HMRC R&D Tax Claim Transparency AI: The Elsbury Tribunal Ruling That Changes Everything

HMRC R&D Tax Claim Transparency AI

Executive Summary: What Happened

In August 2025, the First-tier Tribunal ruled in Elsbury v The Information Commissioner that HMRC must disclose whether it used AI when assessing R&D tax relief claims. Judge Alexandra Marks overturned both HMRC’s refusal and the Information Commissioner’s Office (ICO) support of that refusal, declaring the public interest in transparency outweighed fraud prevention arguments.

The ruling represents the first time a UK tax authority has been legally compelled to reveal AI deployment in taxpayer assessments, with profound implications for the £7.6 billion R&D tax relief system and 46,950 annual claimants.

The HMRC R&D Tax Claim Transparency AI Controversy: Origins

Suspicious Patterns in HMRC Correspondence

Thomas Elsbury, R&D tax specialist and co-founder of Novel software platform, identified anomalies in HMRC compliance letters throughout 2023:

• American spellings (e.g., “analyze” instead of “analyse”)
• Em-dash punctuation (—) uncommon in UK government writing
• Generic responses that didn’t address specific claim facts
• Formulaic language suggesting template generation

These patterns triggered industry-wide concerns that HMRC was using large language models like ChatGPT without disclosure, potentially exposing confidential business information to public AI systems.

The Freedom of Information Request

On December 14, 2023, Elsbury submitted a Freedom of Information Act (FOIA) request asking HMRC to confirm:

1. Whether AI systems were used by the R&D Tax Credits Compliance Team
2. What AI technologies were deployed
3. What selection criteria governed AI use
4. What safeguards protected taxpayer data

HMRC initially confirmed it held the requested information but refused disclosure under Section 31(1)(d) FOIA, citing prejudice to tax assessment and collection.

The Legal Battle: HMRC’s Shifting Positions

Initial Refusal: The Fraud Prevention Argument

HMRC claimed revealing AI use would enable fraudsters to “game the system” by understanding compliance review methods. The agency argued transparency would:

• Help bad actors circumvent detection mechanisms
• Reduce effectiveness of compliance checks
• Increase fraudulent claim success rates
• Undermine the £7.6 billion relief scheme’s integrity

The “Neither Confirm Nor Deny” Pivot

After Elsbury requested internal review, HMRC pivoted to a “neither confirm nor deny” (NCND) stance, refusing to acknowledge whether it even held relevant information about AI deployment.

Judge Marks later described this reversal as “untenable,” “beyond uncomfortable,” and “like trying to force the genie back in its bottle.”

ICO Support Collapses

In November 2024, the Information Commissioner’s Office backed HMRC’s NCND position, agreeing that confirming AI use posed “unsubstantiated and unevidenced” fraud risks.

The Tribunal Ruling: Transparency Wins

Judge Marks’ Compelling Reasoning

On August 2, 2025, the First-tier Tribunal (General Regulatory Chamber) ruled decisively for Elsbury, finding:

“The balance of the public interest lies in disclosing the information requested.”

According to ACCA reporting, Judge Marks determined that:

1. Public Interest Prevails: Taxpayers have fundamental rights to understand how tax decisions affecting their businesses are made, especially when AI influences those decisions
2. Trust Undermined: HMRC’s secrecy “reinforces the belief based on indicators in HMRC correspondence that AI is being used by HMRC officers — perhaps in an unauthorised manner — thus undermining taxpayers’ trust and confidence”
3. Policy Objectives Frustrated: Lack of transparency could deter legitimate R&D claimants, directly contradicting the scheme’s purpose to encourage UK innovation
4. Fraud Risks Overemphasized: The ICO gave undue weight to speculative fraud concerns without adequate evidence

Compliance Deadline and Enforcement

The Tribunal ordered HMRC to respond within 35 working days (deadline: September 18, 2025) by either:

• Disclosing whether and how AI was used in R&D claim assessments
• Serving a proper refusal notice under Section 17 FOIA

Failure to comply could trigger contempt proceedings against the tax authority.

ICO Declines Appeal; HMRC “Considering Position”

The Information Commissioner confirmed it would not appeal the decision. HMRC stated it was “carefully reviewing the decision and considering next steps,” language suggesting potential appeal consideration.

HMRC’s Eventual Response: No Generative AI Used

Following intense media scrutiny, HMRC confirmed in September 2025 via Financial Times reporting that:

“The R&D tax credits compliance team did not use generative AI as part of work on R&D tax relief claims” and that “this technology was not approved for use in generating taxpayer letters.”

However, this narrow denial left critical questions unanswered:

• Machine learning systems: Did HMRC use non-generative AI for risk scoring or claim prioritization?
• Informal use: Were individual HMRC officers using public AI tools like ChatGPT without authorization?
• Data processing: What automated systems classify claims as “high-risk” for compliance review?
• Future deployment: What AI capabilities are included in HMRC’s £2 billion AI transformation roadmap?

