The AI Tax Hunter: Inside Austria’s Predictive Tax Audit System

Austria’s tax authorities use AI to select significant part of tax audits. They’ve been doing it for years. Meanwhile, most tax advisors only started exploring AI when ChatGPT launched.

In this episode, Marina Luketina (Johannes Kepler University Linz) explains how predictive analytics works, why tax auditors are under more pressure than ever, and what the „black box problem” means for taxpayers.

We also dive into her research on Large Language Models for VAT – including a surprising finding: RAG outperformed fine-tuning with 90-100% accuracy.

Is AI replacing tax advisors? Marina says no – but it’s transforming the profession. The tasks will get harder. You’ll need deeper knowledge AND AI literacy.

Here you can find Marina’s articles that explores AI and its application in tax law:

https://pure.fh-ooe.at/en/persons/marina-luketina/publications/

https://arxiv.org/abs/2507.08468

This article summarizes a discussion with tax law and AI expert Marina Luketina and does not constitute a sourced academic paper. The analysis presented reflects the conversation’s key insights rather than comprehensive research.

Artificial Intelligence in Tax Law: Predictive Analytics, LLMs, and the Future of Tax Advisory

Introduction

This article summarizes a podcast conversation between Paweł Mikuła, partner and founder of „Halcyon | Tax. Customs. Legal.”, and Marina Luketina, a professor at the University of Applied Sciences of Upper Austria and associate at Johannes Kepler University in Linz. The discussion explores the interplay between artificial intelligence and tax law, covering both the use of AI by tax authorities and its applications in private tax advisory practice.

The Intersection of AI and Tax Law

Marina has been researching the intersection of AI and tax law for six years, starting when AI was not yet a mainstream topic. Her initial work included examining the taxation of robots replacing human workforce and its consequences for social and tax systems built on taxing human labor.

She emphasizes that AI in the tax field extends far beyond chatbots like ChatGPT. Tax law is inherently a data-rich field – financial authorities collect vast amounts of data, which serves as the fuel AI needs to function effectively. This creates opportunities for implementing various AI systems as genuine support tools for tax professionals.

Predictive Analytics in Tax Administration

The Austrian Model

Austria was one of the first EU member states to introduce predictive analytics in tax administration. Approximately ten years ago, Austria established a dedicated Predictive Analytics Competence Center – not merely a software tool, but an entire department combining IT experts, AI specialists, data scientists, and tax professionals. The department is led by an IT expert and continues to grow.

The scope of data collection is comprehensive. The authorities analyze not only data that companies and individuals provide for tax compliance purposes but also scan the entire worldwide web, including social media. The goal is to detect errors, discrepancies, potential tax fraud, and suspicious transactions.

How the System Works

Predictive analytics operates by comparing current data with historical data. Algorithms analyze patterns and generate predictions about potential tax issues based on probabilities derived from past cases. The system then flags suspicious situations for human review.

Marina provides a practical example: if a company has consistently reported VAT transactions with other EU countries for years and suddenly stops declaring any cross-border transactions, the algorithm will flag this change as potentially suspicious and recommend an audit.

Another example involves proportionality checks – if a company reports one million euros in wages but only 100,000 euros in wage tax, the algorithm will identify that the relationship between these figures appears incorrect.

Impact on Tax Audits

The results are significant. More than one-third of tax audits in Austria are now selected by the predictive analytics system rather than by human auditors. This represents a fundamental shift from past practices where all case selection was performed by humans.

However, the system creates new pressures. Tax auditors report being less satisfied with the new approach. Previously, they analyzed financial data themselves and had full transparency about why they were investigating a particular case. Now they receive indicators and hints from the algorithm without complete context. One auditor described the challenge: receiving instructions to check a company’s revenues due to an inconsistency, but struggling to explain to the company why the audit is being conducted.

The Human-Machine Feedback Loop

Marina emphasizes that the system depends on human-machine collaboration. The algorithm generates predictions, but human auditors must verify whether those predictions were accurate. This feedback loop is essential – without auditor feedback on whether predictions were correct, the machine could not learn and improve. The predictions become increasingly accurate over time through this iterative process.

The Black Box Problem

A significant challenge is the lack of transparency – the „black box” nature of these algorithms. Tax auditors cannot always understand or explain why the machine selected a particular company for audit. They receive risk scores and indicators, but not the complete reasoning. In Austria, there is a threshold: all suspicious transactions must be audited if the algorithm indicates a potential tax amount due of at least 10,000 euros.

VAT and Future Scenarios

VAT plays a particularly important role in predictive analytics for several reasons: it is a tax that corporations face daily, it represents a major financing source for states, and VAT chain transactions are notoriously susceptible to fraud.

The goal is moving toward real-time predictions and controls. Marina notes this is not far away, especially considering e-invoicing requirements and the ViDA (VAT in the Digital Age) regulations that EU member states must implement. The combination of e-invoicing data, payment information, and the new VIES system will provide authorities with comprehensive real-time information for analysis.

One key advantage of AI lies not just in analyzing data but in finding structures and relationships between data that humans might miss. When analyzing data from companies that trade together, algorithms can produce pictures of relationships that were previously invisible.

