More
    Autonomous FinanceWhen Two Pricing Agents Collude Without a Word

    When Two Pricing Agents Collude Without a Word

    Categories

    Quick Answer

    Two pricing algorithms don’t need a phone call to end up fixing a price. When each firm hands its pricing decision to software that learns from the other side’s moves — a reinforcement-learning agent, a shared vendor’s repricing engine, or a large-language-model agent given a profit goal — the two systems can drift into a stable, above-competitive price without either owner writing an instruction that says “raise prices together.” Regulators call this algorithmic collusion pricing, and the two live legal fault lines are whether a shared vendor (the RealPage theory) or a documented signal (the Eturas theory) turns that drift into an actionable “agreement” under antitrust law. Independent, uncoordinated learning that never touches a common vendor or a shared data feed remains extremely hard to prosecute today, even when the outcome looks identical to a cartel.

    Why Pricing Collusion Suddenly Needs an AI Chapter

    Retail gasoline stations, apartment landlords, airline seats, rideshare fares, and now a growing share of SaaS subscription tiers are priced by software that adjusts continuously, sometimes dozens of times a day, in response to a competitor’s last move. That is not new — dynamic pricing has existed since airlines built the first yield-management systems in the 1980s. What changed is who is doing the adjusting. A human revenue manager used to sit between the data and the decision, and that person’s judgment call was the thing antitrust law could examine for signs of a wink-and-nod arrangement. Strip the human out and replace them with an autonomous agent that is rewarded purely for maximizing margin, and you get a system that can rediscover the logic of a cartel purely through trial and error, with nobody ever proposing it.

    The shift accelerated for three concrete reasons. First, cloud-based revenue-management vendors now serve enormous swaths of a single market from one codebase — RealPage’s software has been used to help set rents on millions of US apartment units, and a handful of gasoline-pricing vendors cover a large share of German filling stations. When many competitors run the same engine, the line between “independent pricing” and “shared infrastructure” gets blurry fast. Second, reinforcement learning matured to the point where academic researchers could run head-to-head pricing bots in simulation and watch them settle on supra-competitive prices without any collusive instruction in their code. Third, large-language-model agents are now being deployed as general-purpose commercial agents — quoting freight rates, setting ad bids, negotiating vendor contracts — and nobody has fully mapped what happens when two such agents from competing firms are both trained to “get the best deal for my principal” in the same market.

    None of this requires malice. A pricing agent that observes a rival undercutting it, retaliates by cutting further, and then learns that retaliation is costly for both sides, can arrive at “don’t start price wars” as a stable strategy purely through repeated exposure to the same game — the same logic a human oligopolist reaches through years of market experience, compressed into a few million simulated rounds.

    Four Ways Algorithms Slide Into Collusion

    The OECD’s 2017 competition-policy paper on algorithms and collusion, still the most cited taxonomy in this area, sorts every real-world scenario into four buckets. They carry very different legal exposure, and mixing them up is the single most common mistake in this space.

    1. The Messenger

    Here the algorithm is just a fast fax machine. Human executives already agreed to fix a price, and software executes the deal — monitoring for cheating, adjusting instantly if a cartel member strays. The 2015 US v. David Topkins case is the textbook example: poster and wall-art sellers on Amazon programmed pricing software to detect competitors’ listings and hold matching prices, after first agreeing by email to divide up the market. The Department of Justice’s criminal antitrust prosecution treated the code exactly as it would a fax confirming a price-fixing call, because a human agreement came first. This bucket is unambiguously illegal and has been since long before software existed.

    2. Hub-and-Spoke

    Several competitors (“spokes”) each plug into the same third-party pricing vendor (“hub”), and the vendor pools their nonpublic data — occupancy rates, lease terms, win-loss history — to recommend prices that keep everyone’s margins up. No spoke ever emails another spoke. The Department of Justice’s August 2024 civil suit against RealPage rests on this theory: landlords across dozens of metro areas fed private lease data into RealPage’s algorithm, which then recommended rent levels, and the DOJ alleges the arrangement functioned as a facilitated cartel even though tenants never saw a single explicit agreement between landlords. Several large landlord defendants settled with the DOJ and state attorneys general in 2024 and 2025 rather than litigate the theory to a verdict, while the case against RealPage itself continued into 2026.

