Use of Artificial Intelligence in Tax Management

Bruno Lopes de Paulo; Adriana Diniz Gurgel

DOI: 10.22167/2675-6528-202602744

Article derived from a Course Conclusion Work (TCC), with content based on the student’s original work and adapted to the editorial format of the E&S Magazine with the support of the ResumeAI tool, an artificial intelligence solution developed by Instituto Pecege for textual synthesis and organization.

Summary

Tax compliance in Brazil is a central concern for companies in the development of their operational activities. The work conducted an experimental study on the use of artificial intelligence (AI) in tax management, employing the Retrieve-Augmented-Generation (RAG) technique. Legislative texts were selected to create a vector database through embedding in a Python project. A questionnaire with 20 pairs of questions and answers about the legislation for Corporate Income Tax (IRPJ) and Social Contribution on Net Profit (CSLL) was developed. Large Language Models (LLMs), running locally and via APIs (OpenAI and Huggingface), were submitted to evaluate response time, number of tokens analyzed, and legal accuracy. API-based models demonstrated better performance in efficiency and response time, with OpenAI’s model achieving 85% accuracy in responses, significantly outperforming local models. It was concluded that AI can generate gains in time and accuracy in tax management, but requires robust projects, investment in hardware, and the expertise of tax professionals to mitigate risks and ensure the correct interpretation of legislation.

Keywords: CSLL; Embedding; IRPJ; LLMs; RAG.

1. Introduction

Tax compliance in Brazil is a central concern for companies in the development of their operational activities and in maintaining financial health. The cost of tax compliance is one of the main challenges, being instinctively perceived by taxpayers due to the time involved and the inherent concerns (Bertolucci, 2003).

The Brazilian tax system is recognized for its complexity, characterized by a vast field of laws, norms, and regulations that frequently conflict, generating legal uncertainty. Since the promulgation of the 1988 Constitution, more than 5.4 million norms have been issued, with an average of 769 norms per day, which positions Brazil among the nations with the most complex tax relations in the world (Amaral et al., 2017). Legal certainty is a fundamental principle that aims to guarantee stability, being of extreme importance for both taxpayers and the State, as it generates confidence in the stable application of legislation over time (Silva et al., 2024).

Beyond quantitative complexity, Brazilian tax legislation presents an interpretative challenge. The National Tax Code (CTN) dedicates articles 107 to 112 to the interpretation and integration of legislation. Formal interpretation occurs before state bodies, while informal interpretation is the result of doctrine and tax planners (Bittencourt, 1999). Although literal interpretation is provided for in article 111 of the CTN, it should not be confused with isolated grammatical reading, and should be based on the undisputed core of the normative texts established by the legislator, grounded in legal certainty (Bustamante and Alvez, 2012).

Parallel to this tax reality, the technological scenario has experienced rapid advancement. The advent of the internet in the nineties and its evolution in the following decades transformed the speed and globalization of information traffic. The dissemination of artificial intelligence, especially with Large Language Model (LLMs) models like chat GPT, from the end of 2022, revolutionized the way we interact with machines and perform tasks.

Artificial intelligence enables the exploration of new activities, such as image creation and rapid text analysis. Artificial intelligence models can generate outputs from prompts, which are instructions given to the model, acting as intelligent assistants capable of automating repetitive tasks quickly and efficiently.

Given the complexity and volume of Brazilian tax legislation, manual analysis and the search for precise information consume significant time and resources. The gap lies in optimizing this process, where artificial intelligence can offer a solution to manage the vast volume of data and interpretive nuances, mitigating legal uncertainty and associated costs.

The application of artificial intelligence in tax management is justified by the possibility of promoting savings in tax expenditures and optimizing the time dedicated to the analysis of legislative texts. This allows companies to identify the best way to comply with their tax obligations safely. In this sense, the objective of the work is to use artificial intelligence to analyze the legislation of profit taxes in Brazil, which involves income tax and social contribution on net profit (IRPJ and CSLL). The research aims to bring savings in tax expenditures and time gains in the analysis of legislative texts in the search for the best path that a company can seek in carrying out its activities in order to be safe with its tax obligations, regarding profit taxes.

