Quick Answer: Does HMRC Use AI To Assess R&D Tax Claims?
HMRC uses artificial intelligence, automation and advanced data tools across its wider operations. However, HMRC reportedly told a claimant after a 2025 tribunal ruling that its R&D Tax Relief Compliance Team did not use generative AI as part of its work on R&D tax relief claims and that the technology was not approved for generating taxpayer letters.
HMRC also maintained that human officers open R&D enquiries, review claims, communicate with businesses and make final decisions.
It reportedly acknowledged testing an internal AI-related summarisation tool in 2024, but said the limited test involved cases that caseworkers had already assessed.
The First-tier Tribunal did not decide that HMRC had used AI to reject R&D claims. It ruled that HMRC had to respond properly to a Freedom of Information request about the possible use of large language models and generative AI in R&D compliance work.
The decision was therefore about HMRC R&D claim transparency, rather than proof of automated tax decisions.
Key Takeaways:
- The HMRC AI tribunal ruling concerned information rights and public-sector transparency.
- It did not establish that artificial intelligence autonomously approved or rejected R&D tax claims.
- HMRC reportedly denied authorised generative AI use within its R&D compliance team but acknowledged a limited internal tool test.
- Allegations that individual caseworkers used unauthorised AI tools remain attributed claims rather than findings made by the tribunal.
- HMRC’s July 2026 roadmap confirms that AI-powered systems are being expanded across compliance, administration and customer service.
- Businesses using AI to prepare claims remain responsible for every technical statement, calculation and submission.
- Strong evidence is more useful during HMRC R&D compliance checks than speculation about whether a standard-looking letter was AI-generated.
Why Has HMRC AI Use In R&D Tax Claims Became A Transparency Issue?
R&D tax relief claims can contain commercially sensitive information about experimental technology, internal processes, product designs and unresolved scientific or technological problems. Businesses provide this material to HMRC expecting it to be handled securely and assessed fairly.
Concerns about HMRC AI use in R&D tax claims arose after advisers reported seeing enquiry correspondence containing unusual wording, apparent factual inconsistencies and features they believed could indicate generative AI use.
Those observations created suspicion, but they were not proof that an AI system had been involved.
The resulting debate covers several different questions:
- Does HMRC use algorithms to select R&D claims for examination?
- Can automated tools summarise evidence for caseworkers?
- Are generative AI systems permitted to draft correspondence?
- Does a human independently evaluate an AI-generated recommendation?
- How is confidential technical information protected?
- Can a claimant discover whether an automated system influenced a decision?
These questions should not be collapsed into a single claim that “HMRC uses AI.” Risk analysis, document recognition, summarisation, drafting and autonomous decision-making are materially different activities.
HMRC’s current transformation plans reinforce that distinction.
Its July 2026 roadmap says the department is developing systems to assist and accelerate compliance work, improve document-fraud detection and provide caseworkers with AI-powered tools, including generative AI systems that give guidance. The roadmap does not say those systems autonomously decide R&D claims.
What Happened In The HMRC AI Tribunal Case?
The case was Thomas Elsbury v The Information Commissioner [2025] UKFTT 00915 (GRC).
On 14 December 2023, Thomas Elsbury submitted a Freedom of Information request concerning the use of large language models and generative AI within HMRC’s R&D Tax Credits Compliance Team.
The request asked about:
- How the technology was being used
- Whether it supported data analysis or decisions
- The models and providers involved
- Data privacy and security
- Human oversight
- Policies and ethical controls
- Its impact on enquiry outcomes
- Plans for future use
HMRC initially confirmed that it held information covered by the request but withheld it under section 31(1)(d) of the Freedom of Information Act 2000, which concerns prejudice to the assessment or collection of tax.
During the Information Commissioner’s investigation, HMRC changed its position and sought to neither confirm nor deny whether it held the information.
The Information Commissioner accepted HMRC’s revised position. Elsbury appealed to the First-tier Tribunal.
The tribunal heard the case on 27 June 2025 and issued its decision on 2 August 2025. It allowed the appeal, finding that the Commissioner’s decision was not in accordance with the law and that the public-interest balance had been wrongly assessed.
