While small businesses debate whether to adopt AI, tax authorities have already deployed it. At scale. And the gap between government AI capability and taxpayer sophistication is growing.
Understanding what the other side of the table is doing is not optional. It is essential context for anyone who files a tax return.
The IRS: Agentforce and Beyond
The United States Internal Revenue Service has been one of the most public adopters of AI for tax enforcement. The agency's deployment of Salesforce's Agentforce platform to handle taxpayer interactions marked a visible step toward AI-augmented operations.
But the more significant investment is behind the scenes. The IRS has been building data analytics capabilities for years, using machine learning to score returns for audit potential, identify patterns of non-compliance, and detect fraudulent refund claims. The Inflation Reduction Act of 2022 provided approximately 80 billion dollars in additional funding, a significant portion earmarked for enforcement technology.
The IRS approach focuses on risk scoring: using algorithms to assess the likelihood that a particular return contains material errors or understated income. Returns that score above a threshold get routed for examination. The scoring models consider hundreds of variables: income patterns, deduction ratios, industry benchmarks, historical compliance, and relationships with other taxpayers and entities.
HMRC: The Connect System
Her Majesty's Revenue and Customs in the United Kingdom operates Connect, one of the most sophisticated data cross-referencing systems in the world.
Connect integrates data from an extraordinary range of sources: employer payroll submissions, bank and building society interest reports, property transactions from the Land Registry, company formation data from Companies House, VAT returns, self-assessment returns, and third-party data from payment processors and online platforms.
The system uses network analysis and machine learning to identify discrepancies. If your tax return declares income of 30,000 pounds but your bank received 60,000 pounds in deposits, Connect will flag it. If you report no rental income but the Land Registry shows you own three properties and letting agents report payments to you, Connect will notice.
HMRC has attributed billions in additional tax revenue to Connect since its deployment. The system processes over a billion data points and can identify patterns of non-compliance that would be invisible to human inspectors reviewing returns individually.
Malta: Digital Initiatives
Malta's Commissioner for Revenue (CfR) operates in a smaller market, but the trajectory is similar. The CfR has been investing in digital infrastructure, electronic filing, and data analytics capabilities.
Malta's size is, paradoxically, both an advantage and a limitation for AI-driven enforcement. The advantage: in a small economy, the total data set is manageable, and relationships between entities are more visible. The limitation: the investment budget for AI technology is proportionally smaller.
However, Malta's participation in EU-wide data sharing frameworks means that data from across the European Union flows into Maltese tax authorities. A Maltese resident receiving income from a German company will have that income reported to the CfR automatically, whether or not the individual declares it.
EU-Wide Data Sharing: The DAC Directives
The European Union's Directive on Administrative Cooperation (DAC) has been the quiet engine of cross-border tax transparency in Europe. Each successive version has expanded the scope of automatic information exchange:
DAC1 established the framework for automatic exchange of information on employment income, pensions, and certain other income categories.
DAC2 implemented the Common Reporting Standard (CRS), requiring financial institutions to report account holder information to tax authorities across participating countries.
DAC3 required automatic exchange of advance tax rulings and advance pricing arrangements.
DAC6 mandated reporting of cross-border tax arrangements by intermediaries (accountants, lawyers, financial advisers).
DAC7 extended reporting obligations to digital platform operators (Airbnb, Uber, Etsy, and similar platforms must report seller earnings to tax authorities).
DAC8 extends the framework to crypto-asset transactions.
The cumulative effect is a web of automatic data sharing that makes it increasingly difficult to earn income in one EU country without the tax authority in your country of residence knowing about it. And AI is what makes it possible to actually process and act on this volume of incoming data.
AI-Powered Risk Scoring for Audit Selection
Tax authorities cannot audit everyone. Even with increased budgets, the number of auditors is a fraction of the number of taxpayers. AI-powered risk scoring determines who gets audited.
