← All insights
The Novel Shift

Is AI transforming HMRC's R&D claim processes?

Should R&D advisers use AI when preparing claims, with traditionalists and modernists often clashing over how, or even if, it should be relied up

Belle Chippington
·
November 13, 2025
·
Last updated
·
5 min read
Is AI transforming HMRC's R&D claim processes?

If you’re familiar with the R&D tax credit industry, you’ve likely seen the ongoing debates over whether advisers should use AI when preparing claims, with traditionalists and modernists often clashing over how, or even if, it should be relied upon. Until recently, however, these discussions focused almost entirely on advisers’ use of AI. What hadn’t been explored was HMRC’s potential use of AI to assess R&D claims.

That changed following a two-year back-and-forth between Tom Elsbury, R&D tax expert and co-founder of Novel, and HMRC, during which Tom pursued a Freedom of Information (FOI) request to determine whether HMRC has used AI in its R&D compliance and enquiry activities.

After a series of appeals, extensive correspondence, and even representing himself at the First-tier Tribunal, the court ultimately ruled in favour of Tom, requiring HMRC to disclose whether it has used AI in this context, marking a significant milestone in the pursuit of transparency around HMRC’s use of AI in R&D enquiries.

What was the verdict?

You might be wondering, after such a long time, whether HMRC’s response would finally provide insights on how AI is being used, offer justifications, or give assurances about client data protection, but it didn’t.

In their official response:

“We can confirm we hold the information requested. Language Model (LLM) or generative AI was not used by the R&D Tax Credits Compliance Team as part of work on R&D tax relief claims, and this technology was not approved for use in generating taxpayer letters. In addition, at the time of the request, there were no planned expansions or changes to the use of AI or LLMs in R&D compliance.”

Call me sceptical, but it’s surprising that such a simple response took years to arrive. This information could have been provided much earlier, saving everyone considerable time and effort.

It’s also hard to ignore the evolution in HMRC’s compliance check responses. What were once jumbled, jargon-heavy responses have become streamlined, ‘Americanised’ responses. This shift suggests that AI may have played a role, whether officially sanctioned or not. Even a few rogue agents using AI independently could influence how R&D claims are evaluated and assessed.

Is AI helpful or harmful?

It would be remiss of me, writing on behalf of Novel, to criticise HMRC solely for using AI. AI can be an incredibly useful tool for a variety of tasks. I’ll admit I use it myself to fix typos, correct grammar, and eliminate repetition. After all, I encourage advisers to trial Novel Intelligence, our AI tool designed to generate robust, compliant technical narratives. As a time-saving measure, especially within a stretched government sector, it’s understandable that AI would be used to streamline workflows.

But while I welcome AI for editing and tidying, we need to understand how, or if, it’s being used to evaluate the substance of R&D claims. My concern is that if HMRC inspectors rely on AI to make eligibility decisions, it could produce convincing-looking but entirely unverified rejections. A simple prompt like “Can you reject this R&D claim with a list of reasons why?” could generate a response which is difficult to contest. If this "computer says no" approach is used to shape outcomes, it could have a substantial impact on claimants, particularly where penalties are involved.

Another elephant in the room is client data privacy. Feeding sensitive client information into a public LLM, such as GPT-5, raises major confidentiality concerns. If these public models are trained on confidential submissions, there is a real risk that proprietary project details, financial information, or other sensitive data could be exposed. While we use AI tools within Novel, our LLM is never trained on client data, ensuring that all information remains confidential. Without an officially sanctioned LLM at HMRC, how can advisers reassure clients that their R&D claim remains secure when AI tools are involved?

What are the potential implications of HMRC using AI?

To understand why HMRC’s decision-making approach matters, it’s helpful to look at the 2025 R&D tax statistics. Following the R&D scheme’s reforms, R&D claims are down sharply, with a 26% overall reduction, including a more pronounced 31% drop among SMEs in just one year. The number of first-time claimants continues to fall as well. On the surface, this may look like a win for HMRC, given the sector’s higher rates of error and fraud.

However, my concern is that the decline in claims may not only reflect the removal of ineligible applications, but could also indicate that some genuine businesses are being discouraged from applying. With heightened scrutiny, tougher compliance requirements, and the potential of client data leaks, there is a real risk that eligible claimants may face unintended barriers that prevent them from submitting eligible R&D claims.

If the R&D scheme is to truly support innovation, HMRC must balance fraud prevention with fair treatment, ensuring AI is used responsibly, client data remains secure, and genuine claimants aren’t inadvertently penalised.

What’s next?

In next month’s edition, I’ll shift focus to explore how different industries perceive R&D tax relief credits. We’ll examine how recent changes to the scheme and government funding have influenced these perceptions, which sectors are most affected, and whether these shifts could alter how R&D tax relief is applied in practice.

The Novel Shift

Enjoyed this issue? Get the next one on LinkedIn.

Subscribe on LinkedIn
Book a walkthrough

See how Novel fits the way your team already works.

Book a 30-minute walkthrough