2. Material and Methods

The methodology employed in the present work was characterized as experimental research, according to Gil’s (2022) classification. The study focused on the application of artificial intelligence for the analysis of legislation concerning profit taxes in Brazil, specifically the Corporate Tax/ Income Tax On Legal Person (IRPJ) and the Social Contribution on Net Income (CSLL). The analyzed variables included the response time of the models, the amount of legislative text processed, measured in tokens, and the quality of the responses obtained.

For the research, the Retrieve-Augmented-Generation (RAG) technique was employed, developed for knowledge-intensive natural language processing tasks (Lewis, 2020). The procedure involved selecting legislative texts, creating a vector database, developing a questionnaire with question-answer pairs, submitting this data to various Large Language Model (LLM) models, and analyzing performance variables. A Python project was developed to execute these steps (Lopes, 2026).

The legislation selected for the database covered taxes on profit, including the laws governing the federal tax system and the calculation rules for IRPJ and CSLL. In addition to ordinary laws, decisions from the Administrative Council of Tax Appeals (CARF), decrees, normative instructions, complementary laws, and consultation solutions were incorporated. In total, 920 legislative text segments were indexed in the vector database, aiming to build a robust body of information.

The creation of the vector database was carried out through a Python code, chosen for its simple syntax and readability, being suitable for artificial intelligence work (Ramalho, 2022). The LangChain library (LangChain, 2026) was used, which offers an environment for artificial intelligence agent engineering. The database was implemented with SQLite, a lightweight database management system.

The first stage of the code consisted of creating the vector database. For the indexing of legislative texts, the WebBaseLoader, UnstructuredPDFLoader, and DirectoryLoader classes from the LangChain library were used. These classes were responsible for requesting links, reading unstructured PDF files, and scanning directories for various file formats (PDF, TXT), storing them in the database.

After loading the documents, the texts were divided into pieces, called “chunks”, for indexing. A Large Language Model (LLM) was used to transform these texts into numerical values, a process known as embedding. Embeddings are vector representations that capture the meaning of natural language units (Pilehvar and Camacho-Collados, 2021). For indexing, the “text-embedding-3-small” model from OpenAI was used, accessed via API.

To test the models, a questionnaire was developed consisting of 20 pairs of questions and answers about the taxation of IRPJ and CSLL. The questions were formulated based on the objective application of legislation, addressing tax rates, calculation bases, presumption percentages, and ancillary obligations. The construction of the questionnaire avoided questions that depended on jurisprudential interpretations or that could be subject to legitimate controversy (Carvalho and Rennó, 2020).

Large Language Models (LLMs) were submitted to Python code for processing and comparison. The tests were performed using two approaches: the first, with models running locally on the machine, using the Ollama tool, and the second, via APIs from OpenAI and Huggingface. The local models included llama3, mistral-nemo, phi3:medium, and gemma2:9b. The API models were deepseek-ai/DeepSeek-V3, Qwen/Qwen2.5-72B-Instruct, CohereLabs/c4ai-command-r-08-2024, and gpt-5.1.

The analysis of the collected data focused on three main variables: response time, quantity of tokens processed per second, and legal accuracy of the responses. For the evaluation of response quality, an evaluator model, OpenAI’s “gpt-5-nano”, was used. This model verified the fidelity of the response to the source provided by the vector database, establishing a hit or miss score, without analyzing the argumentative quality or legal support.

As methodological considerations, the study constituted a technical feasibility trial, not seeking to exhaust the judgment of the tax legal system. The indexed legislative base, although comprehensive, did not represent the entirety of applicable legislation or jurisprudence. The tests of the local models were carried out on a computer with 32GB of RAM, AMD Ryzen 5 5625U processor (2.3GHz) and AMD Radeon™ Graphics card (496MB), configurations that influenced the results. The legislation and versions of the models used were in effect during the period of the work’s development.

3. Results and Discussion

Experimental research on the use of artificial intelligence in tax management revealed important insights into the viability and performance of Large Language Model (LLM) models in analyzing Corporate Tax/ Income Tax On Legal Person (IRPJ) and Social Contribution on Net Income (CSLL) legislation. The models were evaluated in three distinct approaches: local execution on a computer, use via API from Hugging Face models, and access to paid OpenAI models. This methodology allowed for a comprehensive comparison of response time variables, number of tokens processed per second, and, crucially, the legal accuracy of the responses, providing a solid basis for understanding the capabilities and limitations of AI in this context.

