Artificial Intelligence and the Duty to Professionalize Customs: The Valorization of Human Capital Through Technology and Digital Technologies

1.Introdução

In recent years there has been a growing interest in the customs community for technological issues, which has been evaluated both from the perspective of the Customs State, which has the prerogative and duty to implement such technologies, considering all international agreements concluded for this purpose (1), and from the perspective of the administered, who experience their benefits, but also new forms of submission to the already known customs powers (2). However, it is considered that the phenomenon should not be observed exclusively under these two prisms, as if the customs fact (3) It was constituted solely by the legal relations established between Customs and economic agents.

Customs officers, as representatives of the Administration, must also be prepared to act and interact with smart technologies; learn to think with them and counteract the results they generate (4). Without professionalized and digitally trained officers, customs administrations will not be able to achieve the goals of efficiency, effectiveness, integrity and transparency that are the promises of the ongoing digital revolution.

The purpose of these brief reflections, therefore, is to analyze, from a legal perspective, some of the impacts that the use of artificial intelligence imposes on customs activities and the competencies and skills required of customs officials in these new interactions. To this end, we will illustrate how the use of generative artificial intelligence and machine learning for the fiscal classification of goods can be risky, and the reasons why such tools should serve solely as support for human decisions, both for those governed and especially for customs agents. It is the duty of the State-Customs to professionalize its agents, not only through digital training but, above all, by combining this knowledge with the necessary technical development in essentially customs matters, for the benefit of the entire community.

2. Some challenges related to the use of AI in customs activities: the example of tax classification of goods

One of the most important purposes of tax classification is to identify the customs tariff that must be paid for goods intended to enter a country, as well as to verify whether they are subject to any type of import restriction or prohibition.

To do this, it is essential to know the characteristics of the goods in order to determine their constituent material, function, or purpose. Then, the Harmonized System (HS) must verify the likely sections in which they could be included, as well as the possible chapters or chapters that group them together, in order to subsequently identify the definitive headings and subheadings. Depending on the type of tariff—whether national, regional, or multilateral—the classification may have eight, ten, twelve, or more digits, depending on the type of breakdown applied, by country or region.

Given the inherent complexity, importers have for some time now turned to foreign trade management software to assist them in the task of correctly classifying their goods. Protracted—sometimes interminable—disagreements between customs authorities and importers over the definition of certain classifications within the Harmonized System are not uncommon.

Importers and customs authorities now hope to overcome such complexity by incorporating digital technologies (especially generative artificial intelligence and machine learning) for both the classification of goods and the identification of errors and fraud of this nature, something that has been viewed with varying reservations by merchandise specialists.

Álvaro Fernández-Acebes and Osiris Ramírez Ponce de León carried out a detailed study on the use of artificial intelligence in freight classification activities. (5) The authors listed the ten main problems of merceology activity and, based on them, analyzed how the AIs developed by the customs of some countries (6), as well as some private tools (7) were able (or tried) to solve them.

Once the systems have been tested, the first thing researchers can verify is that these tools draw from multiple sources, including query solutions and historical customs declaration data. That is, they consider the same data sources available to humans, but they look for correlations that might not be available to the human mind. (8).

This, by the way, is a very present feature in algorithms equipped with artificial intelligence: correlation (9). Modeled under a logic called fuzzy (diffuse) (10), AI algorithms are characterized by correlations and predictions that allow them to make decisions anticipating future behavior. The simplest way to characterize fuzzy logic, according to Mazzarese, is to say that it is a logic of approximate reasoning. It is a logic whose reasons use: i) fuzzy truth values; ii) imprecise truth tables; and iii) inference rules whose validity is more approximate than exact (11). The logic of approximate reasoning does not aim to guarantee the certainty of the results obtained, but only a justifiable solution, a plausible decision. Therefore, such correlations are not without risk, especially when the task they must perform involves the interpretation of legal norms, as is the case with the classification of goods.

Continuing with the tests, Fernández-Acebes and Ponce de León confirmed the first of the concerns that motivated their study: despite the multiple data sources, some of the systems examined were not being fed by primary legal sources of classification, such as explanatory notes, WCO criteria, sentences, etc., but only by query solutions based on these primary sources, which could certainly generate distortions in the machine results, considering the quality of the data used and the possible imprecise correlations (12).

