In its antitrust suit against Amazon, the Federal Trade Commission described a pricing tool internally named Project Nessie. The system identified products where competitors were likely to follow an Amazon price increase, raised the price, and held it once rivals matched. The agency alleges the tool generated more than $1 billion in excess profit — and that Amazon paused it during periods of heightened scrutiny, then switched it back on. Amazon disputes this and says the tool was discontinued years ago.
That is the deliberate version of this problem: a company designing a system to anticipate rivals. The harder version is the one nobody designs at all. In 2017, when automated pricing software became widely available to German gas stations, economists later found that in markets where two competing stations both adopted it, margins rose by about 38% — with no meeting, no message, and no agreement between them. Market-level margins didn’t move at all when only one station in a market adopted the software. The rise appeared only when two algorithms were left to set prices, in effect, against each other, a pattern consistent with each one learning on its own that it earned more by backing off.
That study, published in the Journal of Political Economy in 2024, is among the first real-world measurements of a problem previously shown mostly in simulation.
Pricing algorithms can produce the economic outcome of a cartel, meaning higher prices sustained over time, without the conduct antitrust law was written to detect. It matters for any company that has handed pricing to software, because the behavior may not appear on the dashboards used to judge whether the software works.
Executives usually judge competition by the pressure they feel, and a market where prices hold and margins stay comfortable reads as one they have won. Automated pricing breaks that instinct. When autonomous agents set prices, the same calm picture can mean the opposite, a sign that competition has quietly stopped because the algorithms have learned that leaving each other alone pays better than fighting.
The failure that should concern leaders is subtle. An algorithm that sets an obviously wrong price is easy to catch. The harder case is one that does exactly what it was designed to do, optimize margin, and reaches an outcome the company would struggle to justify in public.
Three ways competition quietly disappears
Competition can fade in more than one way. Independently deployed algorithms, each pursuing its own profit, can learn over repeated encounters to stop undercutting one another, with no one designing the outcome and no data changing hands.
Call it the ghost: no agreement, no data exchange, no one who designed it — just two systems that arrived at the same truce independently.
A single firm can instead use software to anticipate how rivals will react, raising a price where it predicts they will follow, a unilateral strategy rather than a pact.
Call it the mirror: Amazon’s Nessie belongs here — no pact, just a system built to predict a rival’s reflection and act first.
Or competitors feed their data into a common provider whose algorithm guides them all, the pattern enforcers find easiest to challenge.
Call it the hub: RealPage is the textbook case, and it’s the only one of the three regulators have actually managed to touch.
The first is this article’s subject, the hardest to see and hardest for the law to reach.
The clearest evidence comes from controlled experiments. In a paper published in the American Economic Review in 2020, four economists set reinforcement-learning algorithms to compete in a standard model of repeated pricing. The algorithms could not communicate and were told only to maximize profit. They consistently learned to charge above the competitive level, and to enforce it. When one lowered its price to gain share, the others cut theirs, then returned to the higher level once it fell back into line. The pattern held even when firms differed in cost or demand and when the number of competitors changed.
The four authors, joined by Wharton economist Joseph Harrington, set out the policy stakes in Science later that year. They warned that delegating pricing to algorithms opens a backdoor to collusion, since AI can learn collusive rules with no human oversight or awareness. Harrington has argued that competition law must be rethought for coordination that arises without agreement.
The German gasoline data indicates that this happens in practice and not only in a model. Not every experiment reaches the same conclusion, though, and researchers still debate how readily these results carry over to live markets. That uncertainty is itself a reason for boards to watch behavior now, rather than wait for regulators to settle the question for them.
Why the law struggles with this
Antitrust enforcement was designed around human agreement, evidence of a meeting or understanding between competitors. Coordination a machine learns on its own provides none of that, which is why even the most prominent recent case, built around a shared vendor, proved so hard to resolve.
