Few antitrust investigations find their theory of harm laid out in the target’s sales brochure. Brazil’s investigation of Aprix, a startup that sells pricing software to gas stations, nearly managed the feat. One of the company’s promotional brochures introduced prospective clients to the prisoner’s dilemma, the classic game-theory example in which individually rational choices can leave everyone worse off. It explained how a price war could produce just that result for rival stations, then posed the sales pitch as a question: “Which pricing decision increases the company’s profit with the lowest risk that the whole market ends up earning less? It is to answer this question that our pricing technology exists.”
Aprix’s webinars struck the same theme. They urged clients not to chase sales volume through discounts: “Lowering price almost never pays off. Resist!” A gas station that “attacks the market,” Aprix warned, would invite retaliation until “everyone loses together.”
The brochure and a handful of promotional videos helped prompt Brazil’s leading algorithmic-pricing case. In April 2026, the Tribunal of the Administrative Council for Economic Defense (CADE) approved a settlement (known in Portuguese as a Termo de Compromisso de Cessação (TCC)) with Aprix. The case produced neither a litigated finding of collusion nor a condemnation on the merits. Still, the settlement offers a concrete guide to how CADE may handle future algorithmic-pricing cases. Aprix is not CADE’s only such case, and Brazil’s debate over whether pricing algorithms can facilitate collusion remains far from settled.
Today’s Special: Public Prices, Private Data
The investigation began with a news article. On Jan. 20, 2021, the Brazilian outlet Jornal do Comércio reported that a gaúcha startup—one based in the state of Rio Grande do Sul—had developed a pricing system for fuel retailers. The next day, CADE received an anonymous complaint attaching the article. CADE’s General Superintendence (SG), its investigative arm, opened an inquiry. Over the next three years, CADE gathered evidence about how the software worked and asked its Department of Economic Studies (DEE) to assess the conduct.
Aprix’s product combined a client gas station’s prices, costs, and sales volumes with competitors’ pump prices collected from public sources. It then recommended a daily price for each station. Aprix monitored 12,000 stations by October 2020 and more than 20,000 by June 2022.
The software drew on two distinct layers of information. The first was public but scattered. Aprix used an automated tool to collect daily pump prices from what the company described as open sources. It consolidated prices from thousands of stations across Brazil into a single database and applied statistical filters intended to ensure that each figure reflected the price actually charged at the pump.
The second layer consisted of each client’s private historical data, including its sales, costs, and volumes. Aprix collected that information automatically by integrating its platform with the station’s management software.
The algorithm analyzed both layers and recommended a daily price. The station manager retained the final say and had to approve the recommendation.
Each type of information alone would be unlikely to create an antitrust concern. Collecting competitors’ publicly available prices is routine competitive intelligence, and every firm may analyze its own business data. Their combination made the case more difficult. A single vendor held a curated, nationwide database of pump prices while also receiving private data from numerous competing clients. The investigation therefore asked whether that arrangement simply helped each station make better independent decisions or allowed rivals’ prices to align through a common intermediary.
Statistically Significant, Legally Inconclusive
The economic record centers on Technical Note No. 16/2024/DEE, a difference-in-differences (DiD) study. This method estimates an intervention’s effect by comparing changes over time in an affected group with changes in a similar, unaffected group.
The DEE analyzed weekly retail-price data collected by Brazil’s National Agency of Petroleum, Natural Gas and Biofuels (ANP) between January 2015 and May 2022. It compared Aprix clients, known as the treatment group, with other stations in the same municipality, which served as the control group. Because the DEE lacked precise adoption dates, it used early 2020 as a proxy—the year Aprix reportedly expanded its client base by 500%.
The study estimated that Aprix adoption was associated with gasoline and ethanol price increases of roughly R$0.02 to R$0.03 per liter, or about US$0.0038 to US$0.0058. Those amounts equaled approximately 0.5% to 0.8% of the pre-adoption gasoline price and 0.7% to 1.1% of the pre-adoption ethanol price. For diesel, the estimated increase ranged from R$0.01 to R$0.04 per liter, or about US$0.0019 to US$0.0077, equivalent to 0.3% to 1.3% of the pre-adoption price.
The estimated increase, then, was generally around 1% or less of pre-adoption prices. It was smaller still when compared with the higher prices that prevailed between 2020 and 2022. The study found no consistent effect for compressed natural gas (GNV). Its estimates changed direction across model specifications, which the DEE attributed to a small sample riddled with gaps.