Furthermore, they confirmed another limitation of generative AI, as verified in tests with ChatGPT: the tool does not present the legal basis used to arrive at its results (13). Considering that the activity of classifying goods must obligatorily comply with the steps of (i) description of the goods; (ii) assignment of the corresponding code; and (iii) indication of the legal basis, the indication of a code without referring to the legal basis (General Rules of Interpretation, Chapter or Section Notes) means providing an incomplete classification (14).

Another concern noted in the research by Fernández-Acebes and Ponce de León on the use of smart chats is that they do not always provide the same answer when different chats are consulted for an identical case (15). The authors confirmed a problem of reproducibility when using generative AI for cargo classification, since the same conclusions are not always reached when faced with similar or even identical issues.

And on this point, some clarification is necessary. As with all artificial intelligence models, generative AI also requires training. However, unlike narrow AI, generative AI does not use specific data nor is it designed for predetermined outcomes. Contrary to popular belief, generative AI does not "think," "write," "read," or "draw." In fact, it has no idea of ​​the intrinsic meaning of the content it produces.

The World Customs Organization published a note on research into generative AI and its use by customs at the end of 2023 (16). In this study, the WCO noted that smart chats use unlimited sources of language throughout their entirety. Therefore, what happens is that generative AI they learn how language works, especially from two basic principles: the distance between words and the relationship between them.

Therefore, while in narrow AI we face the problems of data bias, algorithmic biasesIn generative artificial intelligence this limitation is called hallucination, or hallucination, in which errors are made due to the use of nonexistent references. The unpredictable nature of these errors, as well as their potential impact, requires rigorous review by administrators and customs agents, both to confirm the results and, especially, to refute them.

There are, of course, various other problems related to the use of artificial intelligence in tax classification activities, some of which are not even known at this time (17).

But beyond the tools created for the purpose of assisting economic agents well-intentioned (18) In the assignment of a classification, there are those that have been developed by customs to manage risks, identifying errors and frauds in the course of customs clearance and, based on their results, applying sanctions and penalties. The Brazilian customs office is an example of this.

As we have already highlighted on another occasion (19), the Brazilian Federal Revenue Service has been using artificial intelligence through machine learning in its control activities for at least ten years, with the promise of reducing repetitive workload and valuing customs agents. This system has recently undergone modifications, such as the inclusion of violation notifications and query solutions in its database, so that, from them, the machine can infer the nomenclatures that should be assigned to the goods subject to future customs declarations. This improvement to the dataset is undoubtedly positive, but it still seems insufficient to us compared to the risks that its inadequate application can entail for those administered. This is due to the existence of structured data, that is, available information of evident relevance that has been ignored in the SISAM feedback: these are the customs declarations submitted to administrative or judicial litigation. If the system considered the outcome of these trials, in which the dispute passed through the control of the adversary and the decision of an impartial and independent authority, there would be greater chances that the tool would be better trained, in addition to contributing to the reduction of customs litigation (20). The system has also been recently modified and trained to perform correlations between the presence or absence of errors in the declaration attributes, taking advantage of all customs declarations rectified by determination of the customs authority due to alleged discrepancies (21).

However, due to the dynamics of foreign trade, which is always marked by the urgency of clearing goods, many importers accept tax reclassifications during the clearance process, because the damage suffered due to logistical expenses and the breach of commercial contractual obligations can be significantly greater than the tax difference and the corresponding customs fine.

The time factor does not always allow foreign trade stakeholders to request a tax classification consultation in advance, or even a technical opinion that supports and ratifies the choice of the NCM used in their import documents. Due to this emergency nature, many declarations end up being rectified and cleared using known incorrect nomenclatures. And, in this case, the tool designed to learn and indicate historical errors may be neutralizing and repeating errors committed by the customs agents themselves in the past, and not only. errors committed by importers, as assumed.

It is clear that the problems highlighted here are not intended to suggest the abolition or prohibition of such tools. The point of these reflections is to demonstrate that systems powered by artificial intelligence, in any field of knowledge, should be solely instrumental; financial for decision-making by human beings, the true agents of transformation.

3. Artificial intelligence and the duty to professionalize customs: the valorization of human capital through technical and digital training

The implementation of digital technologies has been recognized since the end of the last century as one of the crucial strategies for modernizing customs, both to make the customs service more efficient and to limit discretion and even increase levels of trust in the institution.

Historically, customs agents who performed control activities, especially cargo inspection, performed these tasks manually. Records were also kept in large ledgers, which could take many days from the arrival of the goods to their delivery at the final destination. Over time, customs offices have implemented new operational strategies to reduce these times, aiming to offer society customs services that promote security and also facilitate the international exchange of goods.