The scoring models use supervised machine learning trained on historical audit outcomes. Returns that share characteristics with previously audited returns that yielded adjustments receive higher risk scores. The features typically include:
- Deviation from industry norms (a consultant claiming motor expenses three times the industry average).
- Year-over-year changes (income dropping 50% while living standards appear unchanged).
- Third-party data mismatches (reported income not matching data from employers, banks, or platforms).
- Network connections (transactions with entities previously found to be non-compliant).
- Behavioural indicators (late filing, late payment, frequent amendments).
These models are effective. They focus limited audit resources on the returns most likely to yield results, increasing both the efficiency and the revenue impact of enforcement activity.
Real-Time VAT Fraud Detection
VAT carousel fraud, where goods are traded in a circle across EU borders with VAT charged but never remitted, costs EU member states billions annually. The European Commission estimated the VAT gap (the difference between expected and actual VAT revenue) at over 60 billion euros across the EU.
AI is being deployed to detect carousel fraud in near real time. By analysing VAT transaction data across borders, ML models can identify the circular patterns, shell company involvement, and timing anomalies that characterise carousel schemes.
The EU's Transaction Network Analysis (TNA) tool, used by tax authorities across member states, applies network analysis to VAT data to identify suspicious trading chains. Individual countries have also developed their own systems; Belgium and the Netherlands have been particularly advanced in this area.
Pre-Populated Tax Returns: The Nordic Model
Several Nordic countries have effectively removed the burden of tax filing for the majority of their citizens. In Sweden, Denmark, Norway, and Finland, tax authorities pre-populate returns using data they already hold: employer payroll reports, bank interest statements, property records, and investment income reports.
The taxpayer receives a pre-completed return and simply confirms it is correct. If all data sources are accurate and the individual has no additional income or deductions to declare, the entire process takes minutes.
This model works because the tax authority has comprehensive data feeds and the analytical capability to process them. It represents the logical endpoint of the data exchange frameworks described above, combined with sufficient computing power to reconcile all the data for every taxpayer.
Malta is not there yet. But the infrastructure is being built, piece by piece.
What This Means for Taxpayers
The implications are straightforward and significant:
Your data is being cross-checked. Whether you know it or not, information about your income, assets, and transactions is flowing to tax authorities from multiple independent sources. AI is being used to reconcile this data and identify discrepancies.
Underreporting is increasingly detectable. The days when a small unreported income stream might go unnoticed are numbered. When banks, employers, platforms, and foreign tax authorities all report data independently, omissions stand out.
Accuracy is your best defence. If your tax return accurately reflects your income and legitimate deductions, then cross-referencing works in your favour. The data from third parties will match your return, your risk score will be low, and the probability of being selected for audit decreases.
Timeliness matters. Late filing and late payment are themselves risk indicators. Meeting deadlines consistently signals compliance.
The Asymmetry Problem
Here is the uncomfortable reality: governments are investing billions in AI-powered enforcement technology. The IRS alone is spending tens of billions on modernisation. HMRC processes over a billion data points. The EU's data sharing frameworks grow more comprehensive with each new directive.
Meanwhile, many self-employed professionals and small businesses are still managing their finances with spreadsheets, shoeboxes of receipts, and annual visits to an accountant.
This asymmetry is unsustainable. When the enforcement side has AI and the compliance side has a spreadsheet, the compliance side will lose. Not because they are doing anything wrong, but because disorganised records produce inaccurate returns, and inaccurate returns get flagged by AI-powered risk scoring systems.
The solution is not to fear the technology. It is to recognise that accurate, complete, and timely financial records are no longer a nice-to-have. They are the minimum standard for operating in an environment where tax authorities have the technology to check.
The businesses that embrace this reality and invest in proper record-keeping and modern accounting tools will find that government AI is not a threat. It is simply irrelevant, because their returns are already accurate.
The businesses that do not will find themselves explaining discrepancies to an auditor who was selected by a machine that already knows the answers.
Michael Cutajar, CPA — Founder of Accora.