Although the DiD estimates were statistically significant, the DEE described their magnitude as “low.” It also cautioned that, like any empirical model, the analysis “depends on a series of premises that, if altered, can significantly influence the results.” One of those premises is the parallel-trends assumption: Without Aprix, prices at client and nonclient stations would have followed similar paths.
The DEE also rejected the inference that a statistically significant price difference settled the legal question. Higher average prices at Aprix stations “does not imply causality,” it warned, because “other factors can influence the prices charged.”
The study estimated the price effect associated with adopting Aprix’s software but like most economic evidence, it did not directly prove an antitrust violation—i.e., coordination among competitors. The DEE found a modest price increase and identified a risk of coordination. It neither established a cartel nor claimed to have done so.
When the Algorithm Suggests—and When It Schemes
Algorithmic pricing has generated a substantial—and unsettled—antitrust literature. Its starting point is straightforward: Pricing software is not inherently suspect.
Software that tracks public prices and adjusts prices as conditions change can expand output, clear inventory, and help smaller firms compete with sophisticated incumbents. Treating such tools as presumptively collusive would amount to a soft form of price control, chilling innovation that can benefit consumers. (See Alden Abbott, “Legal Challenges to Algorithmic Pricing” and Mario Zúñiga, “A Primer (and Some Questions) About the RealPage Antitrust Case.”) Firms generally may collect competitors’ public prices and use them to set their own. That conduct alone does not violate antitrust law.
The analysis changes when software incorporates nonpublic, competitively sensitive information or helps competitors reach an agreement. The legal dividing line does not depend solely on the technology, a common software vendor, pooled data, or parallel prices. The central questions are whether competitors have relinquished independent control over pricing and whether the system provides a concrete means of coordination.
The International Center for Law & Economics (ICLE) recently summarized three scenarios in its comments on the U.S. Department of Justice (DOJ) and Federal Trade Commission (FTC) guidance on business collaborations:
First, competitors may agree to use a shared pricing algorithm to coordinate prices. In that scenario, the algorithm itself is legally irrelevant; the unlawful agreement is the antitrust violation, and existing law already addresses it.
Second, competing firms may each share competitively sensitive information with a common third-party platform that generates pricing recommendations for all participants. In that circumstance, the platform can function as a coordinating mechanism because each participant receives recommendations informed by rivals’ nonpublic data, even absent direct communication among competitors.
Third, independent algorithms may arrive at similar pricing outcomes simply because they respond to the same market conditions and economic signals. That form of conscious parallelism is not unlawful under U.S. antitrust law absent an agreement among competitors.
As ICLE argues in its amicus brief in Cornish-Adebiyi v. Caesars, competitors’ use of the same pricing algorithm does not establish the horizontal “rim” required for a hub-and-spoke conspiracy. That theory requires a central coordinator (the hub) and an agreement connecting the competing firms (the rim). The relevant question is whether the software’s users plausibly struck an agreement with one another, rather than simply making parallel decisions to buy the same product. On this account, “automating lawful commercial activity does not make that activity unlawful,” and courts should not infer an agreement “from the mere fact that competitors happen to use the same commercially available software.”
Herbert Hovenkamp frames the inquiry somewhat differently. He focuses on what the parties agreed to do with the algorithm’s recommendation. Section 1 of the Sherman Antitrust Act, he observes, “does not specify who must do the agreeing.” A hub-and-spoke price-fixing arrangement may therefore arise when each seller separately agrees with the software provider to charge its recommended price, even if the sellers never communicate directly. For Hovenkamp, the clearest cases involve subscribers that commit to following the recommendations or providers that otherwise control subscribers’ prices.
In a recent paper, Hovenkamp and Thibault Schrepel draw the line at “functional control over implementation,” exercised “through delegation, defaults, monitoring, or inducements.” Data aggregation, correlated prices, and conscious parallelism may warrant scrutiny or provide supporting evidence of collusion. They do not, on their own, amount to cartel management or justify the automatic condemnation reserved for arrangements that replace independent pricing decisions.
The two approaches agree on a basic limit. A recommendation without commitment is insufficient, as are parallel prices without a coordinating mechanism. Aprix is best assessed against that common benchmark.
The distinction also matters as a matter of economics. A price increase implemented after a firm adopts an algorithm does not reveal what caused the increase. Recent economic modeling suggests that algorithmic pricing can produce procompetitive effects in markets where available capacity is sold and replenished over time. Sharing information about capacity use can intensify competition at each level of remaining capacity, while high utilization can reduce the potential gains from coordination.
