R&D tax relief •12 min read

Get Onbord Ltd v HMRC

MJ

Matthew Jones ACA CTA

Last reviewed September 2026 · Editorial standards

Tribunal
First-tier Tribunal (Tax Chamber)
Decided
9 July 2024
Judge
Judge Mark Baldwin and Mr Mohammed Farooq

The company won. A tribunal accepted that an artificial intelligence and machine learning project to automate KYC verification was R&D, and accepted a competent professional who held no formal qualification in the field. It is the decision software and AI claimants cite most. It is also narrower than it is usually made to sound. HMRC ran one objection only, the tribunal decided it on one witness, and the passage advisers quote most often is the passage the tribunal said formed no part of its reasoning.

What was at issue

Get Onbord Limited claimed an R&D tax credit under section 1054 CTA 2009 — the old SME scheme — for building an automated process to verify identity and profile risk for “know your client” (KYC) checks. By the time the decision was released the company was in liquidation, with the appeal continued by its joint liquidators. The judgment states no accounting period and no claim figure, so neither is given here. It does record that the credit had already been paid: applied against the company’s unpaid PAYE liabilities in June 2021 (paragraph 68).

The adviser’s letters to HMRC of 29 September 2021 describe the work at length, and the detail matters more than the label. The screening engine ran against the OFAC, United Nations, Interpol, EU and UK sanctions lists, the company’s own politically-exposed-persons database and a news feed (paragraph 32). The harder part, on the same account, was identity resolution: establishing that a person appearing across those sources was one person and not several. On the model choice, the letters say decision-tree models were preferred to “black-box” alternatives because “white-box” techniques “trade some predictive power for transparency, which is useful for explaining to a regulator” (paragraph 40). None of that is a tribunal finding. It is the claimant’s own description, reproduced in the narrative part of the decision.

HMRC raised one objection and only one: that the work did not advance overall knowledge or capability in the field (paragraph 1). Two questions sat inside it. Could the company’s witness, a former director with no formal qualification in software, speak as a competent professional? And did building on open-source components mean nothing was advanced?

How the dispute got to a hearing

The adviser filed a report on 4 August 2021. HMRC replied on 15 September saying the work was not R&D. On 12 November 2021 it called the platform “a complex platform … impressive” while remaining “still unclear as to what advance has been achieved as per para 6 of the BIS guidelines” (paragraph 34). On 7 January 2022, after two further exchanges, HMRC concluded that no R&D had taken place; the decision records that “No reasons were given for that view” (paragraph 42). A review conclusion letter of 14 October 2022 upheld the refusal.

The appeal, reference TC/2022/13281, was heard by remote video on 3 January 2024. The company had gone into liquidation in October 2023, which nobody told the tribunal until HMRC did so on 10 January 2024. The joint liquidators asked to adopt the arguments the former director had already run. HMRC said the hearing should be held again, because he had appeared without standing. The tribunal held that the liquidators could adopt the argument and refused HMRC’s application (paragraphs 11 and 16). Written submissions closed on 5 April 2024; the decision was released on 9 July 2024.

What each side argued

HMRC’s case, in written submissions at paragraph 70, was not that AI work cannot qualify. It was that the company had argued the wrong test. Market functionality is not the statutory question; paragraph 6 of the Guidelines defines an advance by reference to overall knowledge, not by what competitors sell. Automating a manual process may use existing technology or may require an advance in the underlying technology, and the company had to show the latter. The consequence of the company’s reading was, in HMRC’s words, that “Given that pretty much every computer programme written has unique coding and unique or bespoke functionality, this would therefore extend R&D relief to all coding activity, which is clearly not what was intended.” Two evidential points ran alongside: no plans, milestones or correspondence showed a hypothesis-and-test cycle, and the company’s own software lead gave no evidence.

The company accepted almost all of that framing and disputed its application (paragraphs 66 and 71). KYC, as distinct from identity verification, remained largely manual; products existed for individual functions but none operated holistically. Existing code and APIs are used in nearly every software development, and treating that as disqualifying would remove all software from the relief. The claimed advance was not in compliance practice: “The company has never claimed that it was an advancing compliance or KYC/AML mechanisms” (paragraph 71(7)). On automation, the company said that “the fact that a human can drive a car does not mean that developing a self-driving car is not a massive innovation and technological advance” (paragraph 71(6)).

