Article

Ângela Cristina Moreira de Santana; Álvaro Freitas Faustino Dias

DOI: 10.22167/2675-6528-202602295

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

Abstract

The advent of artificial intelligence as a disruptive technology has reconfigured economic, institutional, and legal processes, conferring centrality to contractual automation in the management of complex relationships. Contractual automation has proven to be a legally non-neutral phenomenon, with decision-making opacity, automated generation of abusive clauses, and difficulties in assigning responsibility, shifting contractual drafting to the legal-normative domain. The study analyzed how artificial intelligence in contractual automation intensified the collision between technological efficiency and legal certainty, and to what extent mechanisms of significant human oversight and algorithmic governance functioned as instruments of normative balancing to preserve principles of Brazilian contract law. The research adopted a qualitative approach with multiple case studies, including documentary analysis of four paradigmatic cases: Moffatt v. Air Canada, United States v. RealPage, Deloitte/DEWR, and INSS/TRF-4ª Região, representing consumer, competition, contractual advisory, and administrative contexts. The analysis identified five recurring structural patterns – decision-making opacity, substitution of normative rationality by probabilistic inferences, accountability gap, informational asymmetry, and semantic incomprehension – that emerged systemically. The fragmented outcome of the RealPage case reinforced the conclusion that jurisprudence on decision-making automation remains in formation. Effective algorithmic governance has demonstrated the need to integrate auditability, accountability, and human oversight mechanisms, in accordance with ISO 31000, COSO ERM, and ISO 37301 standards, as a necessary condition for the legitimacy of automated contractual management systems.

Keywords: Contracts; Algorithmic governance; Artificial intelligence; Legal certainty; Human oversight.

1. Introduction

The advent of artificial intelligence (AI) as a disruptive technology has inaugurated a paradigm shift on a global scale, reconfiguring economic, institutional, and legal processes, and conferring increasing centrality to contractual automation in the management of complex relationships. Historically, Law adapts to technological innovations, straining traditional concepts and driving normative evolution (Lessig, 2006; Zuboff, 2019). The exponential increase in contractual volume and regulatory complexity has stimulated the adoption of AI-based contractual automation systems, making it economically irresistible. However, this pressure tends to undervalue diffuse and long-term costs, such as the erosion of human oversight and decision-making opacity, distorting the efficiency calculation (Han, 2018; Posner, 2014; Shavell, 1980).

Contract automation has proven to be a legally non-neutral phenomenon, with decision-making opacity, automated generation of abusive clauses, and difficulties in assigning responsibility. At the international regulatory level, Regulation (EU) 2024/1689 – the European AI Act – imposes obligations for transparency and human supervision. In Brazil, Bill 2338/2023, approved by the Federal Senate, was awaiting a vote in the Chamber, and the national legal system operates in “eloquent normative silence,” keeping the matter under the general regime of the Civil Code, the Consumer Defense Code, and the General Data Protection Law. AI systems are not technologically neutral, incorporating values and biases (Coeckelbergh, 2020). The drafting of contracts, which requires normative interpretation and balancing of interests, cannot be entirely replaced without compromising legal certainty and principles such as objective good faith and contractual balance (Mackaay and Rousseau, 2020; Schreiber, 2022).

Despite advances in the analysis of the risks of decision-making automation (Floridi et al., 2018; Busuioc, 2021; Wachter, Mittelstadt, and Floridi, 2017), the systematic articulation between algorithmic governance and the foundations of private law remains limited. This gap in the Brazilian context manifests in the absence of a consolidated doctrinal typology to address algorithmic discrimination in contractual and labor relations (Mendes and Mattiuzzo, 2019), evidencing a mismatch between technological incorporation and the legal system’s capacity to absorb its effects. The theoretical relevance of this study lies in the need to examine the compatibility between the inferential rationality of artificial intelligence systems and the structuring principles of contract law, adopting a critical perspective to highlight the limits of technological compatibility with normative rationalities (Schreiber, 2022; Mackaay and Rousseau, 2020; Hildebrandt, 2015).

The central hypothesis holds that artificial intelligence-based contractual automation does not constitute mere instrumental innovation, but rather operates a silent ontological transformation on the contractual institute itself. This transformation is “silent” because it initially manifests as a gain in operational efficiency, but its normative effects, such as the substitution of legal reasoning by probabilistic inference, the erosion of human supervision, and the emptying of protective categories, only reveal themselves a posteriori, when litigation emerges or damage has already occurred (Hildebrandt, 2015; Surden, 2014).

The risks associated with decision-making automation do not stem from episodic failures, but from structural characteristics of these systems, revealing a permanent tension between the probabilistic logic of AI and the assumptions of the Rule of Law. Recurring structural patterns identified include decision-making opacity, the substitution of normative rationality with probabilistic inferences, the accountability gap, informational asymmetry, and semantic incomprehension (Goodfellow, Bengio, and Courville, 2016; Russell and Norvig, 2021; Mittelstadt et al., 2016; Yeung, 2018).