Why HMRC R&D Tax Claim Transparency AI Matters

The Scale of R&D Tax Relief

According to UK Government statistics (September 2025):

• £7.6 billion total R&D tax relief claimed (2023-24)
• 46,950 claims submitted (down 26% from prior year)
• £46.1 billion qualifying R&D expenditure
• 17% compliance check rate (up from 10%)
• £441 million incorrectly claimed relief identified
• 77% adjustment rate on reviewed claims

The AI Deployment Context

HMRC operates within aggressive government AI adoption mandates:

• £2 billion allocated over four years for public sector AI implementation
• AI Opportunities Action Plan (January 2025) fast-tracking government AI projects
• Transformation Roadmap commits HMRC to “harness the immense potential of artificial intelligence”
• 70% of global tax authorities already use AI systems (OECD data)

The National Security Dimension

Elsbury raised critical concerns about AI exposure of classified R&D projects:

Defense contractors claiming R&D relief for weapons systems, surveillance technology, or military innovations could inadvertently expose sensitive details if HMRC uploads claim data to public AI platforms like ChatGPT.

Example scenario: An HMRC officer copies text from a defense R&D claim into ChatGPT to draft a rejection letter. That data enters ChatGPT’s training corpus, potentially accessible to adversaries through prompt engineering or data breaches.

STEP professional body reporting highlighted this creates “catastrophic consequences” if Ministry of Defence R&D details leak through AI systems.

The Technical Reality: How AI Could Be Used in R&D Claims

Risk Scoring and Claim Prioritization

AI systems excel at pattern recognition across massive datasets. HMRC likely employs machine learning to:

• Flag anomalous cost ratios (e.g., subcontractor costs exceeding sector norms)
• Identify templated narratives suggesting low-quality advisor involvement
• Detect keyword patterns associated with historically fraudulent claims
• Score compliance risk to allocate investigator resources efficiently

Automated Correspondence Generation

Large language models like ChatGPT could theoretically:

• Draft initial enquiry letters based on claim data
• Generate technical questions for specific industries
• Summarize complex claim documentation for case officers
• Create rejection rationales citing relevant legislation

The “Garbage In, Garbage Out” Problem

Tax specialists at Crowe UK demonstrated that AI outputs depend entirely on input prompts:

Test case: Asking Microsoft Copilot to produce two reports on whether identical assets qualified for capital allowances treatment—one arguing “yes,” one arguing “no”—generated equally compelling, contradictory analyses drawing from actual HMRC manuals and case law.

Critical question: If HMRC used AI, what prompts were officers using? “Analyze this claim objectively” versus “Draft a letter denying this claim” produce vastly different outcomes.

Implications for R&D Tax Claimants

Immediate Actions for Businesses

1. Document Quality Matters More Than Ever

With 17% of claims now subject to compliance checks (up from 10%), advisors recommend:

• Project-specific narratives: Avoid templated descriptions; detail unique technical uncertainties
• Contemporaneous records: Maintain real-time documentation as R&D progresses, not retrospective reconstructions
• Cost bridges: Clearly link claimed costs to specific R&D activities with transparent methodologies
• Technical depth: Demonstrate competent professional challenges, not routine problem-solving

2. AI-Resistant Claim Preparation

Assume both human reviewers and AI systems will analyze claims:

• Avoid generic language: AI flags templated phrases across multiple claims
• Sector context: Explain why costs deviate from industry norms before HMRC questions arise
• Baseline knowledge: Articulate what was already known versus what your project advanced
• Failure documentation: Show iterative experimentation, including unsuccessful approaches

3. Challenge Formulaic HMRC Responses

Tax dispute advisors note the Elsbury ruling “opens a fresh line of attack where decisions appear formulaic, automated or unreasoned.”

If HMRC correspondence contains:

• Generic objections not addressing your specific technical work
• Boilerplate language inconsistent with claim details
• Unusual formatting or spelling patterns

Request confirmation whether AI was involved in generating the assessment.

4. Submit Additional Information Form (AIF) Diligently

The mandatory AIF introduced April 2023 feeds HMRC’s compliance algorithms. Incomplete or inconsistent submissions trigger automated risk flags.

The Broader AI Governance Crisis

HMRC’s AI Transformation Plans

HMRC’s July 2024 Transformation Roadmap commits to “embedding GenAI in the tax authority’s operations,” yet:

• No AI Charter exists despite calls from professional bodies
• Privacy notice updates acknowledge AI use “where the law allows,” but legal boundaries remain undefined
• Legacy systems constrain AI deployment effectiveness (Public Accounts Committee findings)
• Governance frameworks lag technological capability

The Executive Liability Question

As AI influences tax decisions affecting billions in corporate finances, the question of accountability intensifies. Legal analysis from Ropes & Gray notes:

“Tax is an area where, given the sums at stake, the potential for iterative errors, and the intersection with issues of citizenship and human rights, [transparency] is especially acute.”