Data Privacy Concerns

Marina raises an important concern about taxpayer protection. Austrian tax law contains a provision stating that tax authorities may use algorithms and machines to collect and analyze data – but without clear boundaries. She questions where the protection for taxpayers lies when authorities have such broad powers.

AI Applications in the Private Sector

There is currently an imbalance between tax authorities and private sector practitioners. Authorities began implementing AI systems years ago, while most companies and tax advisors only started exploring AI three years ago when ChatGPT became popular. This creates a disadvantage for taxpayers who face AI-powered analysis by authorities but may lack equivalent tools for their own protection and preparation.

Fine-Tuning vs. RAG: Training AI for Tax Purposes

Marina’s research examined how to specialize Large Language Models (LLMs) for tax law applications, comparing two approaches:

Fine-tuning involves training the model on many examples of how cases were resolved, enabling it to predict solutions for future cases based on learned patterns. This approach requires significant time and resources.

RAG (Retrieval-Augmented Generation) involves providing the model with legal documents and sources, from which it generates answers. The model retrieves information from the provided documents rather than relying solely on its training.

Research Methodology

Marina’s team provided their model with:

• The Austrian Value Added Tax Act
• Official interpretations from the financial authority
• Textbook cases used for teaching students
• Real-world cases collected from practice, structured with questions and answers

They tested the model on determining the place of supply for goods and services – a core VAT concept.

Surprising Results

Contrary to prevailing research suggesting that fine-tuning was necessary for legal AI applications, Marina’s study found that RAG performed better than fine-tuning for their VAT use cases. This is significant because RAG is less resource-intensive – you can provide a model with documents and obtain accurate answers without extensive training.

The accuracy rates ranged between 90% and 100%. Marina notes that while some practitioners dismiss anything less than 100% accuracy, this perspective misses the point. A 90%+ accuracy rate often exceeds what a human tax advisor would score, especially one without 15 years of experience.

The Limitations of Current AI

Despite impressive results for straightforward questions, AI still struggles with nuanced situations. Paweł shares an example where he provided ChatGPT with extensive research on a VAT rate question. The AI failed to identify that recent binding rulings from October 2025 had changed the established approach – it drew conclusions based on the majority of documents indicating one rate, missing that a small number of recent documents represented a fundamental shift.

Marina explains this reflects the nature of AI: it operates on probabilities and mathematics, not understanding. AI lacks intuition about the world and human functioning. It cannot recognize when a recent development should override established patterns.

The Future of Tax Advisors

Marina does not see AI replacing tax advisors but rather transforming their work. Tasks will become more complicated. She cites a partner at a major Austrian tax advisory firm who noted that 20 years ago, she wrote statements about basic invoice requirements – questions no one asks anymore because they are too simple.

The future requires a combination of AI and human knowledge. Tax advisors will need to:

• Deepen their substantive knowledge
• Stay current with AI developments
• Understand which tools to apply and their limitations
• Assess every AI output critically

Human strengths – experience, intuition, understanding of context and relationships – remain essential. AI can handle calculations, statistics, and data structuring, but understanding the world and making judgment calls remains human territory.

Marina also notes a challenge for the profession: if AI handles entry-level tasks previously performed by junior staff, how do professionals develop the experience needed to reach senior levels? This represents a structural challenge the profession must address.

Hybrid AI: Combining LLMs with Logic

Marina sees significant potential in „Hybrid AI” – combining LLMs with logic-based systems. Her research includes:

Knowledge Graphs for Chain Transactions: Using LLMs to extract information from contracts and emails about VAT chain transactions, then producing knowledge graphs showing corporations involved, transactions, and identifying movable versus non-movable supplies. Interestingly, LLMs failed completely when trying to extract chain transaction information from court rulings due to unstructured data, but performed well with contracts and structured business documents.

Transfer Pricing Methods: Combining LLMs with Prolog (a logic programming language) to determine appropriate transfer pricing methods. The LLM collects and retrieves necessary data and performs risk analysis, while logic rules based on OECD and Austrian transfer pricing guidelines determine the appropriate method. This approach provides transparency through documented logic rules and can automatically generate required transfer pricing documentation. Marina received a research grant in December to expand this project with the University of Linz, University of Vienna, and PwC.

Conclusions

Marina offers key takeaways for tax professionals:

First, deepen your knowledge – it will be needed more than ever. Second, engage with AI actively. It is not just the future; it is the present, and it will become increasingly important and capable.

She cautions against focusing only on AI failures and problems. While risks exist, the advantages outweigh the disadvantages when AI is properly implemented as a support system rather than a replacement for human judgment.

Paweł adds an important observation: in Austria, AI has actually increased the need for human tax auditors by providing more information about potential issues requiring investigation. The relationship between AI and human work is not simply one of replacement.

The message is clear: AI and human expertise must work together. AI excels at processing data, finding patterns, and handling routine tasks. Humans provide judgment, intuition, context, and the ability to recognize when established patterns no longer apply. The future of tax advisory lies in mastering this collaboration.

---