HMRC was ordered to state whether it held information responding to each part of the request. Where information was held, it had to supply it or issue a legally valid refusal notice.
Why Did The Tribunal Favour Greater Transparency?
The tribunal found HMRC’s reversal difficult to sustain because the department had already confirmed twice that it held the requested information. The panel described the later attempt to use a neither-confirm-nor-deny response as untenable.
The tribunal recognised legitimate public interests on both sides.
HMRC and the Information Commissioner argued that disclosure could expose compliance methods, assist abusive claimants and prejudice tax collection.
The tribunal accepted that protecting public revenue and supporting HMRC’s compliance activity were important considerations.
However, it also gave substantial weight to:
- Accountability for public authorities
- Public understanding of AI use
- Taxpayer confidence in R&D compliance decisions
- Protection of confidential business information
- The risk that distrust could discourage legitimate claims
- The need for informed public debate about AI-assisted decision-making
The panel concluded that the claimed risks of increased fraud had been given too much weight without sufficient supporting evidence, while the societal benefits of transparency had received too little weight.
What The HMRC AI Tribunal Ruling Did And Did Not Establish?
The ruling is important, but it should not be overstated.
What The Tribunal Confirmed?
The tribunal confirmed that:
- The appeal concerned a Freedom of Information request.
- HMRC had first said it held relevant information.
- HMRC later tried to neither confirm nor deny that information was held.
- The Information Commissioner’s decision contained legal errors.
- The balance of public interest favoured requiring a proper response.
- Transparency about AI-assisted government activity was a legitimate public concern.
What The Tribunal Did Not Confirm?
The tribunal did not find that:
- HMRC had used AI to reject a particular R&D claim.
- A generative AI model was making final tax decisions.
- Every disputed HMRC letter had been written by AI.
- Confidential R&D information had been entered into a public chatbot.
- A taxpayer had received a penalty based solely on AI output.
- Any individual officer had breached HMRC policy.
The tribunal referred to the appellant’s concerns and industry suspicions, but reporting a party’s submissions is not the same as making a factual finding that the allegations were true.
That distinction is central to responsible reporting on algorithmic transparency in tax decisions.
What Did HMRC Disclose After The Tribunal Decision?
According to Financial Times reporting in September 2025, HMRC responded that its R&D tax credits compliance team did not use generative AI as part of its work on R&D tax relief claims. It also said the technology was not approved for producing taxpayer letters.
The reported response also addressed an internal R&D summarisation tool tested in 2024. HMRC said the tool was tested on a limited number of cases and that the relevant caseworker assessments and actions had already occurred.
HMRC’s reported position was that a person remained responsible throughout an enquiry: a human officer opened the case, reviewed the claim, communicated with the taxpayer and made the final decision.
This supports the case for human oversight of R&D tax claims, but it does not resolve every transparency question.
For example, a human can make a final decision while relying on:
- Automated risk indicators
- Data-matching results
- Document-classification systems
- AI-generated summaries
- Suggested responses
- Internal guidance produced by generative AI
The presence of a human at the end of a process does not automatically show how influential an earlier automated output was. Meaningful human oversight requires the officer to understand the evidence, test the output and remain able to reach a different conclusion.
What About Allegations Of Unauthorised AI Use?
The Financial Times separately reported claims from advisers that some individual HMRC caseworkers had used generative AI without formal approval. One source claimed that staff had been disciplined in 2023, while another said officers were later trained on permitted use.
HMRC maintained that generative AI was not used as part of its R&D claims work and said misuse of technology could lead to disciplinary action.
These reports remain allegations based on attributed sources. The publicly available tribunal decision did not determine:
- Whether unauthorised AI use occurred
- Which officers may have been involved
- Which cases may have been affected
- Whether confidential information was exposed
- Whether an AI-generated assessment changed a claim outcome
It would therefore be misleading to say that the tribunal “proved” unauthorised AI use.
It would be equally unsafe to interpret HMRC’s statement about approved team procedures as conclusive proof that no individual could ever have used an unapproved tool.
The strongest available description is that HMRC denied authorised use, while separate sources alleged unauthorised use by some individuals.