The recurrence of these structural patterns underscores the urgency of understanding the impacts of AI on legal certainty and the principles of contract law. This study is justified by the need for an in-depth analysis of these challenges, contributing to the development of robust governance mechanisms. The objective of the work is to empirically analyze the impacts of artificial intelligence-based contractual automation on legal certainty, based on a study of four paradigmatic cases in distinct institutional contexts: consumer, competition, advisory-contractual, and administrative, examining the extent to which mechanisms of significant human oversight and algorithmic governance are capable of mitigating structural risks arising from decision-making opacity, the substitution of normative rationality by probabilistic inferences, the accountability gap, informational asymmetry, and semantic incomprehension. From a business perspective, the research seeks to identify implementation parameters capable of reducing regulatory risks, improving contractual governance, and mitigating liabilities arising from the inadequate use of automated systems.

2. Material and Methods

The research was classified as applied, exploratory-descriptive, and qualitative, seeking to understand the legal risks associated with the use of artificial intelligence in contractual management and to propose practical governance parameters for organizations and regulators. A multiple case study design was adopted, suitable for investigating contemporary phenomena in a real context and under distinct institutional arrangements (Yin, 2015). Information gathering techniques included bibliographic research, documentary research, and the survey of public secondary data.

The selection of cases followed intentional theoretical sampling, guided by three cumulative criteria: legal relevance of the controversy (evidenced by a formal judicial, administrative, or regulatory decision, or a publicly documented contractual resolution); use of an artificial intelligence or automation system with an impact on legally relevant decisions (causally relevant function); and the ability to evidence central research categories (decision opacity, substitution of normative rationality, accountability gap, informational asymmetry, and semantic incomprehension), for comparability between units investigated under different institutional arrangements (Yin, 2015). Cases without formal registration, without direct relation to contractual principles, or without sufficient public documentation were excluded.

Searches were conducted in the Scopus, Web of Science, SSRN, and Google Scholar databases, using boolean combinations of descriptors in Portuguese and English, such as “contracts”, “artificial intelligence”, “algorithmic governance”, “algorithmic discrimination”, “contract automation”, “automated decision-making”, “legal certainty”, and “AI legal liability”. The research period was from 2010 to March 2026. The initial milestone of 2010 is justified by the consolidation of machine learning and the emergence of the first records of litigation involving automated systems. The final milestone of March 2026 corresponds to the data collection closing date, ensuring the currency and incorporation of recent precedents, such as the Deloitte/DEWR case (2024-2025) and the TRF-4 judgment (January 2025).

Initially, 27 potential cases were identified. After applying the eligibility criteria, four cases were selected as the main analytical units, accompanied by approximately 40 supporting documents, including judicial decisions, regulatory acts, institutional reports, incident records, and academic articles.

Each category of supporting document was retrieved from a specific source to ensure traceability: judicial and administrative decisions were obtained from the official repositories of the competent courts and bodies (British Columbia Civil Resolution Tribunal, United States Department of Justice, Regional Federal Court of the 4th Region); regulatory acts from the official portals of regulatory and legislative bodies (Official Journal of the European Union, Official Gazette of the Union, Chamber of Deputies, Australian Department of Employment and Workplace Relations); institutional reports from the websites of the involved organizations or their regulators (Deloitte Australia, DEWR); incident records from specialized databases (AI Incident Database, Responsible AI Collaborative); and academic articles from the Scopus, Web of Science, SSRN, and Google Scholar databases. The individualized list of the 40 supporting documents, indicating type, related case, and access source, is presented in Table 1.

Table 1. Distribution of supporting documents by type and by paradigmatic case

Document type	Air Canada	RealPage	Deloitte/DEWR	INSS/TRF-4	Sum
Judicial and administrative decisions	2	4	0	2	8
Regulatory and normative acts	0	2	1	1	4
Institutional reports	0	1	5	3	9
Incident records	2	0	4	1	7
Academic and doctrinal articles	6	3	0	3	12
Total per case	10	10	10	10	40

Source: Original research data.

The data analysis observed the theoretical replication logic to identify recurring patterns in distinct regulatory environments. Relevant information was extracted by a standardized analytical matrix containing: institutional identification, AI system used, decision-making function, type of algorithmic intervention, identified legal risks, affected principles, existing governance mechanisms, documentary evidence, and incidence of semantic misunderstanding. The analysis was carried out through content analysis, with thematic coding guided by a priori defined categories (Bardin, 2011). The results have an analytical and explanatory purpose, not statistical.

The coding process followed the three phases proposed by Bardin (2011): pre-analysis (floating reading and systematization of a priori categories: decisional opacity, substitution of normative rationality, accountability gap, informational asymmetry, and semantic incomprehension); material exploration (manual thematic coding, associating textual excerpts with one or more categories through the identification of units of record and context); and results treatment (consolidated coding in the standardized analytical matrix, replicated for the four cases). The comparative analysis of the results was performed through cross-case pattern matching, a technique recommended by Yin (2015) for multiple case studies. An internal consistency check procedure was also adopted, reviewing the coding of each case at a later time than the first, with a minimum interval of seven days, to verify the stability of the applied criteria. The categories are not mutually exclusive, allowing the same documentary excerpt to be coded simultaneously in more than one category.