In 2024, the Department of Justice and several states sued RealPage, whose software recommended rents using data from competing properties, along with landlords that used it. In November 2025 the DOJ filed a proposed settlement. RealPage paid no penalty and admitted no wrongdoing. The terms mainly restrict the data the software may draw on — barring recent competitor data and the fine-grained local geography that made neighborhood-level coordination possible — and install a court-appointed monitor. The settlement still needs court approval, and the wider litigation continues.
RealPage is the easier case, a common provider pooling competitors’ nonpublic data into one recommendation. The harder case begins when independently deployed systems reach the same result using nothing but the prices they can all observe. There is no hub to point to, and nothing that resembles a meeting.
Two recent appellate rulings, both involving the same vendor’s software, drew this line for us. The Ninth Circuit dismissed a case against Las Vegas hotels because the tool did not pool their confidential data. A year later, the Third Circuit revived a near-identical case against Atlantic City casinos, where competitors did feed nonpublic data into the shared system and followed its output about nine times in ten. Pooled competitor data on one side and independent use of the same tool on the other is the boundary between RealPage and the harder case.
Legislators have not waited, either. Starting with San Francisco in the summer of 2024, cities including Philadelphia, Minneapolis and Seattle banned algorithmic rent-setting tools. New York enacted the first statewide ban in October 2025, and California amended its antitrust law the same month. Days after its DOJ settlement, RealPage sued New York over its ban, casting its pricing recommendations as lawful speech protected by the First Amendment. These questions will take years to resolve, but the practical conclusion is available now. When coordination is learned rather than agreed, the legal categories may not apply, yet the exposure remains. It shifts toward reputational and regulatory risk and falls on the company that deployed the system and set its objective; that responsibility cannot be outsourced to the vendor.
This also shifts responsibility inside the firm. For a decade, pricing software advised and a person decided, which kept accountability clear. Agentic systems act directly, pursuing an assigned objective transaction after transaction, adjusting without waiting for approval. The decision still exists. It has moved into the objective the company set and the limits it chose not to set.
The question leaders skip
Most pricing teams judge their systems on performance. Margins and conversion improve, and the software is called a success. But a coordinated market and a competitive one produce the same figures, so those metrics cannot reveal the risk. The sharper question is behavioral. What has the system learned about competitors, and would the company defend that behavior to a regulator, or to customers who found that rival suppliers had somehow stopped undercutting one another?
A board that cannot explain why prices across its category have converged, beyond pointing to the algorithm, has delegated a decision it never intended to make.
What leadership can do now
Turning the systems off is neither realistic nor necessary. The task is to govern what they are permitted to learn, and the research points to several measures.
The first is to establish where the systems can observe competitors. A pricing agent that reacts to a rival’s price in real time has the input coordination needs. One that relies on internal signals such as cost, demand, and inventory carries lower risk, though competitor behavior can still reach it indirectly through demand. Many companies have never mapped this and cannot say which systems can see competitor prices.
The second is to introduce constraints that make coordination harder to sustain. Some evidence suggests it is more fragile when competing systems differ from one another or face more rivals.
So leaders should treat these steps as risk reduction rather than a guarantee.
The third is to audit behavior rather than results alone. Reviewing only financial performance will not detect this. A board should ask for a clear account of what the system optimized, which signals it relied on, and where it changed strategy in response to a competitor, treated with the seriousness the audit committee applies to financial conduct.
The fourth is to run a counterfactual competition test. Management can periodically replay market conditions under altered settings, such as delayed competitor signals, randomized response times, or no competitor-price input. If margins hold, the gains are more likely to be the company’s own. If they collapse only when the agent can no longer shadow rivals, the board has identified a reason to investigate. Though technically demanding, such tests can help distinguish value creation from faded competition.
The fifth is to require an auditable mandate. Management should document what the system was told to pursue, what it was barred from doing, the data it may use, and every material change to its pricing policy. When prices emerge from rules rather than from individual decisions, those rules are the decision. A company that never defined when responding to competitors becomes impermissible has, in practice, left that line to the algorithm.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