The analysis of the average response time of the models for the 20 questions elaborated demonstrated a clear distinction between the approaches. Models executed locally, such as `oll_phi3`, `oll_mistral_nemo`, `oll_gemma2`, and `oll_llama3`, presented significantly longer response times, ranging from 361.11 seconds to 729.43 seconds. In contrast, models accessed via API, including `gpt_openai_5_1`, `hf_deepseek`, `hf_qwen`, and `hf_cohere_command_r`, were notably faster. The `gpt_openai_5_1` model achieved the best performance, with an average response time of 5.66 seconds, indicating superior efficiency in information retrieval and generation.

This difference in response time underscores the need for dedicated and robust hardware for the efficient execution of LLM models locally. The superior performance of API-based models, even with the limitations of free quotas, suggests that cloud computing infrastructure offers substantial advantages in terms of speed and processing. The implementation of artificial intelligence solutions in the business environment, therefore, requires considerable investment in technological infrastructure to ensure the necessary agility in analyzing large volumes of legislative text, as pointed out by the literature on the optimization of knowledge-intensive tasks (Lewis et al., 2020).

Regarding processing efficiency, measured in tokens per second, the API models also outperformed local models. `gpt_openai_5_1` demonstrated a processing capacity of 773.15 tokens per second, followed by `hf_cohere_command_r` with 480.67 tokens per second and `hf_deepseek` with 452.64 tokens per second. In contrast, local models such as `oll_phi3` and `oll_mistral_nemo` registered rates of 4.03 and 5.62 tokens per second, respectively. This metric is fundamental for evaluating artificial intelligence’s ability to quickly analyze complex tax law contexts, reinforcing the advantage of cloud-based models for tasks requiring high processing speed.

The legal accuracy of the responses was one of the most critical variables evaluated, using an evaluator model (`gpt-5-nano`) to determine the correctness of the 20 questions. The `gpt_openai_5_1` model achieved an accuracy of 85%, which represents a significantly superior performance compared to the other models. The `hf_deepseek`, `hf_cohere_command_r`, and `hf_qwen` models obtained 60%, 55%, and 55% accuracy, respectively. The local models, on the other hand, presented more modest results, with `oll_gemma2` reaching 35%, `oll_llama3` 25%, `oll_mistral_nemo` 25%, and `oll_phi3` only 10% correct.

This disparity in accuracy highlights the importance of training quality and the models’ interpretation capabilities for handling the complexity of Brazilian tax legislation. The hermeneutic presumption that underpinned the questionnaire’s development focused on the objective application of legislation, such as rates and presumption percentages, allowing for a direct verification of the answers against the legislative text. The superiority of the OpenAI model suggests a greater ability to understand and apply legal nuances accurately, which is essential for mitigating legal uncertainty and tax compliance costs (Bertolucci, 2003; Silva et al., 2024).

To illustrate the models’ performance, the question about the basic Corporate Tax rate and the applicable additional percentage was considered. The expected answer was 15% for the basic rate and 10% additional on the portion of profit exceeding R$ 20,000.00 per month. The `gpt_openai` model provided a correct and detailed answer, citing the basic rate of 15% and the 10% additional on the limit of R$ 20,000.00, based on the relevant legislation. Similarly, the `mistral-nemo` and `phi3` models also presented correct answers to this question, demonstrating the ability to extract direct information from the vector database.

Another relevant example was the issue regarding the taxation of capital gains on the sale of fixed assets in the Presumed Profit regime. The expected answer indicated that capital gains are taxed in full, applying the Corporate Tax/ Income Tax On Legal Person (15% + additional) and Social Contribution on Net Income (9%) rates to the positive difference between the sale value and the book value. The `gpt_openai` model detailed that capital gains are integrated into the tax base for Corporate Tax/ Income Tax On Legal Person and Social Contribution on Net Income as non-operating income, calculated by the positive difference between the sale value and the book value, according to legislation. The `mistral-nemo` and `deepseek` models also provided answers aligned with legislation, albeit with different levels of detail and article citations.

The question about the CSLL presumption percentage for trading companies under the Presumed Profit regime, with an expected answer of 12% on gross revenue, was also analyzed. The `gpt_openai` model confirmed the 12% percentage, citing Art. 20 of Decree-Law No. 1,598/1977. The `llama3` and `qwen` models also presented correct answers, reinforcing the capacity of API models to extract and apply specific legislative information. The variety in the models’ responses, even when correct, reflects the quality of information stored in the vector database and how each model processes and addresses the raised question.