And in this process of introducing technologies, it is true that customs administrations have faced all kinds of resistance, especially from those customs agents who have become accustomed to working in the same way for long periods of time. Problems such as lack of adaptability, fear of reduced income, job loss, and even the loss of control or power by specific groups within the organization are just some examples of these obstacles (22).

On the other hand, customs agents point out that the lack of professional recognition and the inadequate allocation of profiles for internal functions within the customs administration have also contributed to this uncertain and negative environment, which ultimately generates such fears and even a lack of interest in these technological innovations.

Although the WCO is the leading international body establishing guidelines for customs administrations, the organization does not have a specific document that explains in detail the requirements for recruitment and the formation of functional staff. However, through guidelines, recommendations and standards, the Organization has established general principles that should guide member countries, such as the Arusha Declaration, a reference for good governance and integrity for customs around the world (23). The principles mentioned in the Arusha Declaration focus mainly on ethics, integrity, transparency, impartiality and confidentiality in conduct. In this regard, the Declaration recommends that customs establish clear and binding standards for the expected behavior of customs agents.

But beyond these principled guidelines, customs administrations must also establish and maintain an updated human resource management system, referred to by Lopez as the set of policies, processes and practices designed to maximize the performance of customs agents and align their efforts with the institution's strategic objectives. This system includes hiring and selection, training, performance evaluation, compensation and benefits (24).

In 2019, the World Customs Organization created the project called BACUDA – Band of Customs Data Analysts (25), which aims to raise awareness and train members of customs administrations in data analysis, one of the most relevant topics when implementing smart technologies in customs control activities. The project team works closely with research groups from academia and research institutes to develop methodologies and disseminate good practices in data analysis, including the provision to customs authorities of algorithms specifically designed for the detection of fraud in customs value as well as for the detection of errors in tax classification.

The Brazilian government has also instituted a digital training program aimed at all public servants, regardless of their level and power, with the goal of training new data scientists. The platform, developed by the National School of Public Administration (ENAP), offers more than fifty courses, some of which teach everything from the fundamentals of computer vision (26) to how to be more productive using data mining (27) and generative AI (28).

Therefore, in terms of digital training, there is no shortage of national and international initiatives aimed at improving the human-computer interaction we discussed earlier. However, the necessary professionalization we are referring to is not limited to mastering these technological resources and tools.

We have demonstrated some of the risks that the use of artificial intelligence in tax classification can entail for both the governed and the customs agents themselves. For customs professionals to be able to extract the best results from these tools, it is necessary to develop critical thinking, which can only be achieved if these agents have been adequately and sufficiently trained in the essentially customs disciplines. Only in this way will they be able to criticize and eventually refute erroneous or even nonexistent results of the tools. chatbots and machine learning machines, especially when such automated decisions directly impact the rights of those governed, users of customs services.

When examining the call for applications for the latest training course for tax auditors who will join the tax and customs careers of the Federal Revenue of Brazil (29), it was found that the employees received only 48 hours of training related to customs matters. Of these 48 hours, it was found that the training did not cover technical subjects, such as tax classification of goods, customs valuation and rules of origin, but only procedural or procedural content (30).

At this point, Regina Maria Macedo Nery Ferrari's contributions are valuable, emphasizing that when speaking of public service efficiency through the professionalization of public servants, the objective of professionalization must be understood as the effective improvement of state action. This depends on the combination of factors economy y celerity, covering both productivity and perfection of work, its technical adequacy to achieve the ends pursued by the State, with evaluation of the results. This means that the State, in demanding efficiency from the server, must provide all the means and incentives necessary for its improvement and growth in the professional career (31). Considering that artificial intelligence is an empowering tool, but only complementary Among the tasks entrusted to the customs service, the valorization of human capital is an imperative measure for the fair and sustainable modernization of customs today and tomorrow. The obligation of the Customs State to professionalize customs agents also stems from its duty to ensure the protection of the legal order that emanates from it, based on the interests of society, preventing the social regression of the rights guaranteed to citizens.

---

1.A Revised Quioto Convention, as well as the Agreement on the Facilitação do Comércio são os
There are main international agreements that prestige the implementation of technologies in customs control activities.

2.These aspects are addressed by us in recently published work. Reis, Raquel Segalla. Risk management in customs clearance of imports: artificial intelligence as an instrument and control agent. São Paulo, SP. NSM Editora/Caput Libris, 2024.