3. Results and Discussion

Theoretical and Technological Foundation: Artificial Intelligence and Contractual Automation

Artificial intelligence (AI) is a field that develops systems capable of capturing information and acting based on defined objectives, approximating human cognitive abilities through data-driven learning (Russell and Norvig, 2021). Its foundation is Machine Learning (ML), where algorithms extract regularities from large databases and adjust their performance (Mitchell, 1997). Building upon ML is Natural Language Processing (NLP), which enables both predictive AI, for classifying and analyzing legal scenarios, and generative AI, for producing documents that appear human-written based on statistical correlations (Jurafsky and Martin, 2024).

In the legal context, this integration drives the automation of contractual management on “document automation and contract analytics” platforms, integrated into “legal analytics” (Abbott, 2020). Generative AI drafts clauses and predictive AI assesses risks, but both find limits in understanding the normative and axiological content of legal relationships (Ashley, 2017). Its statistical rationality, based on historical correlations, ignores the interpretation of precepts, judgments of validity, and weighing of principles, shifting the decision from the normative to the probabilistic.

Automation enhances efficiency in repetitive tasks with low regulatory complexity, but biased or outdated data can compromise legal certainty. The impact on the contractual cycle is transversal, but not homogeneous: in the pre-contractual and execution phases, economies of scale are real and risks are mitigable by qualified human supervision. However, in the formation phase, where “drafting” requires regulatory interpretation and capture of business intent, algorithmic substitution introduces structural risks of undue homogenization and uncritical reproduction of historically asymmetric patterns, straining the compatibility between probabilistic logic and the ethical and regulatory requirements of contract law.

Contemporary Contractual Theory, Constitutionalization of Private Law and Algorithmic Automation

The contract, in civil doctrine, is a bilateral or plurilateral legal transaction based on the agreement of wills to create, modify, or extinguish patrimonial legal relationships (Gonçalves, 2020; Diniz, 2021). The classic tradition of Contract Law, based on the autonomy of will and minimal state intervention, was strained by the finding that formal equality does not guarantee material balance (Atiyah, 1979).

In Brazil, the 1988 Constitution and the 2002 Civil Code subjected private autonomy to the limits of the social function of the contract and the preservation of contractual balance (Diniz, 2021; Tartuce, 2023). Objective good faith, social function, and contractual balance act as criteria for validity, interpretation, and standards of conduct for corrective intervention, especially in situations of informational asymmetry (Martins-Costa, 2018; Marques, 2021).

The clash between the relational conception of the contract, based on trust, and the economic analysis of law, focused on efficient risk allocation, reveals a conflict of rationalities. Economic analysis, aligned with the algorithmic logic of AI, reduces the individual to performance metrics, converting the contractual subject into a statistical profile (Posner, 2014; Han, 2018). In contrast, the classical tradition, which defines the human being as a political animal, emphasizes the deliberative capacity regarding justice and the common good (Aristotle, Politics, Book I, 1253a). Practical wisdom, or phronesis, which shapes the general norm to the concrete case, is replaced by probabilistic inferences in AI systems.

The uncritical incorporation of AI compromises principles such as good faith and contractual balance by replacing legal interpretation with mathematical probabilities (Surden, 2014), altering the essence of the legal act from a normative and consensual instrument to a predictive-actuarial mechanism (Hildebrandt, 2015; Zuboff, 2019). This empties the protective categories of contemporary private law, which presuppose subjects capable of deliberation, not objects of statistical inference.

Legal Risks in Contract Drafting: Decisional Opacity, Biases, Informational Asymmetry, Semantic Misunderstanding, and “Accountability Gap”

Decision opacity, or the black-box phenomenon, characterizes the difficulty of interpreting the internal decision-making processes in AI systems, especially in deep learning models (Goodfellow; Bengio and Courville, 2016; Russell and Norvig, 2021). In the legal context, this impossibility of reconstructing the decision-making path empties requirements for motivation, transparency, and accountability, limiting the imputation of responsibility and the effectiveness of judicial protection (Mittelstadt et al., 2016).

Algorithmic biases are systematic distortions that produce unfair, discriminatory, or disproportionate decisions or recommendations, amplifying structural inequalities (O’Neil, 2016). The reproduction of biases in databases compromises contractual balance by reiterating asymmetric patterns, favoring economically stronger agents, which can challenge the social function of the contract. In extreme cases, algorithmic exploitation of informational asymmetry may constitute a situation analogous to lesion (art. 157 of the Civil Code of 2002).

In the four cases analyzed, the bias does not manifest with equal clarity. In the RealPage case, price standardization approached doctrinal risk by favoring landlords over tenants. In the Air Canada, Deloitte/DEWR, and INSS/TRF-4 cases, the problem was informational error, semantic misunderstanding, and decision opacity, with no evidence of differential treatment between individuals. Empirical proof of bias requires evidence of disproportionate distributive impact, which explains the concentration of litigation on decision opacity and informational error, rather than on proven discrimination.

The decisional opacity also creates conditions for error regarding the content of the declaration (art. 139 of the Civil Code/2002) and, if there is conscious omission of relevant information, it may constitute fraud (art. 147 of the Civil Code/2002). Formally valid automatic clauses, but incompatible with the parties’ intention, are voidable. If abusive (art. 51 of the Consumer Defense Code), they are null and void.