HMRC AI Use In R&D Tax Claims: Current Position At A Glance
| Question | Confirmed Or Reported Position | What Remains Unclear |
| Did the tribunal prove HMRC used AI to reject claims? | No. It ruled on HMRC’s response to an information request. | Whether any individual historic cases involved unauthorised tools. |
| Does AI officially make final R&D decisions? | HMRC reportedly says human caseworkers make final decisions. | How supporting analytics or AI-generated guidance may influence those decisions. |
| Did HMRC test an R&D-related AI tool? | HMRC reportedly acknowledged a limited internal summarisation test. | Whether a similar tool will be developed or deployed more widely. |
| Is HMRC increasing its use of AI? | Yes. The 2026 roadmap describes wider AI deployment and testing. | Which future systems may apply specifically to R&D tax compliance. |
| Must public-sector algorithmic tools be disclosed? | Certain influential tools used by government departments fall within the mandatory transparency standard. | Whether a particular HMRC tool is in scope, exempt or not yet operational. |
How HMRC Plans To Expand Artificial Intelligence?
The HMRC AI strategy 2026 is broader and more advanced than the position described during the original information dispute.
HMRC’s July 2026 Transformation Roadmap says more than 28,000 Microsoft Copilot licences had been issued to colleagues by March 2026, with plans to scale availability further.
The stated uses included drafting and summarising documents and emails, as well as summarising meetings.
The roadmap also says HMRC has:
- Created test environments using synthetic data
- Piloted AI-generated call summaries with human review
- Used AI simulations for customer-service training
- Developed more advanced document-recognition tools
- Held a private-sector competition for AI and data-science compliance solutions
- Awarded two businesses 12-month pilot contracts
- Planned AI-powered guidance systems for caseworkers
HMRC says its AI adoption will follow ethical, safety, technology, accessibility and Charter controls.
Does The Roadmap Announce Automated R&D Decisions?
No specific passage reviewed for this article says that HMRC will introduce an autonomous AI system to decide R&D tax relief claims.
However, the roadmap does state that HMRC is developing tools to assist, automate and accelerate compliance work. It also describes AI-powered systems intended to provide rapid and consistent guidance to caseworkers.
This creates an important future transparency issue.
Businesses do not necessarily need technical details that would allow bad actors to defeat HMRC’s risk systems. They do have a legitimate interest in understanding:
- Whether AI has a material influence on their case
- What types of data are processed
- Whether outputs are independently checked
- How inaccuracies can be corrected
- Who remains accountable
- How commercially sensitive information is protected
What Is Algorithmic Transparency In Tax Decision-Making?
Algorithmic transparency means giving the public understandable information about how and why an algorithmic tool is used, what role it plays and how it affects a broader decision-making process.
The UK Algorithmic Transparency Recording Standard, known as ATRS, provides a template for public-sector organisations to publish this information.
Government guidance says the standard is mandatory for government departments where an algorithmic tool has a significant influence on a decision-making process with public effect or interacts directly with the public.
The required transparency is not intended to expose every technical feature or reveal information that would enable fraud.
The guidance allows operational, cybersecurity and intellectual-property concerns to be managed through proportionate disclosure and justified exemptions.
It also says that explaining a tool’s place in the overall decision process can remain appropriate even when sensitive technical details cannot be released.
Why Human Approval Is Not The Same As Human Oversight?
A human officer clicking “approve” after reading an AI-generated summary would not necessarily provide meaningful oversight.
Effective review should involve the officer:
- Checking the source evidence
- Recognising the limitations of the system
- Identifying unsupported conclusions
- Considering contrary evidence
- Recording independent reasons
- Retaining authority to disregard the output
This distinction matters because an inaccurate summary can frame a case incorrectly before a human reaches the evidence.
Can Businesses Use AI To Prepare R&D Tax Claims?
Businesses can use AI-assisted software to help organise information or produce an initial draft. However, an AI-generated R&D tax claim must not be treated as a finished technical or tax assessment.