3.Segundo Andrés Rohde Ponce “customs duty” is an expression that consigns the entry of an object in a certain space that is subject to a certain legal order. In: Rohde Ponce, Andrés. The Fundamental Elements of Customs Law. In: Pardo Carrero, Germán (Ed.); Marsilla, Santiago Ibáñez; Yebra, Felipe Moreno (Coordinator). Customs Law, Volume I. Bogotá: University of Rosario; Tirant lo Blanch, 2019, p. 119. For José Lence Carluci, the “customs fact” is a complex of legal and economic facts of varied nature from which the customs relationship is derived from the complex of bodies, activities and administrative structure that, together, make up the customs reality. In: CARLUCI, José Lence. An introduction to the customs clearance. São Paulo: Customs, 1997. p. 20. 4 World Customs Organization.

4.World Customs Organization. Research & Policy Note on Generative Artificial Intelligence for Customs. Available in: chrome extension://efaidnbmnnnibpcajpcglclefindmkaj/https://www.wcoomd.org/-
/media/wco/public/global/pdf/topics/research/papers/researchpolicynote_genai_en_06122023.pdf?la=fi

5.FERNÁNDEZ-ACEBES, Álvaro; PONCE DE LEÓN, Osiris Ramírez. Customs Tests: 10
Mercological-tariff classification problems that artificial intelligence can't help with. Foundation for the Dissemination of Knowledge and Customs Law. Arola Editors: Tarragona, 2024.

6.United States, as its “Census Bureau Schedule B Search Engine”, available in:
https://uscensus.prod.3ceonline.com/ Acesso em: 10 fev. 2025; México, Bahamas foram algumas das aduanas cujos softwares foram analisados. A ferramenta gratuita fornecida pela OMA também foi explorada pelos pesquisadores.

7. Primarily or chatpgt, Google's generative artificial intelligence, in addition to tools such as Experta, from Mexico, Smart HS, from the United States.

8.FERNÁNDEZ-ACEBES, Álvaro; PONCE DE LEÓN, Osiris Ramírez. Customs Tests: 10
Mercological-tariff classification problems that artificial intelligence can't help with. Foundation for the Dissemination of Knowledge and Customs Law. Arola Editors: Tarragona, 2024, p. 150.

9. We deal with more details of these questões in recently published works. Reis, Raquel SegallaRisk management in import customs clearance: artificial intelligence as a control instrument and agent. São Paulo, SP. NSM Editora/Caput Libris, 2024.

10. From a mathematical point of view, fuzzy logic consists of the ability to approximate the real world, in situations where there are only extreme answers. Fuzzy logic gives space in the short term, presenting the possibility of measuring the degree of approximation of the exact solution and thus inferring something that is necessary. Unlike classical logic, which attributes only true or false values, fuzzy logic consists of a range that proposes that it is only a grau quest. In: MARRO, Alessandro Assi et al. Fuzzy logic: concepts and applications. Natal: Universidade Federal do Rio Grande do Norte (UFRN), 2010. p. 2.

11.Mazzarese, Thecla. Fuzzy Logic and Judicial Decisions: The Danger of a Rationalist Fallacy. 1996, p. 210. Available at: <https://rua.ua.es/dspace/bitstream/10045/10475/1/doxa19_12.pdf>.
Accessed on: February 08, 2025.

12.For this purpose, Brazilian customs developed a new system called CLASSIF. According to its representatives, it analyzes the classification of markets using NCM codes, which are based on codes from the Harmonized System of Description and Codification of Markets (HS), with two additional digits. Importers and exporters can insert a description of a market in the system, which will then show them information, such as the NCM codes that contain the words provided and the legal notes related to these codes, as well as the Explanatory Notes of the OMA for the Harmonized System of the OMA, consultation solutions issued for these codes and Classification suggestions that you see directly from SISAM. In: JAMBEIRO FILHO, Jorge; COUTINHO, Gustavo Lacerda; MORGERO, Kelly. how Brazil transformed Customs control through artificial intelligence and other Technologies. WCO News 105 – Issue 3 / 2024 > Dossier: Illicit Trade.

14.It is necessary to register, therefore, that we test mentioned by authors or ChatGPT recommends that national or regional legislation be reviewed, as appropriate, and also consult a customs specialist with all information about the product.

15.In addition to ChatGPT, a type of generative AI tested by the authors, various analog products are available on the market, such as ChatSonic, JasperChat, Youchat, Elicit, AnonChatGPT, Socratic, etc. Each has its own peculiarities, but all are based on the same AI principle. In: FERNÁNDEZ-ACEBES, Álvaro; PONCE DE LEÓN, Osiris Ramírez. Customs Trials: 10 tariff classification problems that artificial intelligence can't help with. Foundation for the Dissemination of Knowledge and Customs Law. Arola Editors:Tarragona, 2024, p. 154.