Information asymmetry constitutes an autonomous risk, as even with auditable systems, the vulnerable party rarely has access, time, or knowledge to verify the rationality of algorithmically generated clauses or prices (Yeung, 2018).

A structural limitation of AI is the distinction between syntax and semantics in the legal domain. NLP models identify linguistic patterns, but their performance is restricted to the syntactic level. The Chinese Room thought experiment illustrates that a system can manipulate symbols with syntactic perfection without understanding their normative meaning (Searle, 1980). In law, this means that NLP systems rearrange symbols with formal sophistication, but remain incapable of understanding their normative sense (Hildebrandt, 2015).

Legal reasoning requires the attribution of normative meaning based on principles, values, and institutional contexts. The formal correctness of clauses or algorithmic inferences does not ensure material conformity with the legal system, generating a formal simulation of legality that conceals material inconsistencies. This limitation is clear in highly complex contracts, such as mergers and acquisitions (M&A), where algorithmic standardization tends to uncritically reproduce market clauses without adaptation to the specific case, producing dysfunctional or litigious formulations.

AI’s inability to grasp normative context generates serious consequences in the face of extraordinary and unpredictable events. The Theory of Unforeseeability (arts. 317 and 478 of the Civil Code/2002) presupposes an agent capable of perceiving that circumstances diverge from those contracted. A system trained on historical data would apply obligations as written, without distinguishing ordinary non-performance from force majeure, due to its lack of normative contextualization. Significant human oversight is, therefore, essential in moments of rupture.

The algorithmic homo economicus would apply the fine, while the Aristotelian political animal, with phronesis and cognitive flexibility, would renegotiate. The creation of win-win solutions in crises requires context recognition, consideration of interests, empathy, and normative imagination, faculties that define the irreplaceability of the human agent.

The accountability gap, arising from the fragmentation of responsibility among developers, operators, and users, is addressed in the algorithmic governance section, which examines mechanisms for explainability, auditing, and bias assessment (Barocas and Selbst, 2016; Goodfellow, Bengio, and Courville, 2016).

Algorithmic Governance: Theoretical Foundations and Regulatory Dilemmas

Algorithmic governance, according to Yeung (2018), is the set of institutional, technical, and normative mechanisms that structure, guide, or correct decisions and conduct by algorithmic systems. The shift from regulation by norms to regulation by code (“code is law”, Lessig, 2006) displaces the definition of lawful and unlawful to the technical architecture of systems, generating “techno-regulation” (Hildebrandt, 2015).

This shift generates four regulatory dilemmas. The first is opacity versus efficiency: the requirement for full explainability can reduce the predictive performance of deep learning models (Pasquale, 2015; Kroll et al., 2017). This was evident in the INSS/TRF-4 case, where the lawfulness of the tool’s use did not exempt the Administration from the duty of intelligible explanation.

The second dilemma is distributed accountability (accountability gap): algorithmic decisions result from chains involving multiple agents, requiring the overcoming of the bipolar author/victim paradigm of classic civil liability (Bodó et al., 2018; Kroll et al., 2017).

The third dilemma is inverted regulatory capture: the technical complexity of the systems may exceed the capacity of control bodies to understand and oversee them, a phenomenon described as “black box society” by Pasquale (2015). This was evident in the RealPage case, where users themselves did not understand the pricing logic.

The fourth dilemma is normative-temporal: the speed of technological development surpasses the legislative pace, creating an “eloquent normative silence” in Brazil. Effective governance depends on the combination of principled state norms and sectoral self-regulation with verifiable external accountability (Teubner, 2012).

The articulation of these dilemmas with the five structural patterns identified in the empirical analysis demonstrates that algorithmic governance is not an isolated technical response, but a multi-scalar institutional arrangement that combines general legal norms (Civil Code, CDC, LGPD), international technical standards (ISO 31000, ISO 37301, ISO/IEC 42001, NIST AI RMF) and internal organizational control mechanisms (compliance, audit, ethics committees), in line with COSO ERM (2017).

Paradigmatic Case Studies: Practical Analysis and Evidence

The empirical analysis examined concrete situations where automated systems produced legally relevant effects in consumer, competition, advisory-contractual, and administrative contexts. The cases show that AI introduces decision-making opacity, intensifies informational asymmetries, reproduces structural discrimination, and hinders the imputation of responsibility, requiring the reinterpretation of institutes such as objective good faith, transparency, and the duty to inform (Katsh and Rabinovich-Einy, 2017).

Automated Customer Service and Liability: The Air Canada Case and the Legal Binding of Erroneous Information Provided by Chatbots.

In the case “Moffatt v. Air Canada” (2024 BCCRT 149), the British Columbia Civil Resolution Tribunal discussed legal liability for the use of a chatbot that provided incorrect information about funeral fares and reimbursement. The consumer, relying on the information, purchased tickets and had their reimbursement denied. The airline argued that the chatbot was merely an informational tool, without binding character. The tribunal rejected this thesis, recognizing that the system integrates the provider’s communication structure, imputing the informational errors to the organization.