UK tax administration operates under a 1970 statutory regime predating modern computing, let alone AI. The Elsbury case “only increase[s] the sense that reform is overdue.”

International Context: Global Tax Authority AI Adoption

OECD Data: 70% AI Adoption Rate

According to ACCA reporting on OECD research, approximately 70% of global tax authorities already deploy AI systems for:

• Fraud detection and risk assessment
• Taxpayer segmentation and targeting
• Compliance intervention prioritization
• Data matching across information sources

Comparative Transparency Standards

The UK tribunal ruling stands out internationally:

• European Union: GDPR Article 22 requires notification when solely automated decisions affect individuals, though tax authorities often claim human involvement
• United States: IRS uses AI extensively but faces limited disclosure requirements beyond Freedom of Information Act requests
• Australia: ATO publicly acknowledges AI use in risk assessment but provides minimal operational detail

The Elsbury precedent may pressure other jurisdictions toward greater transparency.

The R&D Tax Relief Compliance Environment

HMRC’s Anti-Fraud Offensive

Since creating the dedicated R&D Anti-Abuse Unit (July 2022), HMRC has dramatically escalated enforcement:

2022-23 Analysis revealed:

• 25% of SME scheme claims were erroneous or fraudulent
• £759 million in error and fraud (9.9% of total relief)
• Compliance check rate doubled from 10% to 17%

2023-24 Compliance Results:

• £441 million incorrectly claimed relief identified
• 77% of checked claims required adjustment
• 26% overall claim volume decrease (deterrence effect)

The Collateral Damage Concern

Professional bodies and tax advisors warn that anti-fraud measures create “collateral damage” where:

• Legitimate innovators face rejection due to documentation technicalities
• Smaller companies lack resources for compliance-grade claim preparation
• First-time claimants discouraged by perceived bureaucracy (41% decline in new claimants)
• R&D-intensive startups delay claims due to Additional Information Form complexity

The Merged Scheme Era: April 2024 Forward

Fundamental System Restructuring

For accounting periods beginning April 1, 2024 or later, the UK operates under the new Merged R&D Expenditure Credit:

Rate Structure:

• 20% “above the line” credit (taxable)
• 15% effective relief for profit-making companies (after 25% corporation tax)
• 16.2% effective relief for loss-making companies
• 27% for R&D-intensive SMEs via Enhanced R&D Intensive Support (ERIS)

Key Changes:

• Single scheme replaces separate SME and RDEC tracks
• Subcontracted R&D restrictions loosened (can now include companies)
• Overseas R&D expenditure limited
• 30% intensity threshold for ERIS (reduced from 40%)

2025-26 Statistical Preview:

HMRC’s September 2026 publication will reveal merged scheme impacts for first time, providing critical data on:

• Claim volume stabilization or continued decline
• Average claim value trajectories
• ERIS uptake among R&D-intensive SMEs
• Sectoral and regional distribution shifts

What’s Next: The AI Transparency Agenda

HMRC’s Crossroads

The tax authority faces strategic choices following the Elsbury ruling:

Option 1: Appeal

• Challenge tribunal interpretation of FOIA exemptions
• Argue national security or fraud prevention justifications
• Risk further reputational damage and policy criticism

Option 2: Limited Disclosure

• Confirm no generative AI in claim assessment (already done)
• Remain silent on machine learning, automated scoring, or informal officer use
• Maintain opacity on future AI deployment plans

Option 3: Comprehensive Transparency

• Publish AI Charter detailing permitted and prohibited uses
• Disclose risk scoring methodologies and algorithmic criteria
• Implement mandatory AI disclosure in compliance correspondence
• Establish independent oversight mechanism

The Parliamentary Dimension

The Public Accounts Committee has called for:

• Government-wide address of public concerns over AI data sharing
• Transparency and accountability frameworks before widespread deployment
• Legacy system modernization enabling proper AI governance
• Clear legal boundaries for tax authority AI use

Professional Body Pressure

Organizations including STEP, ACCA, CIOT, and ATT are demanding:

• Formal HMRC AI Charter
• Mandatory disclosure when AI influences tax decisions
• Human review requirements for AI-generated assessments
• Data protection safeguards for taxpayer information in AI systems

Practical Guidance: Navigating the AI Era

For R&D Claimants

Assume Algorithmic Review:

• Prepare claims that satisfy both human experts and pattern-recognition AI
• Avoid language or structures that could trigger automated risk flags
• Document technical uncertainty with sector-specific terminology

Request AI Disclosure:

• When receiving HMRC enquiries, ask whether AI was involved in assessment
• Cite Elsbury v ICO [2025] UKFTT 915 (GRC) as precedent
• Challenge decisions that appear formulaic or generic

