Artificial intelligence and risk models in tax administration

Just a few years ago, talking about artificial intelligence seemed like a topic reserved for technology specialists. Today, the reality is very different, as it's increasingly common to use these tools to write documents, analyze information, translate texts, generate images, or automate tasks that previously required several hours of work. In other words, artificial intelligence has gone from being a futuristic innovation to becoming part of the daily lives of millions of people and businesses.

However, the use of this technology is not limited to the private sector. Governments around the world have also begun to incorporate it as a tool to improve the delivery of public services, optimize processes, and strengthen decision-making; tax administrations are no exception.

Indeed, in a context where tax authorities receive millions of data points daily from declarations, tax receipts (CFDI), foreign trade operations, financial information, international reports and various compliance obligations, the real challenge is no longer obtaining information, but transforming it into useful knowledge.

According to the Organisation for Economic Co-operation and Development (OECD), artificial intelligence has the potential to improve the efficiency of public institutions by analyzing large volumes of information, automating processes, and supporting decision-making.[1]

In this regard, the OECD makes a very relevant point: it is not just about making certain tasks faster, but about identifying patterns and relationships that, due to their complexity or volume, could hardly be detected through traditional human analysis.

In tax matters, this change has a particularly interesting consequence, since for many years, the exercise of verification powers depended, to a large extent, on relatively traditional selection processes, even manual reviews by the tax authorities.

However, the current international trend is moving towards oversight schemes based on massive data analysis and the development of models capable of identifying behaviors that could represent a higher risk of non-compliance. These systems are known as risk models.

However, it is important to note that a risk model is not a program that automatically determines how much tax a taxpayer must pay or that replaces the "review" work of the tax authorities.

On the contrary, its function is much simpler, but also much more powerful, since it analyzes huge amounts of information to detect inconsistencies, atypical behaviors or patterns that justify a more detailed review.

A very simple and easy-to-understand example could be a company that submits its declarations, notices, and reports on time, but the information contained in its CFDI is not consistent with other reports submitted to the authority or with information provided by third parties.

From a "formal" standpoint, previously, the tax authority had to individually review each document to identify inconsistencies. While this might seem easy, it became increasingly complicated with a large number of taxpayers. However, when all these documents are analyzed together using computer systems or AI, the inconsistency is identified almost immediately. This type of analysis is precisely what risk models allow to be performed much more efficiently.

This change represents an important evolution in the way we understand tax auditing, because for years, many companies focused their compliance efforts on addressing each obligation individually, such as filing returns, issuing tax receipts, keeping documentation or responding to information requests; however, that view is beginning to prove insufficient.

This is because, in an environment where information can be analyzed comprehensively, the consistency between the different data generated by a company becomes increasingly important; it is no longer enough for each obligation to be formally fulfilled; it is also essential that all the information is consistent with each other.

This means that tax compliance is no longer just a legal or accounting exercise, but also a data management exercise. The quality, integrity, and traceability of information will become increasingly important as tax authorities continue to incorporate more sophisticated technological tools.

Naturally, the use of artificial intelligence in the governmental sphere also poses significant challenges. As the OECD itself has pointed out, these tools must be developed and used under principles of transparency, human oversight, responsible risk management, and accountability, especially when they can influence decisions that affect citizens' rights.

In this sense, in tax matters, this means finding a balance between taking advantage of technology and respecting fundamental principles such as legal certainty, the proper justification and motivation of administrative acts and the protection of personal data.

Beyond the technology used, the trend seems clear: future oversight will depend less on random reviews and more on the ability of authorities to analyze information, identify risks, and allocate their resources more efficiently.

For companies, this scenario also represents an opportunity to strengthen internal processes, periodically verify the consistency of information, and adopt a preventive approach to tax compliance. This not only helps reduce risks from the authorities but also improves the quality of information used for decision-making within the organization itself.

Artificial intelligence will not replace legal knowledge or the judgment of tax authorities; however, everything indicates that it will transform the way in which risks are identified and decisions are made about where to start a review.

Therefore, understanding this evolution will allow companies to better prepare for an environment where information, more than ever, will become the main input for auditing.

[1] Organisation for Economic Co-operation and Development (OECD), Governing with Artificial Intelligence, OECD Publishing, Paris, 2025, chap. 5 (“AI in Tax Administration”).