When AI Prices, Who Is Liable? A Brazilian Perspective

A New Enforcement Frontier

In recent weeks, the Brazilian antitrust authority - the Administrative Council for Economic Defense (CADE) - has led two significant developments signaling a new frontier in antitrust enforcement: scrutiny over the use of pricing algorithms and artificial intelligence (AI) tools as potential facilitators of anticompetitive conduct. What makes these cases particularly significant is not each investigation in isolation, but the pattern they reveal together: CADE is actively building an enforcement framework for algorithmic coordination, one that addresses both the conduct of companies that use pricing tools and the obligations of the third parties that supply them. Building on comments from the CADE official handling the Aprix case at a recent event, they allow for the identification of preliminary guidelines regarding how the authority is approaching these issues.

Investigation Against GOL and Latam

Administrative Proceeding No. 08700.007894/2023-88

In April 2026, CADE’s General Superintendence (SG) opened an investigation against GOL and Latam for alleged conscious parallelism in pricing in the domestic air transportation market, with explicit indications of concerns regarding the use of AI and machine learning instruments. The SG is investigating whether the companies indirectly shared competitively sensitive information, thereby making their pricing strategies mutually observable, even without direct communication between them. According to the SG, a third-party company allegedly served both airlines simultaneously, collecting, processing, and potentially redistributing fare data, with the airlines having contractually authorized unrestricted access to their real-time fares and the use of such information in services provided to other market participants.

Based on the public information currently available, the cartel investigation targets the contracting companies (GOL and Latam). The third-party company, whose identity remains confidential, is not currently under investigation, although it is identified as an essential link in the implementation of the alleged conduct. The technical note opening the investigation refers to a possible theory of culpa in vigilando, that is, the possibility of holding companies liable for negligence in supervising the collusive impact of their contracted algorithms.

Settlement Agreement with Intelprice / Aprix

Administrative Proceeding No. 08700.006280/2024-60 and Settlement Agreement No. 08700.010553/2024-71

Also in April, CADE's Tribunal entered into a Cease-and-Desist Commitment Agreement (Termo de Compromisso de Cessacao, TCC) with Intelprice Solucoes de Precificacao Ltda., developer of the Aprix software, in an investigation concerning anticompetitive practices in the Brazilian fuel retail market.

According to the case records, Intelprice, through Aprix (an AI-based pricing solution), collected and processed competitors' pricing data, historical sales data, and operational information from user gas stations, and allegedly used advertising material and communications containing potentially anticompetitive language, instructing clients to resist price reductions and warning about the risks that individual pricing decisions could pose to the market as a whole. Notably, the SG's attention to this marketing language indicates that regulatory scrutiny extends beyond the algorithmic conduct itself to how pricing tools are marketed and positioned, a signal relevant both for companies evaluating such tools and for the providers themselves.

Under the agreement, Intelprice committed to cease the practice, pay a fine, and undertake the following additional commitments: (i) adoption of a contractual clause with its clients requiring the confidentiality of information and data regarding third parties; (ii) implementation of an antitrust compliance program; (iii) granting CADE access to its facilities and allowing monitoring through external audits, to be funded by Intelprice itself; and (iv) notification to CADE if it begins providing services to companies accounting for 20% of a given market (in this case, a municipality). The rationale behind this parameter would be to assess Aprix’s potential influence over pricing dynamics – a provider may hold a relatively small market position while still exerting significant influence over prices within a given geographic area. According to the CADE official, this threshold may vary upward or downward depending on the market and should not necessarily be treated as a benchmark for other cases.

Key Considerations Regarding CADE's Approach

The two cases, read alongside the CADE official's considerations, allow for several preliminary observations that are already shaping how companies should think about their algorithmic pricing tools, even before CADE's enforcement framework fully consolidates.

