1. Introduction: A new paradigm of Transfer Pricing control? 

The Tax Administrations (TAs) are going through a process of unprecedented technological transformation, aimed at the “3.0” conception promoted by the OECD. In the context of increasing digitalization of the economy, internationalization of business and sophistication of tax planning schemes, the control of Transfer Pricing (TP) continues to be one of the most complex and challenging areas of tax control. 

Traditionally, this control has been conditioned by structural factors, such as the asymmetry of information between Taxpayers and Administrations, and the high technical and operational burden involved in comparability analyses. However, in recent years there has been a qualitative evolution in the use of artificial intelligence (AI) applied to tax management. 

We are moving from a first stage of predictive and analytical models, mainly oriented to the segmentation and prioritization of risks, to a new paradigm based on “agentic” AI systems, capable not only of analyzing information, but also of reasoning, interacting with multiple regulatory sources and executing complex end-to-end control processes, always under human supervision. 

This change is not incremental. It implies a redefinition of the way in which TAs conceive, design, and execute control processes in the field of TP. 

 2. The bottleneck of the traditional model: why is manual monitoring unsustainable? 

The conventional TP analysis presents structural limitations that are difficult to overcome exclusively with the current availability of affected human resources. A complete TP audit can take, on average, between 6 months and up to more than a year of specialized work per case, without considering recursive or litigious instances. Everything will depend on the complexity of the transaction and the amount and quality of data available. 

This reality generates clear operational consequences: 

• • Limited control capacity, which forces TAs to focus on a small number of high-risk taxpayers. 

• • Partial effectiveness, with modest – and even unknown – detection rates of cases with relevant deviations. 

• • Technical inconsistencies arising from differences of criteria between teams or auditors. 

• • Operational obsolescence, due to the difficulty of incorporating regulatory changes, new international guidelines, and recent case law in real time. 

The OECD has pointed out that the strategic use of AI in TAs allows analyzing large volumes of data with a depth and speed unattainable by manual methods, also enhancing the value of the information obtained, through the growing international data exchange (OECD, 2025). 

For its part, CIAT has recently made a significant contribution to this issue with the publication of its “Study on Artificial Intelligence Applied to Transfer Pricing,” an enlightening document that analyzes how new technologies can optimize the taxation of international transactions.  

 3. What is really a tax AI agent? Beyond the “traditional” chatbot 

A common mistake is to equate AI agents with simple conversational assistants. In reality, a smart tax agent is a modular system that combines cognitive, regulatory, and operational capabilities1. Its typical architecture integrates these essential components: 

1. The cognitive nucleus (language model)

Large-scale language models (LLM) such as GPT or Claude – provide semantic understanding, contextual reasoning, and the ability to operate on large and complex texts such as those addressed by TP. 

2. The Normative memory (RAG)

Using Recovery Augmented Generation (RAG) techniques, the Agent accesses a dynamic document database that includes national legislation, OECD guidelines, national regulations, administrative resolutions, and relevant case law. This allows operating on a correct applicable regulation. 

3. The reasoning and planning engine

Through approaches such as ReAct (Reason + Act), the agent can decompose a complex problem into a logical sequence of actions: identify transactions, select methods, search for comparables, execute calculations and validate results. 

4. The operational tools

The agent connects to commercial databases (for example, Orbis or Amadeus), statistical engines, spreadsheets, and programming languages (such as Python), guaranteeing technical accuracy and traceability in the results. 

The adoption of locally deployed AI models, eventually adjusted to one’s own tax domains, allows – together with appropriate governance frameworks – strengthen control over data subject to tax secrecy and, in high-demand scenarios, reduce long-term operating costs. 

 4. The workflow of an AI agent in Transfer Pricing: an illustrative case 

To understand its potential, we are going to propose a hypothetical audit of a local subsidiary dedicated to manufacturing or distribution activities. 