    3. The Predictable Agent

    Each firm runs its own independent algorithm, but the algorithms are simple and transparent enough — for example, “always match the lowest listed competitor price plus two percent” — that rivals can reliably predict each other’s response and settle into a stable, elevated price without ever exchanging data. This is closer to old-fashioned conscious parallelism, a legal gray zone that predates computers: firms can lawfully watch a competitor’s public prices and match them, and matching in itself is not proof of an agreement.

    4. The Autonomous Machine

    This is the scenario regulators have no settled playbook for. Each firm deploys a reinforcement-learning or LLM-based agent with a profit-maximization objective, no communication channel to the rival firm exists, and yet the two systems learn, purely from repeated market feedback, to avoid price wars and hold prices above the competitive level. There is no vendor to subpoena and no message to find in discovery, because the “coordination” lives entirely inside each firm’s own model weights.

    Inside the Machine: How Reinforcement Learning Finds a Collusive Equilibrium

    The clearest empirical demonstration of the autonomous-machine scenario comes from a 2020 American Economic Review paper by Emilio Calvano, Giacomo Calzolari, Vincenzo Denicolò, and Sabina Pastorello, “Artificial Intelligence, Algorithmic Pricing, and Collusion.” The authors built simple Q-learning agents — the same class of reinforcement-learning algorithm used in early game-playing AI — and set two of them against each other in a repeated Bertrand-style pricing game with realistic demand curves. Each agent had no knowledge of the other’s strategy, no message-passing channel, and no instruction resembling “collude.” It only received a reward signal (its own profit) after each round and adjusted its pricing policy to chase higher rewards over time.

    The consistent result across a wide range of parameter settings: the agents learned to price well above the competitive (Nash equilibrium) level, and they did it using a strategy that looks a great deal like classic cartel discipline — hold a high price, and if the rival ever undercuts you, punish them with a temporary price war before returning to the high price once the rival falls back in line. That reward-and-punishment loop, known in game theory as a trigger strategy, is exactly the mechanism economists have long argued sustains human cartels. The paper’s agents rediscovered it from scratch, through nothing more than reward-seeking trial and error, which is the kind of independent multi-agent coordination researchers have also documented in cooperative settings — a related dynamic to how multiple AI agents learn to work toward a single objective together even without being told the specific steps to take.

    A separate, empirical (not simulated) line of evidence comes from the German retail gasoline market. Economists Stephanie Assad, Robert Clark, Daniel Ershov, and Lei Xu studied what happened after German filling stations adopted algorithmic repricing software between 2014 and 2017. Stations that switched to algorithmic pricing raised their margins relative to stations that stayed manual, and the effect was concentrated specifically in local markets with only two or three competing stations — exactly the market structure where tacit coordination is easiest to sustain, human or machine. Nobody needed to prove an agreement; the algorithms simply made stations better and faster at holding a shared elevated price once one station’s software learned that undercutting invited quick retaliation from the other station’s software.

    What makes the autonomous-machine pathway hard to regulate is that the algorithms involved are frequently not colluding on purpose in any sense a person would recognize. They are optimizing a single-firm objective — maximize my own long-run profit — and the collusive-looking outcome is simply the mathematically optimal policy once the game is repeated enough times against a rational-acting opponent. Removing the behavior would mean either banning firms from using learning algorithms that condition on rival prices, which is hard to define and easy to route around, or building in an artificial floor on how responsive a pricing agent is allowed to be to a competitor’s moves.

    The Legal Test Nobody Wrote With Robots in Mind

    US antitrust law under Section 1 of the Sherman Act criminalizes agreements in restraint of trade, and the operative word has always been “agreement.” Since the Supreme Court’s 1954 ruling in Theatre Enterprises v. Paramount, conscious parallelism — two firms independently reaching the same price because each is rationally watching the other — has not, by itself, been enough to prove an illegal agreement. That doctrine was written for humans who read a rival’s newspaper price list. It was never tested against a machine that reads a rival’s price feed every ninety seconds and reoptimizes.