The provision the case turns on

Entitlement came from section 1054(1) CTA 2009, in the form in force before the merged scheme: “A company is entitled to an R&D tax credit for an accounting period if it has a Chapter 2 surrenderable loss in the period (see section 1055).” That loss runs back through sections 1055, 1044 and 1051 to section 1052(3), as it then stood: “Condition B is that the expenditure is attributable to relevant research and development undertaken by the company itself.”

The definition sits outside the Act. Section 1041 CTA 2009 defers to section 1138 CTA 2010, then to section 1006 ITA 2007, which authorises regulations made by reference to guidelines. Regulation 2 of SI 2004/712 then provides that activities treated as research and development in accordance with the Guidelines “are research and development”, and that activities not so treated “are not research and development”. The operative words are therefore the Guidelines’ own. Paragraph 6: “An advance in science or technology means an advance in overall knowledge or capability in a field of science or technology (not a company’s own state of knowledge or capability alone).” Paragraph 20 measures that against what “is publicly available or is readily deducible … by a competent professional working in the field”. HMRC restates both at CIRD81300, amended 19 January 2026.

The condition the tribunal did not reach is section 1057(4A): “A company is not a going concern at any time if it is in administration or liquidation at that time.” HMRC sought permission to argue it, then withdrew: the credit had been applied against PAYE in June 2021, so section 1057(3) preserved the claim (paragraph 68).

What the tribunal decided

The appeal was allowed (paragraph 98). The company’s complaints about HMRC’s handling were put aside: “Questions as to whether HMRC have applied the law fairly and consistently are, if anyone seeks to raise them, for the Administrative Court” (paragraph 73). The burden was held to stay on the company throughout, subject to the shifting evidential burden drawn from Brady v Group Lotus and Wood v Holden (paragraphs 74 to 77).

On the person, both sides’ witnesses were assessed. HMRC’s officer was honest but, having “no technology experience or expertise”, his evidence “was of no real help to us in deciding the issues before us” (paragraphs 64 and 65). Of the company’s witness, the tribunal noted that he “does not have any formal qualifications in this area”. It was nonetheless “completely satisfied that [the witness] has experience (including in coding) and up-to-date knowledge of software capabilities, albeit perhaps only in the area he works in, to be a ‘competent professional’ for our purposes” (paragraph 82).

The tribunal then declined to decide whether evidence from a competent professional is required at all. It was sceptical of the contrary view: “We are not sure that we would have agreed with Mr Lewis with the same degree of alacrity as the FTT did in Flame Tree Publishing”. It was, it said, “a relief not to have to express a conclusion on this issue” (paragraph 84).

On the technology, there was a project, because the objective “defined the boundaries of what GOL was doing” (paragraph 88). The uncertainty was whether an AI system could do the work at all, not anything about the compliance rules themselves (paragraph 89). The tribunal was careful that “the test in the BEIS Guidelines is not whether a new function has been developed”. What it accepted was that creating a new function and “solving a real-world problem in a new and creative way using technology, is at least an indication that there has been an appreciable technological advance” (paragraph 93). On open source: “We do not consider that, of itself, the use of open source, or other existing, materials is an indication that a particular development is routine or readily discernible from publicly available materials” (paragraph 94). Paragraph 95 states the findings in statutory form — a project with a defined aim, technology “not already publicly available or readily deducible”, more than “routine” copying or adaptation, and uncertainties “which a competent professional working in the field could not have readily resolved”.

The part people over-read

Paragraph 96 is quoted more than any other sentence in the decision, usually to suggest the burden moves to HMRC once a claimant has said enough. The tribunal said the opposite about its own use of it: “To the extent that the shifting of the evidential burden might suggest a lower state of satisfaction (and we are not sure that it does), that forms no part of our analysis.” Nor is this a decision that AI projects qualify. HMRC had conceded everything except the advance, so the case never tested costs, records or apportionment — the grounds on which claims are more often lost.