The decision protected legitimate trust and the duty of adequate information, reaffirming objective good faith in the pre-contractual phase. The information influenced the contractor’s will, configuring culpa in contrahendo (art. 422 of the Civil Code/2002) and conduct prohibited by venire contra factum proprium. The case highlighted the risk of algorithmic disinformation and the inadequacy of business strategies to dilute responsibility through technology. In governance, the precedent indicates the need for robust controls over automated service systems, with content review, human supervision, auditable records, and normative updates. In Brazilian law, this aligns with the objective liability of the supplier (arts. 6, III, and 14 of the CDC) and the review of automated decisions (art. 20 of the LGPD). The principles affected were objective good faith, legitimate trust, duty of information, and prohibition of venire contra factum proprium. The algorithmic intervention was one of substitution, with the chatbot taking the place of the human attendant without prior supervision of the response. Factual informational error was identified, not semantic misunderstanding in the strict sense.

RealPage Case: Algorithmic Pricing and Impacts on Contractual Equilibrium

In the case United States v. RealPage, Inc. (No. 1:24-cv-00710, M.D.N.C., 2024), the United States Department of Justice questioned the use of AI-based rental pricing software in the real estate market, within the scope of competition law. The platform collected and processed aggregated market data, including strategic information shared among competitors, to suggest rental prices that maximized revenues. Although formally recommendations, these suggestions exerted relevant influence, reducing decision-making autonomy in price formation and operating as indirect alignment of economic conduct.

The legal consequence was the imputation of possible violation of antitrust norms, on the grounds that the sharing of sensitive data and the use of a common algorithm could produce effects equivalent to concerted price fixing (Ezrachi and Stucke, 2016). In private law, standardized pricing restricts the autonomy of tenants, who do not access the algorithmic logic nor can they ascertain whether the price results from competitive formation or indirect coordination, approaching injury due to exploitation of the condition of necessity (art. 157 of CC/2002) when applied structurally. The principles affected were free competition, free initiative, transparency, objective good faith, and private autonomy. The identified risk was that of algorithmic cartelization, with opaque systems synchronizing economic decisions. As a governance lesson, the case demonstrates the need for algorithm auditing, limits on sharing sensitive data, and competition compliance programs adapted to AI. In Brazil, the conduct could be examined in light of Law No. 12.529/2011.

In a subsequent update (November 24, 2025), the Department of Justice and RealPage entered into an agreement to settle the federal lawsuit without admission of liability, subject to seven-year behavioral restrictions. In parallel, the private multidistrict litigation in Tennessee reached settlements exceeding $359.9 million, also without admission of fault. State actions in Kentucky and New Jersey had their dismissal requests rejected in February and March 2026, with the courts recognizing as plausible the same algorithmic coordination theory that the federal action failed to adjudicate. The outcome is, therefore, fragmented and not consolidated.

Deloitte/DEWR Case: Consultative Elaboration with Generative AI and Responsibility for Fabricated Citations.

In December 2024, the Australian Department of Employment and Workplace Relations (DEWR) engaged Deloitte to review the Targeted Compliance Framework (TCF), an automated system for the application of administrative penalties. The contractual object was a technical-legal report on the legality and regulatory adequacy of the system. Following its publication, researchers identified non-existent academic references, fabricated judicial citations, and material errors. Deloitte acknowledged the use of generative AI, presented a corrected version, and returned part of the remuneration.

Legally, the case evidenced defective fulfillment of an obligation of result. The service was accepted after corrections, with a negotiated reduction in price, in addition to reputational and financial impact for the consultancy. The precedent demonstrates that the “accountability gap” can fall upon the operator itself, who, by adopting generative AI without an output verification protocol, absorbed the costs of the absence of qualified human supervision. The principles affected were objective good faith, transparency, legitimate trust, professional diligence, administrative efficiency, and accountability. The identified risk was that of semantic incomprehension of generative AI: production of legally plausible, but materially false, content. The intervention was characterized as a substitution, with AI taking the place of human legal research and drafting without prior verification of references. As a governance lesson, the precedent shows that the use of AI in technical services with high normative density requires qualified human supervision, integral validation of sources, authorship traceability, disclosure of AI use, and specific contractual controls.

INSS Case: Decisional Automation in Public Social Security Administration

A paradigmatic precedent was issued by the Regional Federal Court of the 4th Region in case No. 5001305-13.2023.4.04.7005 (January 28, 2025), discussing the validity of a social security administrative decision issued in the context of the INSS’s digital service model. The facts stemmed from the denial of a social security benefit after predominantly automated processing, with reduced human intervention. The plaintiff alleged a lack of sufficient motivation and the impossibility of understanding the criteria for the denial. The court recognized the lawfulness of the Administration’s use of technological tools, but affirmed that automation does not remove the constitutional and legal duty to provide motivation: the administered party must receive a clear and intelligible explanation of the reasons for denial, under penalty of invalidity of the decision and violation of due administrative process. In terms of data protection, the hypothesis constitutes the situation described in art. 20 of the LGPD, which guarantees the data subject the right to request a review of decisions made solely based on automated processing.