    Two legal workarounds have emerged to keep the Sherman Act relevant. The hub-and-spoke theory, at the center of the RealPage litigation, argues that when competitors funnel private data through a common vendor that then recommends aligned prices, the vendor itself becomes the connecting thread of an unlawful agreement, even without any spoke-to-spoke contact — courts have accepted this theory in non-algorithmic contexts for decades, and the DOJ is now applying it squarely to software. The second workaround comes from European Union competition law, which is structurally more permissive for enforcers than US law. Article 101 of the Treaty on the Functioning of the European Union prohibits not just “agreements” but also “concerted practices,” a deliberately looser standard. The 2016 Court of Justice of the EU ruling in Eturas UAB v. Lithuanian Competition Council is the clearest illustration: travel agencies used a shared online booking system whose administrator sent all participants a message capping the discounts they could offer. The Court held that mere awareness of the message, combined with continued use of the system without publicly distancing themselves from it, could be enough to presume participation in a concerted practice — no reply, no vote, and no explicit acceptance required. Translate that standard to modern pricing software and a firm that keeps using a vendor’s algorithm after being told the algorithm recommends coordinated pricing takes on real exposure in the EU, even absent US-style proof of an agreement.

    Neither workaround reaches the pure autonomous-machine scenario, where two firms use entirely separate, independently trained algorithms with no shared vendor and no shared message. As of 2026, no US court has held that this scenario alone satisfies the Sherman Act’s agreement requirement, and most antitrust scholars doubt current law can reach it without new legislation. The Federal Trade Commission and DOJ have both signaled, including in a 2024 joint statement of interest filed in algorithmic-pricing litigation, that they view the use of a common intermediary to share competitively sensitive data as functionally equivalent to a traditional information-sharing conspiracy. That view stops short of the harder autonomous-machine case, and enforcers have been candid that existing statutes may need updating to reach it cleanly.

    Worked Example: Two Subscription-Pricing Bots Find the Same Number

    To see how the mechanics actually play out, walk through a simplified numeric example modeled on the Calvano-style setup, using round figures for a hypothetical two-firm SaaS market.

    Two competing project-management platforms, Vendor A and Vendor B, each charge a monthly per-seat price. Marginal cost to serve one seat — support, compute, storage — is $6 per month for both. If the two vendors priced at the fully competitive level, where neither could profitably undercut the other because each additional seat sells near cost, the market would settle around $9 per seat. If the two vendors instead behaved as a single monopolist maximizing joint profit, the optimal price would be $22 per seat, reflecting the fact that plenty of buyers are willing to pay well above the competitive floor before demand drops off sharply.

    Now assign each vendor an autonomous pricing agent that starts at a random price, observes the rival’s price and its own resulting sign-ups and revenue each week, and updates its own price to chase higher profit, with no communication channel to the other vendor’s system.

    Simulated pricing path over successive learning cycles
    Learning cycleVendor A priceVendor B priceCombined margin vs. cost
    Week 1 (random start)$11$14$13
    Week 6 (undercut phase)$8$9$5
    Week 14 (retaliation cycle)$10$8$6
    Week 30 (learning stabilizes)$16$15$18.5
    Week 45+ (steady state)$17$17$22

    By week 45, both agents have converged on $17 per seat — nowhere near the $9 competitive price, and much closer to the $22 joint-monopoly price than either vendor’s finance team ever instructed the software to target. The path matters as much as the destination: the agents pass through an undercutting phase around week 6, discover that undercutting triggers retaliation and erodes both sides’ margin by week 14, and gradually learn that holding a high, stable price is more profitable than fighting for share. That reward-punishment rhythm is the trigger-strategy signature economists look for in real pricing data.

    The bar comparison below shows where the steady-state price of $17 sits relative to the two natural benchmarks.