The Conduct: Scope, Risk Factors, and Legal Characterization

Although other types and uses of algorithms may become subject to investigation, CADE's current priority focus is on pricing algorithms and pricing mechanisms. In both cases, the databases feeding these tools contained not only public information or information belonging to the economic agent itself, but also competitors' confidential data, which appears to be a central characteristic in the identification of competitive risk. That characteristic may also develop into a more severe legal treatment as certain forms of algorithmic coordination might be analysed as a by-object infringement, that is, as conduct whose very object is anticompetitive without requiring proof of actual market effects.

The Structure: Intermediaries and Architectural Risk

Both cases involve, to varying degrees, the participation of service providers external to the affected market, acting as a coordination link or facilitator of information sharing. Importantly, the fact that the third party is not itself a competitor in the affected market does not reduce the antitrust risk for the contracting companies: the intermediary role in enabling the flow of competitively sensitive information is sufficient to raise competitive concerns. Alongside who is involved, the architecture of the tool itself is equally relevant. The CADE official emphasized that mere price suggestion tends not to be problematic, provided that companies remain independent in deciding the prices actually charged. On the other hand, tools that automatically determine and publish prices set by the AI or algorithm represent the highest-risk configuration. Along similar lines, the CMA has already confirmed, in the context of automated price-monitoring tools used by online sellers, that deploying such systems provides no liability shield for the companies using them, and the same principle applies where the automation runs through a shared third-party platform. The existence of an additional step involving human decision-making or a confirmation button allowing the company to define the price actually charged may function as a mitigating factor.

Enforcement: Evidence and Liability

Currently, economic evidence has been used to demonstrate the existence of the conduct itself, rather than to verify or measure its potential effects. This reflects a practical difficulty: the opacity of algorithmic systems makes it hard for regulators to establish what the tools were doing, for whom, and with what information. The CADE official acknowledged this directly, indicating that some form of reversal of the burden of proof is being considered for future cases. That same difficulty shapes the liability theories being developed in parallel: in the airlines case, the SG raised the possibility of holding companies liable for negligence in supervising the collusive impact of their algorithms.

The possibility of requiring confidentiality and non-data-sharing clauses was discussed, although the practical difficulty of verifying effective compliance by suppliers was acknowledged; the CADE official compared the issue to anti-corruption clauses, suggesting that supplier audits could be an alternative, while noting that such an approach might be excessive. In any event, the airlines case suggests potentially greater exposure for companies directly active in the affected markets as compared to the Aprix case, insofar as the conduct is being directly attributed to them, even though the service provider appears as an essential link in the alleged conduct. The Aprix TCC is also relevant in this regard: in addition to the cessation of the conduct and a fine, the agreement outlines a robust set of obligations, including confidentiality clauses, compliance measures, external audits, and notification duties to CADE, which may serve as a reference for future agreements.

Practical Implications

The cases together send a message: antitrust exposure through algorithmic pricing does not require direct participation in a cartel. Liability can arise from the tools a company contracts, the data those tools access, and the degree of oversight (or lack of it) exercised over the supplier's practices. This shifts the compliance question from whether the company is fixing prices to whether it knows what its pricing software is doing with competitors' data, and whether it can demonstrate what it does.

Concretely, companies should assess whether the databases feeding their pricing tools contain competitors' confidential information, and whether supplier agreements include confidentiality, non-sharing, and data use restrictions. The architecture of the tool itself matters: a human decision point before prices are implemented might be a meaningful mitigating factor under the current framework, not merely a procedural formality. Compliance programs that have not yet been updated to account for algorithm-specific risks, including supplier audit mechanisms analogous to those used in anti-corruption matters, may already be operating behind the curve.

Marketing and commercial communications from pricing solution providers also warrant attention. The SG's scrutiny of Aprix's advertising materials signals that how a tool is marketed is itself a relevant factor and may increase companies’ exposure.

Conclusion

These recent decisions indicate that algorithmic coordination is no longer a theoretical concern, and the Brazilian competition authority is actively developing the tools, including evidentiary standards, harm theories, and settlement frameworks, to address it. Companies that have not yet assessed their exposure may be at risk. The question is not whether CADE will continue developing this enforcement agenda, but how quickly, and how far, it will extend.