It is fundamental to consider the limitations of the research for an adequate interpretation of the results. The study constitutes a technical feasibility trial, focused on answering questions based on excerpts of legislation retrieved from a vector database, without exhausting the complexity of the tax legal system, which involves broader hermeneutics. The indexed legislative base, composed of 920 segments, does not represent the entirety of applicable legislation nor judicial or administrative Common/ Case law, although it includes ordinary laws, regulations, and normative instructions of Corporate Tax/ Income Tax On Legal Person and CSLL.

The test environment for the local models also imposed limitations, as specific hardware resources were used (computer with 32GB RAM, AMD Ryzen 5 5625U processor, and 496MB AMD Radeon™ Graphics graphics card). These configurations directly influenced the performance results of the local models, suggesting that more powerful hardware could alter these outcomes. Furthermore, the temporal scope of the legislation and the versions of the models used in the APIs is specific to the period of the work’s development, which implies that reproduction at a later time may generate different results, given the constant evolution of legislation and AI models themselves.

Despite these limitations, the results demonstrate that artificial intelligence, especially advanced models via API, has significant potential to optimize tax management. The ability to generate responses with high accuracy and in reduced time can provide substantial time savings and economies in tax expenditures for companies. However, the successful implementation of such solutions requires robust projects, investment in technology, and, crucially, the expertise of tax professionals to validate interpretations, mitigate error risks, and ensure compliance with complex Brazilian legislation (Carvalho and Rennó, 2020; Bittencourt, 1999).

4. Conclusion

The experimental research sought to employ artificial intelligence to optimize the analysis of Corporate Tax/ Income Tax On Legal Person (IRPJ) and Social Contribution on Net Income (CSLL) legislation in Brazil, aiming for gains in time and accuracy in tax management. It was found that API-based Large Language Model (LLMs) models, notably OpenAI’s, showed superior performance in efficiency and response time, with the `gpt_openai_5_1` model achieving 85% legal accuracy in its responses. In contrast, locally executed models demonstrated significantly longer response times and lower accuracy. This capability to generate precise answers in reduced time represents a significant practical contribution to the optimization of tax compliance, offering companies a tool to mitigate the complexity and costs associated with the vast and dynamic Brazilian tax system.

However, the study was configured as a technical feasibility study, and the indexed legislative base, composed of 920 segments, did not cover the entirety of the tax legal system, including judicial or administrative jurisprudence, which limits the generalization of the findings. The hardware configurations of the test environment for the local models also directly influenced the performance results, suggesting that more powerful hardware could alter these outcomes. For the effective implementation of artificial intelligence in tax management, robust projects, investments in technological infrastructure, and the indispensable expertise of tax professionals are necessary to validate interpretations, mitigate error risks, and ensure compliance with complex legislation. Future studies could focus on expanding the vector database to include jurisprudence and on evaluating models in more powerful hardware environments, as well as exploring applicability in other areas of tax law.

Bibliographic References

AMARAL, Gilberto L.; João Eloi; AMARAL, Letícia M. F. 2016. Quantidade de Normas Tributárias Editadas no Brasil: 28 anos da constituição de 1988. São Paulo, SP, Brasil. Disponível em: <https://static.poder360.com.br/2017/06/Normas-editadas-ibpt-30jun2017.pdf>. Acesso em 25 ago. 2026.

BERTOLUCCI, Aldo V. 2003. Quanto custa pagar tributos. Atlas, Rio de Janeiro, RJ, Brasil. Disponível em: <https://app.minhabiblioteca.com.br/reader/books/9788522472475/.> Acesso em: 20 out. 2025.

BITTENCOURT, Marcelo M. R. 1999. A Interpretação no Direito Tributário. Revista do Ministério Público, n. 9,

SILVA, A.F.; ROCHA, D.A.; SOUZA, E.M.; FILHO, J.I.; ALENCAR, J.M.; LEITÃO, R.D.; PORTO, R.A.. 2024. Direitos Fundamentais do Contribuinte. Foco Jurídico Ltda, Indaiatuba, SP, Brasil.

Article originating from the Course Conclusion Work of the Specialization in Tax Management of the MBA USP/Esalq

To learn more about the course, click here and access the MBX Academy platform