Governing with Artificial Intelligence

AI in tax administration

For many years, tax administrations have been using AI to support activities across their operating model and have been actively exploring its potential to further enhance their operations, improve taxpayer services, increase tax compliance and prevent tax fraud. They are well placed to do this; collecting and analysing data are at the heart of many tax administration processes, which has facilitated the early adoption of rules-based AI systems. These systems have allowed tax administrations to analyse and extract insights from large volumes of data, enabling quicker identification of non-compliance and more precise targeting of limited resources towards high-risk cases.

This existing experience underpins investigations into how more advanced techniques can be deployed in a tax administration context. While technological advancements in AI offer significant potential to enhance and transform services, they can also amplify risks related to sensitive data processing and introduce new concerns. Therefore, tax administrations need to place a strong emphasis on privacy, security and the trustworthy application of AI.

Current state of play

Since 2022, the OECD Inventory of Tax Technology Initiatives (ITTI) survey has provided insight into how AI is being applied across tax administrations around the world. The latest survey shows that the main areas of application in OECD countries are detection of tax evasion and fraud, decision-making assistance and improving tax services (Figure 5.1).

Figure 5.1. AI deployments across OECD Members who use AI in tax administration

Note: 29 of the 38 OECD members report using AI in tax administration in the 2024 Inventory of Tax Technology Initiatives.

Source: OECD Data Explorer - Inventory of Tax Technology Initiatives 2024 (https://oe.cd/dx/ITTI2024).

Improving compliance and detection of evasion and fraud

To improve compliance and detect evasion and fraud, AI is frequently used to uncover hidden patterns of behaviour or new connections between transactions, assets or taxpayers within the data sources that a tax administration might already hold. As technological capabilities and internal expertise has grown, AI-driven techniques are now being used to analyse unstructured data sets, such as text from handwritten documents and information from public posts on social media, to further add depth to the analysis and uncover connections about activities that could indicate tax evasion or non-compliance. These techniques are also used to support the detection or prevention of fraudulent attacks on the tax system by criminals who can use AI to create false claims for refunds on a potentially very large scale. As one example, Greece’s Independent Authority for Public Revenue (IAPR) is leveraging AI to combat tax evasion by detecting compliance issues, automating complex procedures and enabling auditors to respond in real time.1

More recent use cases include applying pattern-detecting capabilities of tax administrations’ AI systems to data from imaging systems to allow tax administrations to analyse maps and satellite imagery. By comparing satellite images over time, AI is being used, for example, to identify property alterations or new buildings or taxable assets that may not have been declared to the tax authority (Box 5.1). Greece has also developed an AI model for geo-locating swimming pools that have not been declared for tax obligations.2

Box 5.1. Detecting undeclared property development in France

To improve the process of detecting undeclared constructions or developments, the French tax administration (DGFiP) uses artificial intelligence and data enhancement based on aerial photographs taken by the National Institute of Geographic and Forestry Information (Institut national de l'information géographique et forestière, or IGN) as part of the Innovative Land (Foncier Innovant) project.

The algorithms make it possible to extract the outlines of buildings and swimming pools from the IGN’s public aerial images, which can be consulted on the website geoportail.gouv.fr.

A computer process then verifies, from the owners’ declarations to the DGFiP, whether the elements detected on the images are correctly taxed (for property taxes, in particular). A DGFiP official then systematically verifies each anomaly detected before any action is taken to remind or ultimately tax the owner of the property.

Source: (OECD, 2022[1]).

Assisting administrative decision-making processes

To make administrative processes more efficient, tax administrations are deploying AI to automate their internal processes, such as helping categorise and distribute similar cases or tasks. Similarly, AI is commonly used to differentiate simple, standardised cases, which can be resolved through automation, from complex cases, such as tax disputes, where the expertise and judgement of tax officials are most needed. This application allows a more efficient use of human and technological resources (Box 5.2).

Box 5.2. Improving decision-making in Brazil

Brazil had around USD 140 billion in assessed taxes waiting for decisions in administrative court tax appeals. It takes about six years for the appeal ruling. Under the AI Litigation Project, Brazil employed supervised machine learning (ML) when distributing groups of similar files to the same officers, in order to increase their speed in administering the file and taking decisions. The first trials, conducted with a sample of 2 000 manually labelled files, showed that supervised algorithms can attain sensitivity and specificity of over 80%.

Brazil also employed clustering algorithms to complete files either in full or in part. Additionally, a web-based report assistant tool is being developed to support officers' analysis and help in their goal of reusing blocks of text. The tool’s resources include the presentation of suggested groups of files and paragraphs and the highlighting of sentences that turned out to be important for the clustering process. Officers can label files and paragraphs, and the labels are used to improve future suggestions.