The principles of legality, motivation, transparency, due process, adversarial principle, broad defense, administrative efficiency, and state accountability were affected. The identified risk was that of decision-making opacity: decisions whose rationale cannot be reconstructed hinder judicial review, technical defense, and accountability for administrative errors. Algorithmic intervention was characterized as predominant automation, with reduced direct human intervention, decision-making opacity prevailing over semantic incomprehension. As a governance lesson, the precedent demonstrates that AI or automation systems used by public authorities must incorporate explainability, effective human review, audit trails, technical documentation, and accessible appeal channels. In state matters, technological efficiency does not replace procedural guarantees.

Comparative Analysis, Analytical Synthesis and Conclusion of Cases.

The comparative analysis of the cases allowed the identification of recurring structural patterns in the use of artificial intelligence systems in legally relevant decision-making contexts. The common characteristic across all cases is that the negative normative effects of automation were only perceived a posteriori, when litigation emerged or the damage had already occurred. This confirms the qualifier “silent”: the transformation of the contract into a predictive-actuarial mechanism occurs under the guise of operational efficiency, with no immediate evidence of the underlying normative compromise.

The first pattern is decisional opacity, present in varying degrees in all cases. The difficulty in reconstructing the algorithmic criteria compromises the verifiability of decisions, affects the duty of motivation, and weakens legal control (Pasquale, 2015; Mittelstadt et al., 2016). Its most intense manifestation occurred in the INSS/TRF-4 case, where the lack of explainability compromised the adversarial process and full defense; the most diffuse, in RealPage, where the opacity of pricing prevented the detection of algorithmic coordination by the users themselves.

The second pattern consists of the partial substitution of normative rationality with inferential rationality based on statistical correlations. These systems operate from patterns extracted from historical data and produce formally consistent results, without guaranteeing adherence to applicable legal criteria (Surden, 2014; Kroll et al., 2017). This was evident in the Deloitte/DEWR case, where generative AI produced legal references that were plausible in form but factually non-existent, and in Air Canada, where the chatbot processed the query without access to the normative meaning of the fare rules.

The third pattern refers to the accountability gap, the difficulty in clearly assigning responsibility when technology-mediated decisions arise from chains distributed among developers, operators, suppliers, and end-users (Bodó et al., 2018). The four cases manifest it in distinct ways: in Air Canada, the company attempted to transfer the risk to the user; in RealPage, responsibility was fragmented between the platform and the administrators; in Deloitte/DEWR, the void fell upon the operator itself; in INSS, the Administration invoked the legality of the system to exempt itself from the duty of motivation.

The fourth pattern consists of the amplification of informational asymmetries, especially when only one of the parties holds knowledge or control over the functioning of the employed system. This restricts contestation, affects transparency, and compromises the balance of legal relationships (Yeung, 2018; Zuboff, 2019).

The fifth pattern concerns semantic incomprehension. Generative AI systems manipulate symbols with high syntactic coherence, but without access to the normative meaning of the elements they process (Searle, 1980; Hildebrandt, 2015). In the Deloitte/DEWR case, this translated into the creation of non-existent citations; in the Air Canada case, into the chatbot’s inability to correctly understand the normative content of the applicable tariff rules.

The recurrence of these five patterns demonstrates that the risks associated with decision-making automation do not stem from episodic failures, but from structural characteristics of these systems, which reveal a permanent tension between the probabilistic logic of AI and the assumptions of the Rule of Law, especially justification, control, and normative interpretation (Hildebrandt, 2015).

Table 2. Distribution of support documents by type and by paradigmatic case [TABLE 2]

Source: Original research data.

Table 3. Confrontation of the four cases by analytical category

Category	Air Canada	RealPage	Deloitte/DEWR	INSS/TRF-4
Algorithmic intervention type	Replacement, the chatbot replaced the human attendant as the official interface, without prior supervision of the response	Support with structural influence, recommendations formally presented as suggestions, but with strong adherence from administrators	Replacement, generative AI has replaced human legal research and writing in drafting report excerpts	Predominant automation, decision generated by automated flow, with reduced direct human intervention
Decisional opacity	Diffuse: punctual failure of the chatbot, not of the general system logic	Intense: pricing logic inaccessible even to its own users (administrators)	Moderate: technical report did not reveal the citation generation process	Maximum: lack of explainability compromised contradictory and broad defense
Substitution of normative rationality	Chatbot processed query without understanding the normative meaning of the tariff rules	Price formation shifted from autonomous decision to coordinated algorithmic inference	Generative AI produced legal references that were plausible in form but non-existent in fact	Automated processing replaced individualized motivation with statistical inference
Accountability gap	Company tried to transfer risk to the user; thesis rejected by the court	Responsibility fragmented between the platform (RealPage) and property management companies	Lack of accountability fell upon the operator itself (the consultancy)	The administration invoked the legality of the system to exempt itself from the duty of motivation
Information asymmetry	Consumer trusted the official channel without means to verify the information provided	Tenants without any access to the algorithmic pricing logic	Contracting Party (DEWR) relied on delivery without a prior source verification mechanism	Administered without access to the criteria that underpinned the denial
Semantic misunderstanding	Chatbot manipulated symbols without access to the normative meaning of the tariff rules	Little evidence of a pattern in this case; risk concentrated in price coordination	Most intense manifestation: creation of legally non-existent citations and precedents	Little evident pattern; risk concentrated in opacity and lack of motivation

Source: Original research results.

