The evidence problem in R&D tax relief

Where the time goes

A software R&D claim depends on answering a few questions well: what the team knew at the start, what they couldn't work out, what they tried, and why a competent professional couldn't simply have solved it. The facts usually exist. The problem is getting at them.

The record isn't written for tax. Engineers record work in tickets, comments, pull requests and specifications. They write "sync job times out", not "uncertainty about whether the approach was feasible". Translating one into the other is skilled, slow work.
Interviews depend on memory. Where technical information lives only in the team's heads, advisers draft reports from engineer interviews. Those happen months after the work, with engineers who have moved on to other projects, or other employers.
The Additional Information Form moves the work earlier. Since August 2023, every claim must be supported by an AIF, or the claim is invalid. For one to three projects, all must be described; for more, enough projects to cover at least half the qualifying spend. Each needs the baseline knowledge, the advance sought and the uncertainties.
Enquiries reopen everything. When HMRC opens a check, the adviser has to rebuild the evidence trail behind the narrative, often for a period long past. Commentators note that HMRC typically asks for contracts, project notes, technical diagrams, test logs and timesheets, and expects records from the time, not just a narrative written afterwards.

HMRC's scrutiny in numbers

Scale. About £8.2 billion of R&D relief was claimed in 2024–25.
Error and fraud. HMRC estimated error and fraud at 9.9% (£759 million) for 2022–23. Its 2025–26 annual report puts 2023–24 at 6.4% (£493 million), 11.1% in the SME scheme, and estimates around 5.3% for the two following years. Improving, but still high enough to keep compliance a priority.
Resources. HMRC says it has more than doubled the number of people working on R&D compliance over three years.
Coverage. HMRC has stated it is now checking over 20% of claims, roughly one in five.
Process. On top of the AIF, companies claiming for the first time, or after a gap, must notify HMRC in advance for accounting periods beginning on or after 1 April 2023.

The cost lands on genuine claimants too. The Chartered Institute of Taxation has warned that valid claims are being rejected and that some claimants give up because challenging HMRC costs too much. BDO has noted that HMRC's error estimates may overstate the problem, because claims abandoned for lack of resources to defend them are counted as error.

HMRC and AI: what's known

The Quantexa contract

In May 2026 HMRC awarded Quantexa, a London-based data and AI company, a £175 million, ten-year contract to modernise its core data, giving it a more connected view of information held across separate systems, to identify tax at risk and tighten control. The contract runs to March 2034, with an option to extend.

It is a data-foundation programme across HMRC, not an R&D claims product. Quantexa's chief executive said automated decisions about taxpayers will still be checked by people, and that decisions must be transparent, auditable and explainable. Some R&D specialists read it as a clear signal that data-driven risk profiling will reach every kind of filing, R&D included.

The Elsbury case and HMRC's answer

R&D tax specialist Thomas Elsbury asked HMRC under Freedom of Information whether its R&D compliance team used generative AI, after noticing American spellings, odd formatting and generic reasoning in rejection letters. HMRC refused to confirm or deny. In August 2025 the First-tier Tribunal ordered HMRC to answer, finding that secrecy was undermining trust and could deter legitimate claims (Elsbury v Information Commissioner [2025] UKFTT 915 (GRC)).

HMRC's response said generative AI was not used by the R&D Tax Credits Compliance Team before December 2023 and was not approved for writing taxpayer letters. Advisers pointed out the answer was silent on informal use by individual caseworkers. The Financial Times later reported, citing sources, that some staff had been disciplined for using AI in correspondence. HMRC has said staff found misusing AI face disciplinary action.

The practical point for advisers: whatever tools HMRC uses, its letters may be formulaic, and its reviewers are increasingly data-led. A claim needs evidence that holds up to a specific question, not a narrative that only reads well.

Why advisers are wary of AI

Tribunals have penalised unchecked AI. In Elden v HMRC [2026] UKFTT 41 (TC), a firm of regulated accountants submitted AI-generated case summaries that were inaccurate and irrelevant. The tribunal found they hadn't been properly verified, and described this as professional incompetence. Earlier cases, including Harber and Zzaman, involved invented or irrelevant authorities.
Professional standards now address AI directly. On 19 January 2026 the seven PCRT bodies (AAT, ACCA, ATT, CIOT, ICAEW, ICAS and STEP) published guidance on using AI in tax work. Advisers remain responsible for the output, must keep appropriate oversight, and should be able to explain how a conclusion was reached, even where AI produced the work.
HMRC expects human control in software. HMRC's own January 2026 guidance on AI in software used to submit information to it calls for strong human oversight built in at appropriate stages.
Generic narratives are a known weak point. Practitioners report that AI-written narratives tend towards boilerplate and plausible but unsupported detail. Some advisers warn that claims prepared mainly by AI are unlikely to survive scrutiny without specialist review, and some firms advise clients to avoid AI in claim preparation altogether.

The fear isn't AI as such. It's AI that produces confident text the adviser can't verify, and then puts the adviser's name to it.

What tribunals keep saying

R&D enquiries and appeals turn on evidence from the time. In RDM Wines v HMRC, reported in October 2026, the tribunal found that the company's documents didn't support the timeline or extent of testing it described, and gave little weight to the after-the-fact account. Commentators drew the same lesson: a well-written report can explain a claim, but it can't replace records showing what was actually happening, and those records often already exist in project-management tools, specifications, test results and issue logs.

What this means for a claim

Every statement in the narrative should be traceable to a record made at the time.
Gaps should be found before filing and closed with the people who did the work, not filled with plausible prose.
Routine work should be visibly excluded, so the claim isn't inflated.
Any tool used should leave the adviser able to explain and verify every conclusion.

That is the brief ClaimEvidence is built to. See how it works.

Sources

This page summarises public reporting and guidance for context. It is not legal or tax advice, and some sources are firms with a commercial interest in R&D advice. Where we report claims made by others, such as the Financial Times report, we say so. If anything here is out of date or wrong, tell us at support@claimevidence.co.uk and we'll correct it.