every overseas cost is automatically excluded, but the exception needs evidence.

Leaving evidence until year end

AI projects move quickly. If the evidence is not captured as the work happens, the technical story can become hard to reconstruct later.

AI, grants and Patent Box: where the wider funding picture fits

AI projects should not always be viewed only through the lens of R&D tax relief.

Depending on your stage, risk profile and IP position, other funding routes may be relevant.

Grants for early AI development

Grant funding may be worth reviewing where the AI project has a clear innovation case, technical risk and wider impact. This can be especially relevant for feasibility work, collaborative R&D, sector-specific challenges or projects linked to health, clean energy, manufacturing, transport, defence, space or other high-impact areas.

A grant application usually needs a clear problem, a credible plan, realistic costs, a capable team and measurable outcomes. It is not the same as an R&D tax claim, but the evidence can overlap.

R&D tax relief for qualifying AI expenditure

R&D tax relief can support qualifying expenditure once the company has incurred relevant costs. It is often most useful when the business is actively developing, testing and iterating the technology.

The important point is timing. If the team waits until after the year end, the technical evidence may be weaker and process requirements may be easier to miss.

Patent Box for profitable AI-enabled inventions

Patent Box may become relevant later, if the company has qualifying patented inventions and taxable profits linked to them. The Patent Box allows companies to apply a 10% Corporation Tax rate to profits attributable to qualifying patented inventions, subject to the rules.

Not every AI product will be patentable, and Patent Box is not automatic. But where patent protection is part of the strategy, it is worth thinking early about IP ownership, development activity, income streams and cost tracking.

How to prepare a stronger AI R&D claim

A well-supported AI R&D claim starts with the right filing checks. Before preparing the full claim, check whether your company needs to submit a Claim Notification Form. This is an online form that tells HMRC you intend to claim R&D tax relief. It is not the claim itself, but where the form is required, it must be submitted within the claim notification period before a valid claim can be made.

First, check whether a Claim Notification Form is required. This can apply to first-time claimants and some companies that have not claimed recently. HMRC guidance says the claim notification period usually ends six months after the end of the period of account.

Second, define the project. Be clear about what technical goal the team was working towards and where the R&D boundaries sit.

Third, identify the advance sought. Explain the wider science or technology field, not just the commercial product.

Fourth, describe the uncertainty. Show why standard methods, tools or knowledge were not enough.

Fifth, capture the work. Keep records of experiments, design decisions, failed approaches, tests and outcomes.

Sixth, map the costs. Link staff, contractors, cloud, software and data costs to the qualifying activity.

Seventh, check the scheme rules. Review the accounting period, merged scheme or ERIS position, overseas costs, contracted-out work and contracted-out R&D rules.

Eighth, prepare the Additional Information Form carefully. This is separate from the Claim Notification Form. The technical narrative, cost categories and project details should be consistent with the claim, the CT600 and the company’s records. The Additional Information Form must be submitted before or on the same day as the Company Tax Return.

The aim is not to make the claim sound more impressive. It is to make it clearer, more accurate and better supported.

How we help

We help you understand whether your AI work may meet HMRC’s R&D definition, then build the claim with clear technical and financial evidence.

Our team brings together STEM, sector and tax expertise to review AI projects, identify qualifying activity, map costs and prepare well-documented submissions across R&D tax credits, grants and Patent Box. The aim is straightforward: clear advice, realistic expectations and a funding position that can stand up to scrutiny.

If you are developing AI, using AI to support wider R&D, or reviewing a previous claim, book a funding assessmnet with our R&D funding advisors. We’ll help you understand what support may be available, what evidence you need and the right next step for your business.