Higher prices at Aprix stations could therefore reflect more precise independent pricing rather than collusion. That possibility explains why the DEE’s caution about causation deserves more than fine-print status.
No Verdict, Plenty of Clues
Because the case ended in a settlement, CADE’s Tribunal never issued a final decision on the merits. The Aprix investigation therefore does not definitively establish CADE’s position on algorithmic pricing. Still, the proceedings and public record offer several clues about how the agency views its competitive risks.
Aprix accepted several obligations for a two-year period. It agreed to pay R$70,803.40 (US$13,604.53) into a federal public fund, a payment similar to a fine. It also agreed to establish a competition-compliance program, segregate client stations’ pricing and strategy data, add confidentiality provisions to client contracts, refrain from requiring clients to adopt its recommended prices, and notify CADE whenever a group of clients reached a 20% share of a municipal market. CADE also received audit access, and Aprix provided the agency with the algorithm’s source code.
The settlement did not require Aprix to redesign its algorithm. Instead, the obligations governed how Aprix handled client data and how stations used the software’s recommendations. That choice suggests that CADE’s principal concern lay with control and information flows, rather than the mere use of automated pricing.
Aprix sits somewhere on the easier side of the spectrum of possible algorithmic-pricing cases. Its software influenced pricing but, at least formally, left the final decision to each station. It relied on competitors’ public prices rather than their confidential sales volumes or margins, combining that public information with each client’s private business data. The apparent concern arose from the arrangement as a whole: One provider processed data from numerous competing stations while marketing its product with conspicuous warnings against cutting prices.
International readers should also note an important feature of Brazilian law. The SG did not need to prove a traditional cartel agreement. It charged the conduct under Article 36, Paragraph 3, Item II, of Brazil’s Competition Law, which prohibits “influencing the adoption of uniform commercial conduct.” That standalone offense does not require the same proof as a horizontal cartel. It can therefore allow CADE to find an infringement without establishing an express agreement among competitors. Whether that lower threshold makes for sound competition policy is a separate question.
Same Fares, Different Flight Plan?
Aprix is unlikely to be CADE’s last algorithmic-pricing matter. A pending airline case more directly tests the difference between lawful parallel pricing and technology-assisted coordination.
In November 2023, the SG opened an administrative inquiry into the Brazilian carrier GOL Linhas Aéreas S.A and the Brazilian subsidiary of Chile’s LATAM Airlines. The inquiry concerned identical or nearly identical fares on overlapping domestic routes. In April 2026, the SG converted the inquiry into a formal administrative proceeding to examine whether pricing tools or market-data platforms could explain the pattern. The proceeding remains pending.
GOL and LATAM deny the allegations and maintain that they set fares independently. The case must therefore determine whether the parallel fares resulted from separate responses to the same market conditions or coordination facilitated by pricing technology.
That makes the GOL-LATAM matter harder than Aprix. It has no common software vendor at its center. At least on the public record, it features parallel fares on overlapping routes without an obvious mechanism connecting the airlines’ pricing decisions.
Other competition authorities have examined similar questions without treating algorithmic pricing as an offense in itself. In 2023, the Italian Competition Authority (AGCM) opened a market investigation into airlines’ pricing algorithms on routes serving Sicily and Sardinia. The inquiry examined how revenue-management systems adjust fares over time, how those adjustments affect competition, and whether airlines personalize prices for individual customers.
The AGCM closed the investigation in December 2025 without finding that pricing algorithms had produced collusive outcomes. Its concerns were more prosaic, though hardly trivial: whether consumers could understand and compare the prices they saw.
The Algorithm Jury Is Still Out
Aprix is now Brazil’s leading algorithmic-pricing case, though it arguably raised more questions than it answered. It did not establish that pricing algorithms are illegal per se. Nor did it determine whether algorithmic coordination should face the near-automatic condemnation applied to traditional cartels or an analysis of its actual competitive effects.
The pending GOL-LATAM case presents different facts. It concerns two direct competitors rather than conduct centered on a common pricing vendor. Still, it raises the same basic question: How should CADE distinguish independent pricing decisions from coordination facilitated by technology?
Aprix leaves CADE with several unresolved questions. What evidence is enough to establish an infringement? Does liability turn on aggregating scattered public prices, processing clients’ private data, or placing both types of information in one vendor’s hands? When does algorithmic coordination amount to an old-fashioned cartel, and when does it require a different analysis?
As pricing software becomes more common, CADE must separate efficiency-enhancing uses from collusive ones. Competition authorities worldwide face the same task. An algorithm can sharpen competition or soften it. The hard part is telling which is which.