What it changes for a claim being prepared now

  • A competent professional is made by expertise, not by paperwork. A self-taught engineer with current command of the field can be the right person. What the tribunal wanted was evidence of that command, given by the person themselves and tested in cross-examination.
  • Building on open source does not disqualify the work. The question is whether the capability you set out to create was already publicly available or readily deducible by a competent professional, not whether every part was new.
  • Say what was not deducible, and why. That sentence is the claim. A claim that the work combined existing tools in a new way is not it. The shape, on invented facts: a competent professional in the field, working from the methods published before the project began, could not have deduced whether the components would hold together at the throughput the design required. The company did not know it was achievable until the third architecture was tested.
  • Put the technical case in one document, early. The closing suggestion at paragraph 99 is that both sides “put their scientific cards face up on the table”. For the company that means marshalling its technical evidence in one document, including evidence from a competent professional — “which is clearly highly desirable, whether or not it is strictly necessary”. HMRC then replies with its own analysis and evidence.
  • Watch the going-concern condition if insolvency is in prospect. Get Onbord kept its credit only because it had already been applied against PAYE before liquidation. A company entering administration with a claim still outstanding is in a different position. The condition has since moved: for accounting periods beginning on or after 1 April 2024 it is in sections 1112F and 1112G CTA 2009. Section 1112G(2) still provides that “a company is not a going concern if it is in administration or liquidation”. Section 1112F(6) preserves the same saving for an amount “paid or applied before the company ceased to be a going concern” — see insolvency and R&D tax relief.

Where AI work does and does not qualify is taken further in machine learning and qualifying R&D; the person question in who counts as a competent professional and AI and robotics.

Where it sits against the other decisions

Get Onbord is the positive end of a line that otherwise reads against claimants. It applies the same strict construction as the rest of the register, taking Gripple at [12] through Hadee Engineering (paragraphs 27 and 28). It also quotes AHK Recruitment on evidence rather than assertion (paragraph 78). It reaches a different result because the evidence was there.

Read against Flame Tree Publishing, decided three months earlier and argued by the same HMRC litigator, the two are consistent: one claimant had a witness who could speak for the field, the other did not. Later decisions have followed Flame Tree — Tanglewood Care Services at paragraph 112, and Beer Express Ltd [2026] UKFTT 672 (TC) at paragraphs 89 and 91, where a director found “truthful and credible” was still not a competent professional for the projects claimed.

Two decisions supply the counterweight. BE Studios Ltd v Smith & Williamson Ltd [2005] EWHC 1506 (Ch), a professional negligence claim rather than a tax appeal, found at [45]–[46] that a product being novel and built with technology does not make it R&D. Get Onbord does not cite it; later decisions do, and it answers the over-reading of paragraph 93. On going concern, MW High Tech Projects UK Ltd [2023] UKFTT 1040 (TC) shows the condition unsatisfied: an RDEC claim of £1,934,343.19 extinguished because the latest published accounts were not prepared on a going concern basis.

Has it been appealed?

No onward appeal has been reported, and there is no Upper Tribunal decision in the case on the published record. Applications for permission to appeal, and refusals dealt with on the papers, are not consistently published, so that is not proof that none was made. HMRC has not changed its published guidance as a result: the CIRD manual’s dated change log records no amendment attributed to this decision, and no Revenue and Customs Brief or Agent Update names it.

The decision is First-tier Tribunal: it binds only the parties and sets no precedent. Its weight is practical — it is the decision software and AI claimants reach for, and the one an officer will want to distinguish. It sits with the rest of the register of R&D tribunal decisions, one page per decision.

Sources

Every CIRD paragraph cited above (CIRD81130, CIRD81300 and CIRD81960) is indexed on our CIRD reference index, with HMRC’s own title for each, a line on what it says and the date HMRC last revised it.

First-tier Tribunal decisions bind only the parties to them and set no precedent. This entry describes the decision as it stood at the review date above, as general information rather than advice on your circumstances — see our terms. For an answer on your own facts, talk to us.