Algorithmic pricing is creating cartel-like outcomes in some markets without the meetings, messages or formal agreements traditionally used to prove collusion, raising difficult questions for competition authorities and company boards.
The issue has been highlighted by the US Federal Trade Commission’s antitrust case against Amazon, in which the agency described an internal pricing tool known as Project Nessie. The FTC alleges that the system identified products where rivals were likely to follow an Amazon price rise, increased the price and kept it there once competitors matched.
The agency claims Nessie generated more than $1 billion in additional profit and was paused during periods of heightened scrutiny before being switched on again. Amazon disputes the allegations and says the tool was discontinued years ago.
That case involves a company deliberately attempting to predict and exploit the reactions of its competitors. A more difficult problem arises when separate systems, each pursuing profit independently, learn that they can charge more by avoiding aggressive price competition.
Research into German petrol stations found that margins rose by about 38% in markets where two competing stations adopted automated pricing software. Margins did not change when only one station in a market used the software.
The finding, published in the Journal of Political Economy in 2024, suggested that the increase emerged when both systems were setting prices against one another. Each appeared to learn that it earned more by backing away from price cuts, despite there being no communication or agreement between the businesses.
It was among the first attempts to measure in a real market a phenomenon that had previously been demonstrated mainly through computer simulations. Researchers continue to debate how widely the findings apply, but the results have increased concern about firms handing pricing decisions directly to software.
How algorithmic pricing can weaken competition
There are several ways in which automated systems can produce a similar result to a cartel.
In one scenario, independently deployed algorithms learn over repeated encounters not to undercut each other. No competitor data is exchanged and no person designs the outcome, but prices can remain above competitive levels because each system discovers that restraint is more profitable than a price war.
In another, a single company uses software to anticipate a rival’s response and raises prices where it expects the competitor to follow. Project Nessie, as described by the FTC, falls into this category.
A third arrangement involves several competitors supplying data to a common provider whose software recommends prices to them all. This is the type of conduct regulators have found easier to challenge.
Controlled experiments have offered evidence of how the first scenario might develop. In a 2020 paper in the American Economic Review, four economists set reinforcement-learning algorithms to compete in a repeated pricing model. The systems could not communicate and were instructed only to maximise profit.
They repeatedly learned to charge more than the competitive level and to maintain that position. When one system cut its price to win market share, the others responded with cuts of their own, before returning to higher prices once the deviation ended.
The pattern remained when the firms had different costs or levels of demand and when the number of competitors changed. The authors, later joined by Wharton economist Joseph Harrington, warned in Science that pricing systems could learn collusive rules without human oversight or awareness.
For businesses, the risk is that conventional performance measures may not expose the problem. Higher margins, stable prices and improved conversion rates can appear to demonstrate that a pricing system is working, even when competition has quietly weakened.
Why competition law is struggling to keep pace
Antitrust law has traditionally looked for evidence of an understanding between competitors, such as a meeting, exchange of information or coordinated decision. Independently operating algorithms may create the same economic effect without leaving any of those traces.
The RealPage litigation illustrates the distinction. In 2024, the US Department of Justice and several states sued the property-pricing software company and landlords that used its system. The software recommended rents using information from competing properties.
Under a proposed settlement filed by the Justice Department in November 2025, RealPage would pay no penalty and admit no wrongdoing. The proposed terms would restrict the use of recent competitor data and detailed local information, while also providing for a court-appointed monitor. The settlement requires court approval and the wider litigation continues.
The case is more straightforward for regulators because a common provider was accused of bringing competitors’ non-public data together to produce recommendations.
Two appellate rulings involving the same vendor’s software illustrated the legal boundary. The Ninth Circuit dismissed a case against Las Vegas hotels because their confidential data had not been pooled. The Third Circuit later revived a similar case involving Atlantic City casinos, where competitors had supplied non-public data to a shared system and followed its recommendations about nine times in ten.
That difference — shared competitor information on one side and independent use of a similar tool on the other — may determine whether authorities can establish unlawful coordination. When systems reach comparable prices using only information they can observe publicly, there may be no common provider or obvious agreement to challenge.
Lawmakers have begun responding to the risks, particularly in housing. San Francisco introduced a ban on algorithmic rent-setting tools in 2024, followed by Philadelphia, Minneapolis and Seattle. New York enacted the first statewide ban in October 2025, while California amended its antitrust law the same month.
RealPage sued New York after the Justice Department settlement, arguing that its pricing recommendations amounted to lawful speech protected by the First Amendment. The legal questions are expected to take years to resolve.
For companies, however, responsibility does not disappear simply because conduct was generated by software or supplied by an outside vendor. Exposure may take the form of regulatory or reputational risk even where existing legal categories do not clearly apply.
What companies are being urged to examine
The move from advisory software to systems that act directly has also changed where accountability lies. Previously, software might recommend a price before a person approved it. More autonomous systems can pursue an assigned objective repeatedly, changing their strategy without waiting for individual authorisation.
The decision has not vanished. It is contained in the objective given to the system, the information it is allowed to use and the restrictions imposed by the company.
Boards and executives are therefore being urged to examine behaviour as well as financial results. They should establish whether pricing systems can observe competitors’ prices in real time and whether they react directly to those signals.
Systems relying mainly on internal information, such as costs, demand and stock levels, may carry less risk, although competitor behaviour can still affect them indirectly through changes in demand. Many companies have not mapped which of their systems can see rivals’ prices.
Businesses can also test whether improved margins depend on the ability to shadow competitors. One approach is to replay market conditions with delayed competitor signals, randomised response times or no competitor-price information. If margins collapse only when the system loses access to rival prices, that may indicate that the gains require further investigation.
Such measures are not a guarantee that coordination will be prevented. Differences between systems and a larger number of competitors may make coordinated outcomes harder to sustain, but they do not remove the risk.
Companies are also being advised to audit the system’s conduct rather than simply its results. That means recording what it was designed to optimise, which signals it used, when it altered its strategy in response to a competitor and what changes were made to its pricing policy.
The mandate should be auditable, including the actions the system is forbidden to take and the data it may access. Where prices emerge from rules rather than individual human decisions, those rules are effectively the company’s decision.
A board that cannot explain why prices across a market have converged, beyond pointing to an algorithm, may have delegated a competition-sensitive decision without defining the limits it intended to impose.