    Steady-state price vs. competitive and monopoly benchmarks (per seat, per month)

    Competitive price$9
    Algorithmic steady-state price (both agents)$17
    Joint-monopoly price$22

    Dashed red line marks the $0 baseline for each bar. Bar length is scaled to the $22 monopoly ceiling.

    In this illustration the algorithms captured roughly 62% of the total available gap between the competitive price and the monopoly price — $8 of the $13 spread — purely by learning from repeated play against each other. No vendor, no shared data feed, and no message existed to find in discovery. That combination is precisely what makes the autonomous-machine pathway the hardest one for regulators to touch under current law, and it is why several competition authorities, including the European Commission and the UK’s Competition and Markets Authority, have said openly that they are studying whether existing statutes need updating rather than assuming today’s rules already cover it.

    Reference Table: The Four Pathways Side by Side

    PathwayReal-world exampleLegal exposure (US)Detection route
    MessengerUS v. Topkins, 2015 (Amazon poster sellers)Per se illegal — same as any human price-fixing dealEmails, chats, code comments referencing the agreement
    Hub-and-spokeDOJ v. RealPage, filed August 2024 (rent-setting software)High — treated as a facilitated cartel through the vendorVendor contracts, shared data-feed logs, pricing-recommendation records
    Predictable agentSimple “match the lowest listed price” rules in retail and travelLow under current US doctrine (conscious parallelism)Public price data only; no internal evidence typically exists
    Autonomous machineCalvano et al. 2020 Q-learning simulations; German gasoline algorithm studyUnsettled — no US ruling has reached this fact pattern directlyStatistical price-pattern analysis; no documentary evidence exists to subpoena

    Common Mistakes Companies Make With Pricing Algorithms

    Legal and product teams tend to repeat the same handful of errors when they roll out an automated pricing agent.

    Treating “no explicit communication” as a full defense. Teams often assume that because their algorithm never talks to a competitor’s system, they are automatically safe. The RealPage case shows that funneling data through a shared vendor can create exposure even without direct contact, and the Eturas ruling shows the EU standard is looser still.

    Letting the algorithm see more than it needs. Feeding a pricing agent granular, non-public competitor data — occupancy, win rates, individual deal terms — recreates the exact fact pattern regulators have already prosecuted, even if the intent was purely to forecast demand rather than to fix a price.

    Skipping a documented rationale for pricing logic changes. When an algorithm’s pricing behavior shifts and nobody can explain why in plain business terms, that gap becomes exhibit one in litigation. Firms that can show a legitimate, cost-based or demand-based justification for every material pricing rule fare far better in review than firms whose only answer is “the model decided.”

    Assuming a vendor’s compliance claims transfer automatically. A revenue-management vendor telling clients its software is “compliant” does not immunize the client; each firm using the tool is independently exposed if the tool’s actual function is to align prices across users.

    Ignoring reinforcement-learning drift after launch. An algorithm tuned and reviewed at launch can behave very differently six months later as it continues to learn from live market data. A one-time legal review of the model’s code at launch is not the same as ongoing monitoring of what price the model is actually producing.

    Conflating dynamic pricing with collusive pricing. Charging different prices to different customers at different times, on its own, is neither illegal nor evidence of collusion — airlines and hotels have done this for forty years. The legal risk is specific to coordination with a rival’s pricing, not to responsiveness to demand.

