UK Tax Authority to Deploy AI to Detect Fraud and Tax Return Errors

On 14 May 2026, HM Revenue and Customs announced a £175 million, ten-year contract with Quantexa, a London-based AI and data analytics firm, to overhaul the agency's data infrastructure and deploy artificial intelligence across tax compliance, fraud detection, and customer service [1]. The deal is one of the largest government AI procurement contracts in UK history and arrives as HMRC confronts a tax gap — the difference between taxes owed and taxes collected — that has surged to £46.8 billion [2].

The contract sits at the intersection of several high-stakes policy questions: whether algorithms can reliably distinguish fraud from honest mistakes, whether the technology will disproportionately target vulnerable taxpayers, and whether £175 million spent on AI would recover more revenue than the same sum invested in human investigators.

The Contract: What £175 Million Buys

Quantexa's platform will modernise HMRC's core data infrastructure, creating a unified view across datasets that currently sit in separate systems — income tax, corporation tax, VAT, and customs records [1]. The company's entity resolution technology links records that refer to the same individual, company, or transaction across fragmented databases, then layers machine learning on top to flag anomalies [3].

The contract covers three broad capabilities: data integration, AI-driven compliance analytics, and customer service improvements [4]. HMRC has stated that any decisions arising from the AI system will be "reviewed and signed off by a human" with relevant tax expertise [5].

The £175 million figure should be set against HMRC's broader IT spending. The agency pegs its annual contracts with IT suppliers at over £1 billion, with a further £2 billion technology pipeline under procurement including cloud services, data centres, and digital platforms [6]. HMRC's transformation roadmap targets a "digital-first" organisation with at least 90% of customer interactions happening online by 2030 [7].

The Tax Gap: A Problem That Keeps Growing

Britain's tax gap has expanded significantly in recent years. From £31 billion in 2018-19, it reached £39.8 billion in 2022-23 and has now been revised to £46.8 billion — roughly 5.6% of total theoretical tax liabilities [2][8].

Source: HMRC Tax Gap Reports

Data as of Jun 1, 2025CSV

The composition of that gap matters for evaluating whether AI is the right tool. HMRC's own breakdown by taxpayer behaviour shows that 31% of the gap is attributable to "failure to take reasonable care" and a further 15% to outright errors — categories that describe carelessness rather than criminal intent [9]. Evasion accounts for 14%, non-payment for 12%, and legal interpretation disputes for another 12%. Criminal attacks contribute 9%, the hidden economy 5%, and formal avoidance schemes just 1% [9].

Source: HMRC / GOV.UK

Data as of Jun 20, 2024CSV

In absolute terms, evasion costs the Treasury an estimated £5.5 billion per year, while avoidance accounts for roughly £1.8 billion — though Tax Justice UK has argued that HMRC's methodology likely underestimates both figures [8][10].

By taxpayer type, small businesses account for 60% of the overall gap, with their corporation tax gap exceeding 30% for four consecutive years — up from 11% in 2011 [8]. Wealthy individuals (approximately 800,000 people classified as such by HMRC) account for about 5% of the gap, while all other individuals combined contribute around 6% [11].

How the UK Compares

International comparisons are imperfect because tax authorities use different methodologies. The United States' "gross tax gap" was estimated by the IRS at $688 billion (roughly £545 billion) for tax year 2021, representing approximately 14.8% of total tax liability — significantly higher than the UK's 5.6% rate, though methodological differences make direct comparison unreliable [12]. The European Commission has estimated VAT gaps across EU member states averaging around 5.1% in 2022, broadly comparable to the UK's overall rate [2].

The US offers a useful comparison on AI deployment. The IRS now runs 126 active AI applications across audit selection, fraud detection, and taxpayer services — up from just 10 in 2022 [12]. In the first five months of fiscal year 2026, IRS enforcement revenue rose 12% despite a 25% workforce reduction, a result the agency attributed to technology-driven productivity [13].