16. World Customs Organization. Research & Policy Note on Generative Artificial Intelligence for Customs. Available in: chrome-
extension://efaidnbmnnnibpcajpcglclefindmkaj/https://www.wcoomd.org/-
/media/wco/public/global/pdf/topics/research/papers/researchpolicynote_genai_en_06122023.pdf?la=fi

17.Optou-is the analysis of the topic of tax classification, but it is certain that new technologies are being used for the verification of value frauds, for the confirmation of the origin of markets, among other topics subject to control.

18.Expressão used breeder hair from CLASSIF to differentiate the SISAM – Selection System
Customs by Machine Learning, implemented by Brazilian customs to identify errors and frauds during customs clearance. In: JAMBEIRO FILHO, Jorge; COUTINHO, Gustavo Lacerda; MORGERO, Kelly. How Brazil transformed Customs control through artificial intelligence and other Technologies. WCO News 105 – Issue 3 / 2024 > Dossier: Illicit Trade.

19.Reis, Raquel Segalla. Big Data and artificial intelligence do not control customs: convergence between strategy, law and ethics. Customs News. Available at: https://aduananews.com/pt/big-data-e-
Artificial intelligence in customs control: convergence between strategy, law, and ethics

20.Reis, Raquel Segalla. Risk management in customs import clearance: artificial intelligence as an instrument and agent of control. São Paulo, SP. NSM Editora/Caput Libris, 2024, p. 193.

21.In: JAMBEIRO FILHO, Jorge; COUTINHO, Gustavo Lacerda; MORGERO, Kelly. How Brazil
transformed Customs control through artificial intelligence and other Technologies. WCO News 105 –
Issue 3 / 2024 > Dossier: Illicit Trade.

22.JULIUS, Kugonza; CHRISTABEL, Mugalula. Effectiveness and efficiency of artificial intelligence in boosting customs performance: a case study of RECTS at Uganda customs administration. World Customs Journal, v. 14, no. 2, p. 177-192, 2020.

23.The Portuguese language version of the Arusha Declaration is available in: chrome-
extension://efaidnbmnnnibpcajpcglclefindmkaj/https://www.wcoomd.org/-
/media/wco/public/pt/pdf/topics/integrity/instruments-and-tools/compila%C3%A7ao-de-praticas-de-
integrity.pdf?la=em

24.LÓPEZ, Ivonne Stephany Rodríguez. Considerations and challenges in the modernization process of the Colombian customs authority. Master's degree in international taxation, foreign trade, and customs. Externado University of Colombia. September 2024. Available in:
https://bdigital.uexternado.edu.co/entities/publication/fd3eb1d1-74a1-4839-b8bb-ce5d21eafbfd

25.World Customs Organization. BACUDA Project – Band of Customs for Data Analysis. Available at: https://bacuda.wcoomd.org/#bacuda

26.Federal Government. Virtual School. ENAP. Available in: Fundamentals of Computational Vision – Azure AI

27.Federal Government. Virtual School. ENAP. Available in: Information on Documents and Mining of Knowledge

28. Federal Government. Virtual School. ENAP. Available at: https://www.escolavirtual.gov.br/curso/1093

29.Ministry of Economy. Special Secretary of the Federal Revenue of Brazil. Public Competition for the promotion of wages in the tax and customs billboards of the Federal Revenue of Brazil. Editing of the Call for the Professional Training Course n. 1/2022. Available in: chrome-extension://efaidnbmnnnibpcajpcglclefindmkaj/https://conhecimento.fgv.br/sites/default/files/concursos/convocacao-curso-de-formacao-auditor_v3_semrevisoes.pdf Access in: 08 Feb. 2025.

30. Disciplines offered in the training course that contemplates customs issues: Regras e Fundamentos o Comércio Internacional – RCI*: 4 hours; Customs Administration and Control Model: 6 hours; Gestão de Intervenientes: 2 hours; Cargo Control, Information Flow and Border Management: 4 hours; Operational Customs Clearance: 8 hours; Risk Management and Customs Fiscalization: 10 hours; Differentiated Customs Processes: 4 hours; Authorized Economic Operator: 2 hours; Customs Surveillance and Repression: 8 hours.

31.Ibidem, p. 34.