Tax authorities adopt AI for tax fraud and efficiencies

New models of Artificial Intelligence (AI) and Machine Learning (ML) are being evaluated by global tax authorities to tackle fraud

As the world wrestles with the newest phase of AI – natural language tools based on reinforcement learning, such as Chat GPT – tax authorities have been quietly evaluating their uses to underpin their own objectives. To date this has evolved around 3 goals: fraud detection; improving the taxpayer experience; and efficiencies.

Watch VATCalc’s new AI VAT Advisor deliver tax advice (legislative references and tax cases) on complex VAT questions.

But authorities need to be cautious around deploying the considerable power of AI to such a financially and legally sensitive area as tax – as the shamed Dutch Tax and Customs Administration will attest to with the use of machine learning algorithms on claims. And don’t mention privacy risks, either, which could end-up strangling AI with a mess of regulations as we have already seen in Italy last month.

AI delivers on tax administrators 3 goals

To date, tax agencies have been focused in three aims:

1. Detecting tax fraud and errors. Tax authorities have been using predictive modelling in identifying flags for tax fraud or errors, with cloud-based AI able to review colossal volumes of data. It is proving particularly successful where special grants or seasonal tax deductions and payments are due, quickly identifying irregular funds flows or applications. Long standing Computer-assisted audit tools and techniques (‘CAATT’s’) used to detect international avoidance, and are heavily enhanced by AI.
2. Improving the taxpayers’ experiences. Tax administration is another arm of government which owes the public a duty of care and the fairest of outcomes. Aside from the efficiency gains of tax collected against the cost of the agencies, this can also mean faster and improved communications and outcomes for taxpayers.
3. Boosting internal operational efficiencies. The ability of AI and related technologies to extract and analysis trends from huge amounts of data is cutting down on the costly and unreliable use of manual interventions. And in addition to better accuracy, it can speed-up operations which means a better chance of identifying potential fraud or managing enquiries from the public or government. Typical areas where AI may be used to improve processes in tax administration include:
   • Neural learning for compliance obligations for VAT determination;
   • Prediction of errors and fraud to help target resource allocation
   • Sensitivity analysis on transfer pricing versus VAT
   • Natural language translations of tax laws and regulations to practical guidances
   • Q&A chatbots
   • Augmentative structures to understand patterns in court decisions