Who Is Quantexa?

Quantexa was founded in 2016 by Vishal Marria, formerly the youngest executive director at EY [1]. The company built its reputation in anti-money laundering for banks, with HSBC and Standard Chartered among its early clients [1]. It reported revenue of £126 million for the year ending March 2025, up 49% year-on-year, and raised $175 million in a Series F round in March 2025 at a valuation of $2.6 billion [1].

In September 2024, Quantexa launched a dedicated Global Public Sector Business Unit [14]. Its first US federal contract came through a deal with US Special Operations Command (USSOCOM) for an AI-enabled news intelligence platform [14]. The HMRC deal, however, represents an order of magnitude larger than any prior government engagement.

The procurement process that led to the award has not been publicly detailed. HMRC has not disclosed whether it was a competitive tender, a framework call-off, or a direct award — a transparency gap that may attract scrutiny given the contract's size and ten-year duration [3][4]. The "sovereign AI" framing — Quantexa's data never leaves the HMRC environment — aligns with the government's Sovereign AI Unit, launched in 2025 with £500 million in funding [1].

Marria has emphasised the need for explainability: "In government environments, AI cannot operate as a black box... Decisions need to be transparent, auditable, and explainable" [1].

The Bias Question: Lessons from Robodebt and SyRI

The most pointed criticism of automated compliance systems comes from two international precedents that exposed how algorithmic pattern-matching can disproportionately harm vulnerable populations.

Between 2015 and 2019, Australia's Robodebt scheme auto-generated debt notices to roughly 470,000 welfare recipients by comparing annual tax averages against fortnightly payments — a method that systematically overstated income for anyone with irregular employment [15]. A Royal Commission concluded in 2023 that the scheme was unlawful, caused serious psychological harm, and drove some recipients to suicide [15].

In the Netherlands, the SyRI (Systeem Risico Indicatie) system profiled welfare recipients using 17 datasets, generating fraud-risk scores concentrated in lower-income urban neighbourhoods [16]. The District Court of The Hague ruled in February 2020 that SyRI violated Article 8(2) of the European Convention on Human Rights, citing a lack of transparency in how risk profiles were generated [16].

Academic research on AI in tax administration has flagged a structural problem: lower-income households and gig workers have irregular data footprints because informal economies generate fewer formal records, making them statistically more likely to be flagged as anomalous by pattern-recognition models [15][17]. This does not mean HMRC's system will replicate these failures, but the parallels warrant scrutiny — particularly given that HMRC has not publicly disclosed the false positive rate it considers acceptable, or whether the system has been independently audited for demographic bias.

Legal Framework: What Rights Do Taxpayers Have?

The legal landscape for automated decision-making in the UK shifted significantly with the Data (Use and Access) Act 2025 (DUAA), which received Royal Assent on 19 June 2025 [18]. The DUAA repealed Article 22 of the UK GDPR — which had broadly prohibited solely automated decisions with legal effects — and replaced it with new Articles 22A through 22D [18].

The new framework is more permissive. Organisations may now rely on legitimate interests as a lawful basis for automated decision-making, including where it produces legal or similarly significant effects [18]. However, safeguards remain mandatory: organisations must provide meaningful information about the decision, enable individuals to make representations, and offer access to human intervention and the ability to challenge outcomes [19].

For HMRC specifically, this means taxpayers flagged by the AI system retain the right to request human review of any adverse finding. The ICO has published draft updated guidance on the new Articles 22A-D, though it has not issued HMRC-specific guidance on the Quantexa deployment [19][20].

A Data Protection Impact Assessment (DPIA) is mandatory for this type of high-risk processing [20]. Whether HMRC has completed one, and whether the ICO has been formally consulted, has not been publicly confirmed.

What Happens When the Algorithm Flags You

Under HMRC's existing compliance framework, an investigation typically begins with a letter notifying the taxpayer that their return has been selected for a compliance check [21]. Civil investigations fall under Code of Practice 8 (suspected significant tax loss) or Code of Practice 9 (suspected fraud) [21].