Generative AI may assist with:
- Structuring engineers’ notes
- Identifying missing questions
- Converting interviews into a draft narrative
- Comparing terminology across project records
- Preparing an initial evidence checklist
- Improving readability
It can also create serious risks.
An AI tool may invent project details, exaggerate an advance, describe routine development as qualifying R&D or produce a confident explanation unsupported by the company’s records.
The business remains responsible for the information submitted to HMRC. A software provider cannot verify what engineers actually attempted, when an uncertainty arose or whether a cost calculation agrees with payroll and accounting records.
What Information Must Support An R&D Tax Relief Claim?
HMRC requires an additional information form for new R&D tax relief or expenditure-credit claims. The form must be submitted before, or earlier on the same day as, the Company Tax Return. HMRC states that the claim will be rejected where the tax return is submitted first.
The form requires information about the company, responsible contacts, agents, qualifying expenditure and relevant projects. HMRC’s guidance also asks claimants to describe the field of science or technology and the work undertaken to resolve scientific or technological uncertainty.
Technical Evidence
A defensible claim should identify:
- The relevant field of science or technology
- The existing technological baseline
- The advance the project sought
- The scientific or technological uncertainty
- Why the answer was not readily deducible
- The competent professionals involved
- Tests, trials and failed approaches
- The point at which qualifying work began and ended
Financial Evidence
Businesses should retain:
- Payroll reports
- Staff-time calculations
- Contractor and subcontractor records
- Software invoices
- Consumable costs
- Apportionment methods
- Ledger extracts
- Reconciliation to the Company Tax Return
Governance Evidence
An AI-assisted claim should also record:
- Which tool was used
- What information was entered
- Who checked the output
- Which changes were made
- Who approved the final submission
- How confidential information was protected
Realistic Example: Reviewing An AI-Assisted R&D Claim
A software company develops a data-processing platform intended to handle unusually high transaction volumes without creating unacceptable delays.
Its finance team uses generative AI to turn interviews with engineers into a first draft of the project description.
The draft claims that the company created a “world-first algorithm.” The engineers did not make that claim, and no evidence supports it. The draft also treats a routine cloud migration as part of the technological uncertainty.
A competent professional reviews the output and makes several corrections:
- The unsupported “world-first” statement is removed.
- The existing technological baseline is explained.
- Routine migration work is separated from experimental development.
- Failed load-balancing approaches are documented.
- Project dates are checked against engineering records.
- Staff costs are reconciled with payroll and time records.
HMRC later opens a compliance check and sends a standardised letter questioning whether a qualifying advance existed.
The company does not assume that the letter was written by artificial intelligence. It responds to the actual concerns with architecture diagrams, test results, dated development records and a signed explanation from the competent professional.
The example illustrates the correct approach on both sides of the process: automation may assist, but evidence and accountable human judgement should determine the outcome.
What Should A Business Do During An HMRC R&D Compliance Check?
HMRC describes a compliance check as an examination of a taxpayer’s position to confirm that the right amount of tax has been paid and that allowances or reliefs have been claimed correctly. HMRC normally writes or calls to explain what it intends to check and why.
Review The Questions Carefully
Businesses should separate general template wording from questions that relate specifically to the claim.
For each point, determine:
- What HMRC is asking
- Which qualifying condition is relevant
- Whether the letter contains a factual mistake
- What evidence answers the concern
- Whether technical input is needed
Respond With Evidence Rather Than Assumptions
A repetitive or oddly worded letter does not prove AI involvement.
The stronger response is to:
- Answer every numbered question
- Correct inaccurate statements
- Cross-reference supporting documents
- Explain specialist terminology plainly
- Keep a complete correspondence record
- State where requested information has already been supplied
Ask For Clarification Where Necessary
A claimant may reasonably ask HMRC to explain:
- Which evidence was considered
- What part of the claim remains disputed
- Which eligibility condition HMRC believes was not met
- Whether the latest evidence has been reviewed
- What further information could resolve the issue
These questions focus on the quality of the decision rather than speculation about the software used to prepare a letter.
Can A Business Challenge An AI-Assisted HMRC Decision?
The involvement of AI would not automatically make a tax decision invalid. The central questions are whether HMRC applied the law correctly, considered the relevant evidence and followed a fair process.