Source: (OECD, 2022[1]).

Enhancing risk assessment processes

A significant driver of tax administrations’ embracing AI is its ability to analyse data to score and prioritise risks. With AI, tax administrations can process large volumes of taxpayer data — from historical filings to transaction records and digital payment information — to build sophisticated risk models and then assign risk scores to specific types of behaviour or transactions. These models help tax administrations quickly identify possible non-compliance and focus resources on them. As the predictive capabilities of these systems has increased, tax administrations can now to identify potential issues more quickly, prevent them from escalating, and better support taxpayers in getting their filings right. To ensure reliability of these risk models, it is essential that the training data is accurate. Box 5.3 provides an example in Austria. As another example, in the Greek fuel market, IAPR has developed an analysis system that uses data from various sources to compose over 100 different risk analysis criteria.3

Box 5.3. Improving compliance checks with predictive analytics in Austria

Since 2014, the Austrian tax administration has applied machine‐learning algorithms through the Predictive Analytics Competence Centre (PACC), a specialised unit within the Federal Ministry of Finance. Tasked with modernising risk management, the PACC works to improve tax collection, auditing and fraud detection. Organised into four subject areas (Predictive Analytics, Advanced Analytics, Tax Analytics, and Customs Analytics), the PACC addresses a broad range of challenges across the tax system.

In 2023, the PACC’s risk models analysed around 6.5 million cases across income, corporate and value added tax sectors as well as customs transactions. The analyses detected instances of false reporting in employee tax assessments and identified fraudulent activities, resulting in additional tax revenues of approximately EUR 185 million. Nearly 27.5 million cases were also examined for compliance, with 375 000 cases flagged for further review due to implausible risk profiles. Advanced techniques, including decision trees, regression models and text mining, support both retrospective and real‐time audits while ongoing projects seek to expand analytical capabilities to include, for instance, generative models.

Source: https://www.bmf.gv.at/en/press/press-releases/2024--New/August-2024/BMF-generated-around-EUR-185-million-in-tax-income-from-AI-in-2023-.html.

Improving taxpayer services

Helping taxpayers meet their obligations is a long-standing activity of tax administrations that often uses labour intensive service delivery methods, such as call centres or visits at physical tax offices. To help taxpayers faster with simple enquiries or processes and to free up resources, tax administrations have for many years adopted AI-based virtual assistant technology. Basic assistants use rules-based techniques that can direct taxpayers to relevant information, make payments or check their status. According to recent ITTI data, virtual assistants that follow a set of pre-programmed rules when interacting with taxpayers are a very common use case for AI.4 More recently, tax administrations have begun using more advanced AI techniques, such as large language models (LLMs) to enhance and personalise interactions with taxpayers. The example in Box 5.4 illustrates the range of services that are possible, and the growing levels of sophistication that allow more personalised responses. In another example, Greece’s IAPR is developing an LLM-driven digital assistant to provide advisory support to taxpayers by answering questions during the submission of income tax returns.5

Box 5.4. Enhanced taxpayer services in Singapore

The Inland Revenue Authority of Singapore (IRAS) has employed end-user automation (i.e. tools that simplify things for service users), data and AI tools to deliver seamless and personalised taxpayer services. In 2021, IRAS launched a chatbot to more efficiently handle common queries on individual income tax, corporate tax, goods and services tax, property tax, stamp duty, withholding tax and employer tax matters. The chatbot is powered by the Singapore Government Virtual Intelligent Chat Assistant (VICA) platform.

The IRAS VICA chatbot enhances user engagement through a humanised user experience, featuring interactive elements such as carousels and infographics. In 2023, the chatbot was upgraded with the integration of a Large Language Model (LLM) engine, significantly improving its ability to understand the intent of taxpayers' queries.

Beyond answering queries, the chatbot also enables access to authenticated services. Taxpayers can conveniently check their outstanding tax balance, view payment plans, check payment status, cancel or reinstate payment plans and make tax payments via self-service payment terminals and mobile channels. These bot transactions save taxpayers about 10 minutes per transaction compared to traditional digital service channels. In the financial year 2024, the IRAS VICA chatbot handled about 70 000 transactional queries, potentially saving an estimated 11 666 taxpayer hours.

Source: (OECD, 2022[1]), https://www.iras.gov.sg/digital-services/others/iras-bot, https://www.tech.gov.sg/products-and-services/for-government-agencies/productivity-and-marketing/vica, Government of Singapore officials.

Assisting taxpayers in filing

Additionally, AI has facilitated the extensive pre-population of tax returns; algorithms and other tools examine tax administrations’ extensive data sets and then place the data in the correct part of the tax returns. This can improve the accuracy of tax returns and simplify the filing processes, reducing the time taxpayers spend on filing their returns. In instances where taxpayers complete their returns, tax administrations are deploying pattern detection techniques that can find anomalies and errors and can prompt taxpayers to double check their returns.