Propositions emerging from the confrontation between the cases:

P1: the greater the distance between the system operator and the vulnerable end-user, the greater the probability of fragmentation of the accountability gap among multiple agents (observed in RealPage and INSS).

P2: semantic misunderstanding is manifested with greater intensity in natural language content generation activities (Deloitte/DEWR) than in classification or binary decision activities (INSS).

P3: decision opacity tends to be more easily challenged by the affected party when there is a formal channel for human review provided for by law (LGPD, art. 20 – INSS case) than when such a channel does not exist (RealPage case).

The cases indicate that the incompatibility between AI and law is not absolute: supplier liability, the requirement for administrative motivation, competition control, and contractual sanctions operate as forms of normative absorption of the effects of automation, even if incomplete (Lessig, 2006; Teubner, 2012). However, without decision-making transparency, qualified human supervision, and unequivocally attributed responsibility, the technology compromises legal certainty and converts efficiency gains into institutional risk.

Table 2. Algorithmic Compliance Framework Applied to Contract Management

Category	Air Canada	RealPage	Deloitte/DEWR	INSS/TRF-4
Type of algorithmic intervention	Replacement, the chatbot replaced the human attendant as the official interface, without prior supervision of the response	Support with structural influence, recommendations formally presented as suggestions, but with strong adherence from administrators	Replacement, generative AI has replaced human legal research and writing in drafting report excerpts	Predominant automation, decision generated by automated flow, with reduced direct human intervention
Decision opacity	Diffuse: punctual failure of the chatbot, not of the general system logic	Intense: pricing logic inaccessible even to its own users (administrators)	Moderate: technical report did not reveal the citation generation process	Maximum: lack of explainability compromised contradictory and broad defense
Replacement of normative rationality	Chatbot processed query without understanding the normative meaning of the tariff rules	Price formation shifted from autonomous decision to coordinated algorithmic inference	Generative AI produced legal references that were plausible in form but non-existent in fact	Automated processing replaced individualized motivation with statistical inference
Gap in accountability	Company tried to transfer risk to the user; thesis rejected by the court	Responsibility fragmented between the platform (RealPage) and property management companies	Lack of accountability fell upon the operator itself (the consultancy)	The administration invoked the legality of the system to exempt itself from the duty of motivation
Information asymmetry	Consumer trusted the official channel without means to verify the information provided	Tenants without any access to the algorithmic pricing logic	Contracting Party (DEWR) relied on delivery without a prior source verification mechanism	Administered without access to the criteria that underpinned the denial
Semantic misunderstanding	Chatbot manipulated symbols without access to the normative meaning of the tariff rules	Little evidence of a pattern in this case; risk concentrated in price coordination	Most intense manifestation: creation of legally non-existent citations and precedents	Little evident pattern; risk concentrated in opacity and lack of motivation

Source: Elaborated from ISO 31000:2018; ISO 37301:2021; ISO/IEC 42001:2023; NIST AI RMF 1.0; COSO ERM (2017).

Practical Applicability of the Framework: How to Use the Checklist in the Routine of Contractual Compliance

The framework presented in Table 2 is not just a classificatory instrument, but an operational tool to guide the implementation, monitoring, and auditing of artificial intelligence systems applied to contractual management. Each of the five axes – governance, risk management, operational controls, transparency, and continuous monitoring – corresponds to a concrete stage of the contractual lifecycle, allowing compliance, legal, or technology professionals to identify which control should be triggered.

In practice, the Governance axis corresponds to the phase prior to the contracting of any AI system, requiring the compliance area to approve the responsible use policy and designate an AI Officer, whose absence in the analyzed cases contributed to the materialization of legal risk. The Risk Management axis operates during the contracting or updating of the system, through risk inventory and classification, replicating the criteria of PL 2338/2023 and the European AI Act.

Operational Controls require that any technical-legal delivery generated with the support of generative AI be submitted to documented human validation before transmission to third parties, replicating the lesson from the Deloitte/DEWR case. The Transparency axis mandates informing the contractor that a communication was generated or supported by AI, replicating the understanding from the Air Canada case, which dismissed the thesis of the chatbot as a neutral tool without binding effects.

Finally, Continuous Monitoring should integrate the periodic audit routine, with a record of accuracy, error, and legal impact indicators, allowing for early identification of risk signals that, in the four cases studied, were only perceived a posteriori. The checklist is not a static document, but a decision roadmap to be followed whenever the organization initiates, operates, or terminates the use of AI in a contractual context.

The use of artificial intelligence systems in the automated drafting of contracts gradually compromises the normative rationality of contract law. In the consumer context (Air Canada), the compromise occurs in the pre-contractual phase, due to algorithmic disinformation that frustrates good faith and legitimate trust. In the competitive context (RealPage), due to the opaque substitution of autonomous economic will formation, with a risk of algorithmic cartelization. In the consultative-contractual context (Deloitte/DEWR), due to the emptying of the obligation of result, via semantic incomprehension that produces a formal simulation of legality. In the administrative context (INSS), due to the suppression of the duty of motivation and constitutional procedural guarantees. In all cases, the compromise is silent, initially manifesting as efficiency and revealing its disruptive content only when litigation emerges.