There is no fixed timeline for investigations. A routine compliance check may conclude in weeks; a Code of Practice 9 fraud investigation can stretch for years [21][22]. Penalties depend on the taxpayer's behaviour: careless errors attract lower penalties (potentially reduced to zero with an unprompted disclosure), while deliberate inaccuracies with concealment can carry penalties up to 100% of the tax owed [22].

Taxpayers who disagree with a finding can request a statutory review by an HMRC review officer (typically completed within 45 days), and if dissatisfied, appeal to the First-tier Tribunal (Tax Chamber) within 30 days [23].

HMRC has stated that the AI system will not change this process — it will serve as a triage tool, directing human investigators toward the highest-risk cases [5]. The critical question is whether the pre-investigation phase changes in practice. If AI flags generate preliminary inquiries or informal contact before a formal compliance check is opened, taxpayers may face an opaque and unaccountable screening layer that precedes any formal legal process.

The Steelman Case Against AI Fraud Detection

The strongest argument against prioritising AI fraud detection is structural: the largest share of the tax gap does not stem from the kind of fraud that pattern-matching algorithms are designed to catch.

Nearly half the gap (46%) comes from failure to take reasonable care and errors — problems caused by a complex tax code and inadequate taxpayer support, not criminal behaviour [9]. Avoidance, at just 1% of the gap, primarily involves sophisticated corporate structures that exploit legal ambiguities — precisely the kind of arrangements that resist automated detection because they are technically lawful [9][11].

TaxWatch, an investigative think tank, has argued that HMRC's tax gap methodology likely underestimates large-scale corporate avoidance because it measures only known avoidance schemes, not those that remain undetected [10]. Small businesses account for 60% of the gap, but much of their non-compliance stems from confusion about obligations, not deliberate evasion [8].

The government has also committed to hiring 5,500 additional compliance caseworkers over the next five years [1]. Critics, including the Public Accounts Committee, have questioned whether the £175 million AI investment would generate a higher return if redirected toward human investigators with specialist expertise in corporate tax avoidance — an area where relationship-based investigation, access to legal expertise, and cross-border cooperation are more effective than algorithmic scoring [24].

HMRC itself targets a 20% increase in charging decisions for tax fraud by 2029-30 [2]. Whether that increase comes from better-targeted human investigations or from AI-driven volume remains an open question.

The Case for AI

Defenders of the approach point to the sheer scale of the problem. HMRC processes tens of millions of tax returns annually, and even a modest improvement in detection accuracy across that volume could recover hundreds of millions of pounds. The IRS experience — a 12% enforcement revenue increase with 25% fewer staff — suggests that well-deployed AI can generate significant returns [13].

Quantexa's entity resolution technology is specifically designed to connect fragmented data across systems, making it suited to identifying hidden relationships between entities — precisely the kind of analysis that human investigators find time-consuming [1][14]. If the system works as described, it could surface corporate networks, undisclosed income sources, and VAT fraud chains that would otherwise require months of manual cross-referencing.

HMRC received 93,000 complaints in 2024-25, up from 70,000 in 2020-21 — evidence that the agency is struggling with service quality under existing resources [1]. Automating routine compliance checks could free human caseworkers to focus on complex investigations.

What Remains Unknown

Several critical questions remain unanswered. HMRC has not disclosed: how many taxpayers it expects the AI system to flag in its first year of operation; what false positive rate it considers acceptable; whether the system has been audited for demographic bias; the specific procurement process used to award the contract; or whether a DPIA has been completed and shared with the ICO.

The ten-year contract term also raises lock-in concerns. If the system underperforms or produces discriminatory outcomes, unwinding a decade-long, £175 million technology partnership will be politically and practically difficult.

Britain's tax gap is real, growing, and costly. Whether an algorithm is the right tool to close it depends on answers that HMRC has not yet provided.