Examples of tax authorities using Artificial Intelligence

• UK HMRC is building Large Language Model to help identify VAT evasion
• Greece is establishing a specialist AI unit using MyDATA VAT transactions for detecting fraud.
• Austria AI unit has been targeting missing trader VAT fraud with impressive results.
• Romania claims to have boosted VAT receipts by up to 1% in the last 12 months via AI and robot adoption.
• Italy, possibly the most fervent users of AI for detecting tax evasion, last year identified over 1 million high-risk cases with AI-driven data analysis. This includes a latest algorithm that cross references financial data to identify taxpayers at risk of not paying. Its VeRa algorithm compares tax filings, earnings, property records, bank accounts and electronic payments looking for discrepancies. High-risk taxpayers then receive a letter asking them to explain the differences. The more data VeRa processes, the smarter it becomes.
• Vietnam has announced that it will be adopting Artificial Intelligence before the end of 2023 to help identify tax fraud. This includes, for example, to flag firms that issue invoices too often, for unusually high amounts or in other ways indicating attempts to slash taxable revenue.
• Australia claims to have identified over $530 million in unpaid tax bills and prevent $2.5 billion in fraudulent claims using AI models, including deep learning and natural language models.In addition to detecting underpayments, the ATO’s AI systems have also been utilized to combat GST fraud. This includes the ATO has employing gradient-boosting machine learning models, which have been successful in identifying fraudulent behavior patterns.
• US, the Inland Revenue Service has drawn-up a second-half 2023 plan to adopt Artificial Intelligence technologies and algorithms. It will be adding AI tools to identify taxpayers who make $1 million and up, and have more than $250,000  The IRS’ initial focus will be using AI analysis to replace existing paper-based reporting and returns. Its existing Modernized e-File (MeF) system already accepts 76% of paper tax returns processed without human intervention. This next phase will be about experimenting and adopting AI models to extract valuable information from this exercise.
• India from May 2023 is using AI to identify fraudulent applications for input tax credits via false GST registrations. The central government’s Business Intelligence and Fraud Analyst (BIFA) site, the e-way portal, and the Rajasthan government’s Business Intelligence Unit (BIU) would collaborate to detect GST numbers that appear to be false.
• India‘s Income Tax Department is using AI to identify falsified income tax deductions. It uses algorithms designed to identify unusual ratio’s between income and political or charitable donations.
• Malta, UK, Canada, the Netherlands and Ireland use an AI system that daily compares wealth based on public sources with that declared in their VAT and tax returns. It also sources public registers, bank accounts (in limited circumstances) to identify undeclared assets and spending.
• Sweden deploys AI to identify and highlight tax risk issues when businesses apply for new incorporations. Since 2021, it has been able to review for tax avoidance flags in registration applications. This has also helped speed-up the application process by reducing the manual time required in reviewing documentation.
• Poland’s System Teleinformatyczny Izby Rozliczeniowej (STIR), analysis data provided daily by banks and credit unions report account data and clearinghouse data on a daily basis for all transactions carried out by taxpayers. It enables the National Revenue Administration (NRA) to detect potential carousel frauds in near-real-time, versus the two months that might have been needed previously.
• France uses AI satellite image scanning to identify signs conspicuous consumption. This can include multiple cars or swimming pools appearing at residents of person under tax investigation. This is particularly useful for local direct taxes (real estate tax).
• Singapore’s Inland Revenue Authority has developed an in-house network visualiser with graph database as an underlying technology to address its auditors’ needs. This tool provides auditors with customised functionalities to analyse intricate, multi-layered relationships between entities during audits/investigations. It can also uncover relationships more than 10 connections deep in a real-time manner.
• Xenon is a tool used by six European countries to investigate tax evasion based on internet searches and surveillance. It was originally developed by The Netherlands
• Brazil has been using AI behavioural insights (called ‘High Performance Inspection’ (FAPE)) to analyse the outcomes of varying standard tax letter requests to taxpayers.  It has been evaluating the response of the taxpayer based on their particular background and circumstance to determine the optimum tax communications tone and lever of affirmation. From this, the authorities are able to determine the best approach to take with future taxpayer queries or audits.
• Aside from using AI to detect potential tax fraud or errors, most authorities are now using AI to assist the efficiency of their own compliance and administrative activities. This can include recruitment processes. Countries such as Canada and Singapore are leading the way on this.
• As common on must large private sites, the tax authorities are increasingly using AI-driven virtual assistants. The list of countries includes: Spain, Peru, Australia, Canada, the United Kingdom, Ireland, Finland, Sweden, Latvia, Estonia, the Republic of China, Russia, Singapore, Guatemala, Chile, Mexico, Costa Rica, Colombia and Brazil.

AI black box in Dutch national scandal – kinderopvangtoeslagaffaire

Whilst AI promises many attractions to tax authorities seeking better revenue and service outcomes, they must remain weary of its considerable powers. The Dutch government of 2021 was disgraced by the tax authorities wrongly ‘hunting down’ thousands of families on their family credits.

These credit claims processed by the tax office on self-learning algorithm. In the tax authority’s workflow, the algorithm would first vet claims for signs of fraud, and humans would analyse those claims it flagged as high risk. But the algorithm developed a pattern of falsely labeling claims as fraudulent, and harried civil servants rubber-stamped the fraud labels. So, for years, the tax authority baselessly ordered thousands of families to pay back their claims, pushing many into onerous debt and destroying lives in the process.