Phase 1: Information intake and strategy 

The IA Agent automatically processes contracts, accounting statements, and tax returns, including three-level reports (master file, local file, and BEPS Action 13 CbC report) and other international cooperation reports (CRS and soon CARF, rulings, etc.). It classifies the transactions (goods, services, intangibles) and links each transaction with the applicable national regulations. 

Phase 2: Functional analysis (Functions, Assets and Risks) and selection of the method 

Based on the analysis of functions, assets and risks, the system can characterize the entity — for example, as a limited risk distributor— and suggest the most appropriate method, such as the Transactional Net Margin Method (TNMM). 

Phase 3: Search for comparables and determination of the range 

The Agent filters comparable companies according to sectoral and geographical criteria, calculates relevant financial indicators and determines the arm’s length interquartile range. 

If the taxpayer’s margin is outside this range, the system can automatically estimate the potential adjustment, documenting each step of the calculation. 

Multi-agent ecosystems: towards a collaborative and intelligent control 

Contemporary international control exceeds the capacity of a single isolated system. Therefore, the most advanced architectures adopt multi-agent ecosystems, where different specialized agents collaborate under a coordinated logic, for example: 

• • Coordinating agent, which manages times, priorities, and workflows. 

• • Network analysis agent, aimed at identifying complex corporate structures and risks of erosion of the tax base. 

• • Expert agent in TP, in charge of technical and economic analysis. 

• • Legal agent, who validates the legal and jurisprudential consistency of each conclusion. 

This approach makes it possible to scale up the supervision without losing specialization or control.  

Even, under the appropriate regulatory safeguards, the intervention of TAs from two or more countries could even be considered, to facilitate the detection of international evasion structures through the intelligent exchange of information and synchronizing audits through agents specialized in corporate network analysis and international legal validation. 

Explainability and traceability: condition of validity  

In the field of TP, the explainability of AI systems for the technical and legal validity of their results is crucial. 

AI agents applied to TP must be able to document and reconstruct every relevant decision: from functional characterization and method selection to the choice of comparables and the determination of adjustments. 

Unlike traditional opaque models, agentic approaches allow incorporating regulatory, economic, and procedural traceability, facilitating human supervision and respect for the due tax process. 

Governance, ethics, and the irreplaceable role of the human factor 

The incorporation of AI agents does not imply, in any way, the replacement of the human auditor. On the contrary, their function is revalued, moving from repetitive tasks to validation, criterion, and decision activities. 

The OECD (2025) stresses the need to establish governance, transparency, and accountability frameworks, ensuring the permanent presence of the human in the cycle (human-in-the-loop), especially in the area of TP in critical instances such as: 

• • The functional characterization. 

• • The selection of the method. 

• • The determination of the final adjustment. 

Likewise, the risk of errors or “hallucinations” typical of LLMs are mitigated – in addition to the use of RAG – through the application of deterministic tools: the Agent does not estimate or invent results, but executes verifiable, traceable, and auditable calculations. 

As a conclusion of a topic open for discussion: 

The adoption of AI Agents in the control of TP offers the TAs the possibility of evolving from selective and reactive models towards preventive, massive and technically consistent control schemes. 

AI is no longer limited to reading documents or classifying risks. Today it can reason, execute, and actively assist in the protection of the tax base, always under the strategic guidance of the human factor in critical issues. 

In this new scenario, the true competitive advantage of TA will not lie solely in the technology available, but in their ability to responsibly integrate it into their processes, their regulatory framework, and their organizational culture.

Bibliographic sources consulted. 

Inter-American Center of tax administrations (2025). Study on Artificial Intelligence applied to transfer pricing (DT-06-2025). https://www.ciat.org/dt-06-2025-estudio-sobre-inteligencia-artificial-aplicada-a-los-precios-de-transferencia/   

OECD. (2026). Portal Transfer Pricing. https://www.oecd.org/en/topics/transfer-pricing.html   

OECD (2008). AI in Tax Administration: Governing with Artificial Intelligence. OECD Publishing. https://www.oecd.org/en/publications/governing-with-artificial-intelligence_795de142-en/full-report/ai-in-tax-administration_30724e43.html 

24 total views, 24 views today