Box 5.5. Preventing errors in Australian tax returns

The Australian Taxation Office (ATO) is employing advanced technologies to enhance the accuracy and efficiency of tax return submissions. These include real-time analytics, pre-filled forms and anomaly detection systems to assist taxpayers in meeting their obligations and reducing errors.

As part of its efforts, the ATO provides a pre-filling service that automatically populates individual tax returns with data sourced from employers, banks, government agencies and other third parties. This includes information such as salary, bank interest, dividends and private health insurance details. Taxpayers are required to review and confirm the pre-filled data before submission. This system not only reduces administrative burdens but also improves accuracy by minimising manual entry errors.

Additionally, the ATO uses real-time prompts during the submission process to address potential anomalies. For instance, if a taxpayer’s reported figures are flagged by a model as abnormal, , the system generates a message encouraging them to double-check their inputs. A message might state, for example: “You have not reported any interest income. Please review and ensure any interest earned has been reported.” In 2023-24, more than 636 000 prompts were issued to individuals, helping to protect approximately AUD 78.9 million in revenue.

Source: (OECD, 2022[1]), https://aws.amazon.com/institute/future_of_tax_technology_ai_ml_cloud_security, https://www.austlii.edu.au/au/journals/JCULawRw/2024/5.pdf, https://www.ato.gov.au/individuals-and-families/your-tax-return/how-to-lodge-your-tax-return/lodge-your-tax-return-online-with-mytax/pre-filling-your-online-tax-return.

Managing risks and challenges

Whilst all deployments of AI in public administrations are generally accompanied by challenges, tax administration makes a particularly demanding environment for AI integration because of three critical factors:

1. Ensuring data protection and privacy: Tax administrations have access to a vast amount of sensitive taxpayer data and have the obligation to protect the privacy and confidentiality of this data. Governance and legal frameworks therefore need to be adapted to account for the use of AI.

2. Ensuring taxpayer rights: Taxpayers’ ability to challenge tax decisions made by tax authorities mean that decision-making relying on AI systems and associated administrative processes need to be explainable, transparent and accountable.

3. Maintaining trust: Tax administrations rely heavily on taxpayers’ voluntary compliance, which requires sustained trust in the impartiality and fairness of administrative processes. Taxpayers need to be able to trust the government’s use of AI systems, continue to believe tax systems are fair and remain willing to comply with their tax obligations.

Associated risks

• Inadequate or skewed data in AI systems

• Lack of transparency and explainability

If AI systems rely upon inadequate or skewed data, it could lead to inaccurate or adverse outcomes for some individuals or groups. With regard to tax administration, this could result in inaccurate risk assessments and thereby the improper targeting of some individuals or groups for control measures (e.g. audits). See Box 5.6 for an example.

When adopting AI in tax administration, inadequate and poor data quality poses a significant risk to effective outcomes, which can undermine public trust. For AI systems to return reliable and trustworthy outcomes, the input dataset needs to be accurate, complete, and well-structured. Discrepancies in data formats, incomplete or mistyped records, outdated taxpayer data or data lacking the necessary context can lead to inaccurate predictions, flawed risk models and biased outcomes. Some machine learning (ML) systems do not require data to be as structured, but quality and accuracy are still critical. Incorrect inputs can make systems less efficient, undermine taxpayers’ trust in government and have adverse consequences for taxpayers.

Box 5.6. Challenges with algorithmic errors in tax administration

The “Toeslagenaffaire” was a child benefits scandal in the Netherlands. Between 2005 and 2019, the Dutch Tax and Customs Administration (Belastingdienst) wrongfully accused approximately 26 000 families of fraudulently claiming childcare benefits, due to the use of flawed data and a skewed algorithm. Many of these families were flagged due to administrative errors, such as missing signatures, and disproportionately targeted families with dual nationality or migrant backgrounds. These families were forced to repay tens of thousands of euros in benefits, plunging them into severe financial hardship, with some losing their homes, jobs and marriages. In extreme cases, children were removed from their families. The ensuing scandal resulted in the collapse of the government.

Following the Toeslagenaffaire, the Dutch Ministry of the Interior and Kingdom Relations tasked researchers at Utrecht University to develop Human Rights and Algorithms Impact Assessment (IAMA). In 2022 the Dutch House of Representatives approved a motion to “make it mandatory to conduct this impact assessment before using algorithms when algorithms are used to make evaluations or decisions about people.”

Source: (OECD, 2023[2]).