    A Practical Audit Checklist Before You Ship a Pricing Agent

    • Map every data input the pricing agent receives and flag any field that originates from a competitor, a shared vendor, or an industry data pool rather than your own first-party sales and cost data.
    • Document, in plain language a non-technical reviewer can follow, the business justification for each pricing rule the agent enforces — cost basis, demand elasticity, inventory constraints — separate from the model’s own internal weights.
    • Confirm whether any third-party revenue-management vendor you use also serves direct competitors in your local market, and ask the vendor in writing whether client pricing recommendations are pooled, benchmarked, or shared across clients in any form.
    • Set a hard ceiling on how strongly the agent is allowed to react to a competitor’s price move, and log every instance the ceiling is triggered so legal counsel can review the pattern periodically rather than only after a complaint arrives.
    • Run a quarterly statistical review comparing your realized prices against a simple cost-plus benchmark; a persistent, growing gap with no cost-side explanation is the same red flag an economist retained in litigation would look for first.
    • Keep model version history and retraining logs for at least the period your jurisdiction’s antitrust statute of limitations covers, since “the old model did that, not this one” is not a defense regulators accept without documentation.
    • Brief the pricing and product teams, not just legal, on the hub-and-spoke and concerted-practice theories, since the people configuring the vendor relationship are usually the ones best positioned to prevent the exposure in the first place.

    Key Takeaways

    • Algorithmic collusion pricing does not require two firms to talk to each other; a shared vendor (hub-and-spoke) or a simple, predictable rule can produce coordinated-looking prices without any message ever being exchanged.
    • US law under the Sherman Act still requires proof of an agreement, and pure independent-learning collusion between separately built algorithms remains largely outside settled precedent as of 2026.
    • EU law’s “concerted practice” standard, illustrated by the 2016 Eturas ruling, reaches further than US law and can capture firms that merely stayed silent after being made aware of a shared pricing signal.
    • The DOJ’s 2024 case against RealPage is the clearest modern test of the hub-and-spoke theory applied to software, and its outcome will likely shape how aggressively regulators pursue revenue-management vendors going forward.
    • Academic simulations and real-world gasoline-market data both show that reinforcement-learning pricing agents can independently rediscover cartel-style reward-and-punishment strategies without being told to.
    • Firms deploying pricing agents reduce legal exposure most by restricting the data those agents can see, documenting the business logic behind pricing rules, and monitoring vendor relationships for shared-data risk.

    Frequently Asked Questions

    Is it illegal for two companies to use the same pricing algorithm?

    Not automatically. Using the same off-the-shelf software is not illegal on its own. The legal risk arises when that shared software pools competitors’ non-public data — occupancy, lease terms, win rates — to generate pricing recommendations that keep everyone’s margins elevated, which is the hub-and-spoke theory the DOJ is pursuing against RealPage.

    Can an AI pricing agent be prosecuted for collusion?

    The software itself cannot be prosecuted, but the company deploying it can be held liable for the outcomes it produces. Prosecutors and regulators look at whether a human agreement preceded the algorithm’s deployment, whether the algorithm connects competitors through a shared vendor, and whether the resulting prices show statistical patterns consistent with coordination.

    What is the difference between tacit collusion and an illegal price-fixing agreement?

    Tacit collusion happens when firms independently reach a stable, elevated price by watching and reacting to each other’s public prices, with no communication or shared data involved; under long-standing US doctrine this alone is not illegal. An illegal price-fixing agreement requires proof of a meeting of the minds, whether through direct communication, a shared intermediary, or a court’s finding of a concerted practice under EU-style standards.

    How did the RealPage case change algorithmic pricing law?

    The DOJ’s August 2024 civil antitrust suit against RealPage applied the hub-and-spoke conspiracy theory directly to a modern software product for the first time at this scale, arguing that landlords who fed private lease data into RealPage’s algorithm and used its rent recommendations participated in a facilitated cartel. Several landlord defendants settled rather than litigate the theory, which signaled that regulators intend to treat revenue-management software vendors as a live enforcement target rather than a legal gray area.

    Can reinforcement-learning agents collude without being programmed to?

    Yes, according to peer-reviewed research. The 2020 American Economic Review study by Calvano, Calzolari, Denicolò, and Pastorello found that simple Q-learning pricing agents, given only a profit-based reward signal and no instruction to coordinate, consistently learned reward-and-punishment strategies that held prices above the competitive level in repeated simulations.