HMRC’s guidance says that when it makes an appealable decision, its letter should explain what to do. A taxpayer will generally have three options within 30 days:
- Submit new information to the officer
- Request review by a different officer
- Appeal to an independent tribunal
Alternative dispute resolution may also be available and does not remove the right to appeal.
Professional advice may be appropriate where the dispute involves a substantial amount, penalties, complex eligibility issues or concerns about procedural fairness.
Common Misinformation About HMRC, AI And R&D Claims
“The Tribunal Proved HMRC Rejected Claims Using AI”
This is incorrect. The tribunal ruled on whether HMRC could avoid answering an information request. It did not determine the technology used in an individual claim decision.
“Every Generic HMRC Letter Is AI-Generated”
Standard wording, repeated phrases or an unusual punctuation style may raise questions, but they are not proof of AI authorship.
“HMRC Does Not Use AI Anywhere”
This is also incorrect. HMRC’s 2026 roadmap describes extensive use and testing of AI, including drafting, summarisation, customer-service support and compliance-related systems.
“An AI-Generated Claim Is Automatically Compliant”
No. A polished narrative may still contain false facts, unsupported conclusions or incorrect cost treatment.
“A Human Final Decision Removes All AI Risk”
A human decision-maker can still be influenced by an inaccurate automated output. The quality and independence of the review matter more than the presence of a human name at the end of the process.
Conclusion
The debate over HMRC AI use in R&D tax claims is ultimately about accountability, not simply technology.
The 2025 tribunal ruling established that HMRC had to respond properly to questions about possible generative AI use. It did not prove that artificial intelligence was rejecting claims.
HMRC subsequently denied authorised generative AI use in its R&D compliance work, while reports of unauthorised use by individual officers remain allegations.
At the same time, the HMRC AI strategy 2026 confirms that AI-powered systems will become increasingly common across the department. That makes clear disclosure, secure data handling and meaningful human oversight more important.
Businesses should respond by maintaining accurate technical evidence, controlling their own use of AI and challenging decisions through facts, records and established review or appeal procedures.
How Well Do You Understand the HMRC AI Ruling?
Answer four quick questions to separate confirmed facts from common assumptions.
Frequently Asked Questions
Does HMRC Use AI To Assess R&D Tax Claims?
HMRC uses AI across its wider operations. Its reported position is that generative AI was not used as part of the R&D Tax Relief Compliance Team’s claims work and that humans make final decisions.
Did The Tribunal Prove HMRC Rejected R&D Claims Using AI?
No. The tribunal decided that HMRC had to respond properly to a Freedom of Information request. It did not determine that AI had rejected any claim.
What Was The HMRC AI Tribunal Ruling About?
It concerned HMRC’s refusal to confirm or deny whether it held information about large language models and generative AI in its R&D compliance team.
Did HMRC Test An AI Tool On R&D Cases?
HMRC reportedly said it tested an internal summarisation tool on a limited number of previously handled cases in 2024.
Are HMRC R&D Compliance Letters Written By AI?
HMRC reportedly said generative AI was not approved for generating taxpayer letters. A letter’s wording alone cannot prove whether AI was used.
Can A Business Use AI To Prepare An R&D Claim?
AI may assist with organisation or initial drafting, but responsible people must verify the technical narrative, expenditure and supporting evidence before submission.
Can Confidential R&D Information Be Entered Into An AI Tool?
Only approved tools with suitable security, privacy and retention controls should receive confidential information. Public or consumer AI tools should not be assumed to provide appropriate protection.
Editorial Disclaimer: This article provides general information and is not tax or legal advice. Businesses should check current HMRC guidance and obtain advice based on their circumstances.
Source Links
- Thomas Elsbury v The Information Commissioner – Tribunal Decision
- HMRC Transformation Roadmap: Update 2026
- HMRC R&D Additional Information Form Guidance
- HMRC Compliance Checks: Help And Support
- Algorithmic Transparency Recording Standard Guidance
- Financial Times: Advisers Claim HMRC ‘Smoke And Mirrors’ Over AI Use