AI-based contractual automation does not constitute mere instrumental innovation, but promotes a structural shift of the contract from the normative domain to the predictive-actuarial domain. The erosion of human oversight, the substitution of legal reasoning for probabilistic inference, and the emptying of protective categories only emerge a posteriori, under the initial guise of operational efficiency gains. Semantic incomprehension, identified as an autonomous pattern, confirms that syntactic correctness does not ensure material compliance with objective good faith, contractual balance, and the social function of the contract, and may constitute defects of consent or abusive clauses. In the Brazilian regulatory sphere, the general regime provides partial protection, with the LGPD ensuring review of automated decisions (art. 20), the CDC imposing strict liability for informational defects, and the Civil Code requiring good faith in all phases of the contract. Organizations with formalized algorithmic governance may be in anticipatory compliance when the Legal Framework for AI is enacted, representing a competitive advantage and liability mitigation. The four precedents analyzed are recent (2024-2025) and were issued by first-instance or limited jurisdiction bodies, which calls for caution regarding their generalization, as illustrated by the RealPage case, which had a federal settlement and parallel state actions that recognized the plausibility of the algorithmic coordination theory. The jurisprudence on algorithmic automation is still in formation, indicating the need for doctrine to systematize criteria to avoid over-generalization or underestimation of the legal risks of contractual automation.

4. Conclusion

The study analyzed how artificial intelligence in contractual automation intensified the collision between technological efficiency and legal certainty, and to what extent mechanisms of significant human oversight and algorithmic governance functioned as instruments of normative balancing to preserve principles of Brazilian contract law. It was found that contractual automation, in distinct institutional contexts, generated five recurring structural patterns: decision opacity, substitution of normative rationality by probabilistic inferences, accountability gap, informational asymmetry, and semantic incomprehension. These patterns emerged systematically in the consumer, competition, consultative-contractual, and administrative cases, confirming that the transformation of contracts into predictive-actuarial mechanisms occurs silently, under the guise of operational efficiency, revealing its disruptive effects only when litigation arises. Semantic incomprehension, identified as an autonomous pattern, confirmed that syntactic correctness does not ensure material conformity with objective good faith, contractual balance, and the social function of the contract, and may constitute defects of consent or abusive clauses.

The study’s main contribution lies in proposing an algorithmic compliance framework, which offers practical parameters for the implementation, monitoring, and auditing of artificial intelligence systems in contractual management, aligning with international standards and mitigating identified risks. However, the analysis of recent precedents (2024-2025), issued by first-instance or limited jurisdiction bodies, calls for caution regarding generalization, as illustrated by the fragmented outcome of the RealPage case, which reinforces the conclusion that jurisprudence on decision-making automation is still developing. It is suggested that doctrine systematize criteria to avoid over-generalization or underestimation of the legal risks of contractual automation, deepening the understanding of the compatibility limits between the probabilistic logic of artificial intelligence and the normative rationalities of law.

Bibliographic References

Busuioc, M. 2021. Accountable artificial intelligence: holding algorithms to account. Public Administration Review 81(5): 825-836.

Coeckelbergh, M. 2020. Al ethics. MIT Press, Cambridge, MA, United States.

Floridi, L.; Cowls, J.; Beltrametti, M.; Chatila, R.; Chazerand, P.; Dignum, V.; Luetge, C.; Madelin, R.; Pagallo, U.; Rossi, F.; Schafer, B.; Valcke, P.; Vayena, E. 2018. Al4People: an ethical framework for a good Al society. Minds and Machines 28: 689-707.

Han, B.-C. 2018. Sociedade do cansaço. Vozes, Petrópolis, RJ, Brasil.

Lessig, L. 2006. Code: version 2.0. Basic Books, New York, NY, United States. Link/DOI: https://codev2.cc/

Mackaay, E.; Rousseau, S. 2020. An economic analysis of contract law. Edward Elgar, Cheltenham, England. Link/DOI: https://www.e-elgar.com/shop/usd/contract-law-and-economics
9781847206008.html

Posner, R. A. 2014. Economic analysis of law. 9. ed. Wolters Kluwer, New York, NY, United States. Link/DOI: https://aspenpublishing.com/products/posner-economic-analysis-of-law-9e

Schreiber, A. 2018. Manual de direito civil contemporâneo. Saraiva, São Paulo, SP, Brasil. Link/DOI:
https://www.grupogen.com.br/

Shavell, S. 1980. Damage measures for breach of contract. Bell Journal of Economics 11(2): 466-490.
Link/DOI: http://www.law.harvard.edu/faculty/shavell/pdf/11_Bell_J_Econ_466.pdf

Wachter, S.; Mittelstadt, B.; Floridi, L. 2017. Transparent, explainable, and accountable AI for robotics.
Science Robotics 2(6): eaan6080. Link/DOI: https://doi.org/10.1126/scirobotics.aan6080

Zuboff, S. 2019. The age of surveillance capitalism. PublicAffairs, New York, NY, United States. Link/DOI: https://www.hachettebookgroup.com/titles/shoshana-zuboff/the-age-of-surveillance
capitalism/9781610395694/

Article originating from the Final Course Work of the Specialization in Compliance and ESG from MBA USP/Esalq

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