    References

    • Calvano, E., Calzolari, G., Denicolò, V., and Pastorello, S. (2020). “Artificial Intelligence, Algorithmic Pricing, and Collusion.” American Economic Review, 110(10).
    • Assad, S., Clark, R., Ershov, D., and Xu, L. (2020). “Algorithmic Pricing and Competition: Empirical Evidence from the German Retail Gasoline Market.” Working paper on algorithmic pricing and margins in the German fuel market.
    • US Department of Justice, Antitrust Division. United States v. David Topkins, Plea Agreement (2015).
    • US Department of Justice, Antitrust Division. United States v. RealPage, Inc., Complaint (filed August 23, 2024, US District Court for the Middle District of North Carolina).
    • Court of Justice of the European Union. Eturas UAB and Others v. Lietuvos Respublikos konkurencijos taryba, Case C-74/14 (2016).
    • OECD Competition Committee (2017). “Algorithms and Collusion: Competition Policy in the Digital Age.”
    • US Supreme Court. Theatre Enterprises, Inc. v. Paramount Film Distributing Corp., 346 U.S. 537 (1954).

    Sana Qureshi
    Sana Qureshi
    Sana Qureshi is a fintech and consumer-protection writer who teaches readers how the systems behind money actually work—and how to avoid their traps. Born in Karachi and raised in Leeds, Sana studied Information Systems and later completed a certification in financial compliance. She worked inside a fast-growing payments startup and then with a regional bank’s fraud team, where she designed onboarding flows, risk flags, and plain-language disclosures that real people could understand.Sana’s writing connects the dots between product design and your wallet: how overdraft policies really behave in 2025, the difference between soft and hard pulls, which alerts matter, and why security hygiene is about habits, not paranoia. She reverse-engineers fine print, maps data flows, and gives readers “good friction” checklists—two-factor setups, credit freezes, spend alerts—that reduce risk without turning life into an audit.She also compares everyday tools—debit vs. credit for travel, buy-now-pay-later vs. old-school layaway—and shows how to choose a stack that integrates cleanly. Off the page, Sana drinks too much chai, photographs rainy city streets, and teaches a quarterly workshop on digital self-defense for students and freelancers. Her north star: confidence comes from clarity, and clarity comes from seeing how the pipes are laid.

    LEAVE A REPLY

    Please enter your comment!
    Please enter your name here

    Recent Posts

    More

      Data Poisoning in Financial Model Training: What the Evidence Shows

      0
      Answer Box Researchers demonstrated that corrupting roughly 0.01% of a web-scale training set, for about $60 in expired-domain purchases, was enough to implant a persistent...

      How Prompt Injection Hijacks Financial AI Agents: A Risk Brief

      0
      Quick Answer A prompt injection attack on a financial agent works by hiding instructions inside content the agent is told to read: an invoice, an...

      Fraud Vectors Unique to Agentic Payments: A Risk Brief

      0
      Quick Answer Agentic payment fraud is not card fraud with extra steps. It is a distinct risk class that emerges when an autonomous AI agent—not...

      Know Your Agent: The Compliance Frontier for AI Agents

      0
      A procurement bot renews a software contract at 2 a.m. A travel agent rebooks a flight after a cancellation, without waking anyone up. A...

      Agent Identity: Proving a Bot Really Acts for You

      0
      Quick Answer: A payments team proves a bot is really acting for you by checking three things at once, not one: a verifiable identity...

      More From Author

      More

        Retirement Income Sequencing Across Account Types: A Guide for Multi-Account Retirees

        Quick Answer Most retirees who hold a mix of taxable, traditional (pretax), and Roth accounts save the most in lifetime taxes by spending taxable-brokerage cash...

        Assisted Living Costs by Region: What the 2026 Data Shows

        Quick Answer A private room in assisted living now runs a national median of roughly $6,300 a month, but that single figure hides a spread...

        Data Poisoning in Financial Model Training: What the Evidence Shows

        Answer Box Researchers demonstrated that corrupting roughly 0.01% of a web-scale training set, for about $60 in expired-domain purchases, was enough to implant a persistent...

        How Prompt Injection Hijacks Financial AI Agents: A Risk Brief

        Quick Answer A prompt injection attack on a financial agent works by hiding instructions inside content the agent is told to read: an invoice, an...