Machine learning to spot likely VAT evasion across multiple data sources
Tax authorities adopting AI has been underway for several years. But what about the UK?
For over a decade, HMRC has used advanced analytics through its Connect system to identify potential cases of wider tax evasion and avoidance. But this is increasingly extending to VAT avoiders. The UK VAT gap is estimated to be to be 4.9% of theoretical VAT liability, or £8.1 billion in absolute terms, in the 2022 to 2023 tax year. It is stepping-up use of Large Language Models (LLMs), both in-house and acquired technologies.
HMRC does publish details of its AI uses in any great depth. But it is clear, like other jurisdictions, AI technologies are leveraged extensively to detect patterns of VAT fraud by analysing mass datasets derived from VAT returns submitted by taxpayers, thereby enabling the identification of fraudulent activity with greater precision.
You can see how we at VATCalc are already using it to generate AI VAT advice and item classification or tax codes to get the right taxing rates, country, exemptions and return reporting.
AI is primarily deployed by the UK tax authorities for compliance risk assessment across VAT and other tax revenues. These AI-driven systems are instrumental in risk management processes, evaluating the risk profile of taxpayers upon filing VAT returns or submitting repayment claims, and supporting case selection for compliance investigations.
The Risk and Intelligence Service (RIS), a specialized unit within HMRC’s Compliance team, is understood to spearhead the development and implementation of these AI technologies. The RIS team collaborates with HMRC’s in-house data scientists and analysts, while the Chief Digital Information Office – Data Science (CDIO – Data Science) division also provides expertise to assist various HMRC departments in developing advanced AI-driven solutions aimed at enhancing tax compliance and risk management.
Watch VATCalc’s new AI VAT Advisor deliver tax advice (legislative references and tax cases) on complex VAT questions.
Adopting Large Language Models for VAT
HMRC is further advancing its AI capabilities through the development of a large language models (LLM). Upon completion, the LLM is expected to equip HMRC officers with real-time, AI-generated responses to taxpayer inquiries, while referencing relevant sections of VAT guidance. Although it is not yet confirmed whether the LLM will include a taxpayer-facing interface, it is plausible that such functionality could be introduced in the future to streamline taxpayer queries traditionally addressed through HMRC manuals or helplines.
While HMRC possesses internal capabilities for developing AI technologies, it is also understood to source technology services from third-party vendors to support system enhancement efforts.
Tax data for training AI
Regarding the data used to train these AI models, HMRC primarily relies on taxpayer data collected through its systems. In addition, HMRC may procure supplementary data from external sources, such as merchant acquirers. It remains uncertain whether HMRC uses pre-labeled data from these external providers or undertakes its own data labeling, particularly for supervised learning models.