Challenges regarding the transparency, explainability and interpretability of complex AI systems can pose risks to rule of law and a taxpayer’s legal recourse. The ability to challenge decisions is a fundamental component of tax administration processes. AI’s “black box” nature for many ML systems can create a lack of transparency, explainability and interpretability that can impact on taxpayers’ rights to understand how tax decisions are made, which in turn can limit their rights to dispute a decision.

To mitigate these risks, tax authorities need to invest in governance frameworks. These frameworks should ensure that AI is deployed in a way that respects the balance between the administration’s tax collection and taxpayers’ rights. Therefore, AI systems and associated operational processes in tax administrations need to be sufficiently transparent and explainable and operate with proper oversight and accountability mechanisms. These can include, for example, data governance (see Chapter 4, section on “Creating a strong data foundation”), including in terms of data cleaning, validation processes and regular updates to maintain the integrity and trustworthiness of the data used to inform AI-driven decision-making. Only with high-quality, reliable data can AI truly enhance tax administration by improving accuracy, compliance and operational efficiency for taxpayers.

Implementation challenges

• Skills gaps

• High costs of AI adoption and scaling

• Inflexible or outdated legal and regulatory environments

Taking advantage of the full range of opportunities that AI can bring to tax administration highlights three key challenges.

First, tax administrations report fierce competition for skilled employees who can develop and implement these technologies. Accompanying this is the requirement to train large numbers of existing employees in the use of this new technology and managing the associated changes in process.

Second, there are often significant investment costs associated with the deployment of this technology. The process for securing the budgets can be challenging, as the return on the investment may by uncertain or over the long term. As well as helping to provide new services to taxpayers, this investment is also necessary to help defend the tax system against fraudulent attacks.

Finally, AI has the potential to disrupt existing legal frameworks that support many tax administration processes. Adapting these frameworks needs to be considered in parallel to the guidelines for ensuring the effective design of and deployment of AI. It is also worth noting that whilst countries may have general guidelines for AI deployment, they are often not tailored to the unique needs and specific complexities of tax administration, which needs further consideration.

Untapped potential and way forward

Tax administrations across the globe have already begun leveraging the potential of AI, deploying it in wide range of activities to improve the efficiency of their operations. AI holds to the potential not only to optimise tax administrations’ existing operational models but also to transform them. OECD’s (2020[3]) Tax Administration 3.0 Framework sets out a vision for digital transformation where digital technology is used to create seamless, event-based tax ecosystems. These allow for tax collection and reporting to be built directly into the natural systems of individual and business taxpayers.

To fully realise the transformative potential of AI, tax administrations should adopt a systemic approach to AI application. AI is not a standalone tool; it should be used an integrated component embedded in a broader tax ecosystem, encompassing regulatory frameworks, organisational structures, existing technical infrastructures and human expertise. It is crucial that tax administrations take into account the need for adjusting the rule-making process. For example, applying Rules as Code (RaC) approaches, tax administrations can encode tax laws and regulations into machine-readable formats, enabling AI systems to interpret and apply rules with greater accuracy and consistency (Mohun and Roberts, 2020[4]).6

Given that taxpayer trust is an essential part of a well-functioning tax administration, tax administrations should also ensure that AI deployed is trustworthy, transparent and accountable manner. To support tax administrations in this effort, the OECD Forum for Tax Administration is currently piloting a framework to support tax administrations in their deployment of AI systems. Building on the OECD principles for trustworthy AI, the framework outlines key considerations for each stage of the AI lifecycle, describing learning and experiences from a range of administrations from their own implementations. Through this ongoing co-operation, tax administrations can adopt a structured, risk-based approach to enhancing the use of AI in tax administration, which can also adapt to the changing wider context. The Forum on Tax Administration plans to publish its findings from this work in 2026.

Notes

← 1.  https://www.aade.gr/sites/default/files/2024-12/dt_.11.12.24.pdf.

← 2.  https://www.tovima.gr/2023/05/29/finance/aade-pos-i-texniti-noimosyni-apokalyptei-tis-adilotes-pisines.

← 3.  https://www.aade.gr/sites/default/files/2024-12.

← 4.  https://www.oecd.org/en/data/datasets/inventory-of-tax-technology-initiatives.html

← 5.  https://www.mononews.gr/oikonomia/aade-sta-skaria-chatbox-gia-tis-forologikes-diloseis.

← 6.  Rules as Code (RaC) is the practice of translating legal rules and regulations into machine-readable formats (Mohun and Roberts, 2020[4]). This approach enables automated compliance checking and enforcement, and supports decision-making based on predefined legal frameworks. When applying AI in tax administration, RaC can help ensure that AI supported decision making is transparent, consistent across cases, and legally compliant. RaC can thereby help increasing public trust and accountability in AI application.