Elevate Documentation Standards:

• Treat claim submission as audit-grade compliance exercise
• Maintain contemporaneous project records, not retrospective narratives
• Link costs transparently to qualifying R&D activities

For Tax Advisors

Quality Over Volume:

• HMRC statistics show the market is consolidating toward high-quality advisors
• Generic, templated claims face 77% adjustment probability
• Invest in technical expertise and robust claim methodologies

Challenge Automation:

• The Elsbury ruling creates new grounds for tribunal appeals
• Question HMRC decisions showing AI characteristics (formulaic language, generic objections)
• Demand human review when automated processes appear to drive outcomes

Stay Informed:

• Monitor HMRC’s AI Transformation Roadmap implementation
• Track tribunal decisions involving AI disclosure requests
• Participate in professional body consultations on AI governance

Conclusion: The Transparency Imperative

The Elsbury tribunal ruling establishes that HMRC R&D tax claim transparency AI is not a technical preference but a legal requirement grounded in public interest.

As Judge Alexandra Marks determined, taxpayers possess fundamental rights to understand when and how artificial intelligence influences decisions affecting billions in tax relief for UK innovation.

The ruling’s implications extend far beyond R&D claims:

• Precedent for VAT, PAYE, and Corporation Tax: Other HMRC divisions deploying AI face similar transparency obligations
• International influence: UK leads global tax authority accountability standards
• Democratic governance: Public institutions using AI must operate within transparent, auditable frameworks

For UK businesses claiming R&D tax relief in 2026 and beyond, the message is unambiguous:

Document meticulously. Question formulaic HMRC responses. Demand transparency when AI is suspected. The law now requires it.

Key Takeaways

• First-tier Tribunal forced HMRC to disclose AI use in R&D tax claim assessments, overturning both HMRC and ICO refusals (Elsbury v ICO [2025] UKFTT 915 GRC)
• Judge Marks ruled public interest in transparency outweighs fraud prevention concerns, finding HMRC secrecy “undermines taxpayers’ trust and confidence” and deters legitimate claims
• HMRC confirmed no generative AI used in R&D compliance but left unanswered questions about machine learning, informal officer use, and future deployment plans
• £7.6 billion R&D tax relief system faces 17% compliance check rate with 77% of reviewed claims requiring adjustment, creating high-stakes environment where AI influence matters
• The ruling establishes precedent that UK taxpayers have legal rights to know when and how AI influences tax decisions, extending beyond R&D to entire HMRC operations

Frequently Asked Questions

Does HMRC currently use AI to assess R&D tax claims?

HMRC confirmed in September 2025 that its R&D tax credits compliance team “did not use generative AI” in assessing claims. However, this narrow statement doesn’t address machine learning systems for risk scoring, automated claim prioritization, or informal use of public AI tools by individual officers. The tribunal ruling compels further disclosure about non-generative AI systems.

What was the Elsbury case about?

Thomas Elsbury, an R&D tax specialist, submitted a Freedom of Information request asking HMRC to confirm whether it used AI in R&D claim assessments after noticing suspicious patterns (American spellings, em-dashes, generic responses) in HMRC correspondence. After HMRC and the ICO refused disclosure, the First-tier Tribunal ruled in August 2025 that HMRC must reveal AI use, finding public interest in transparency outweighed fraud prevention arguments.

How does the tribunal ruling affect my R&D claim?

The ruling doesn’t change R&D eligibility criteria but establishes you have the right to know if AI was involved in assessing your claim. If you receive HMRC enquiry letters with formulaic language or generic objections not addressing your specific technical work, you can now request confirmation whether AI was used and challenge automated-seeming decisions in tribunal proceedings.

What are the signs HMRC might be using AI on my claim?

Potential indicators include: correspondence with American spellings or unusual punctuation (em-dashes), generic rejection language not addressing your specific technical uncertainties, boilerplate responses appearing across multiple claimants, unusually fast turnaround times for complex technical assessments, or questions that don’t align with the facts in your claim documentation.

Can I appeal an HMRC decision if AI was involved?

Yes. Tax advisors note the Elsbury ruling “opens a fresh line of attack where decisions appear formulaic, automated or unreasoned.” If you suspect AI influenced your claim assessment, you can request disclosure under FOIA, challenge the decision in First-tier Tribunal, and argue that automated processes violated your right to proper human review of complex technical judgments.

What should I do differently when preparing R&D claims now?

Assume both human reviewers and AI systems will analyze your claim. Avoid templated language that AI could flag across multiple submissions, provide project-specific technical narratives with sector context, explain cost ratios that deviate from industry norms proactively, maintain contemporaneous documentation as work progresses, and prepare claims to “audit-grade” compliance standards given the 17% review rate and 77% adjustment probability.