Agentic AI presents amazing opportunities that will streamline and automate numerous tasks, and numerous agentic AI models and systems have already emerged. As agentic AI models and systems proliferate, however, new harms to competition might emerge, and practitioners should be aware of them. To that end, this article describes several threats to competition that could materialize in the context of agentic AI, including risks that a dominant AI agent provider or developer monopolizes the industry, that AI agents are used to facilitate collusive data collection of pricing or similar commercial data, and that AI agents could themselves enter into anticompetitive arrangements without human oversight. All of these risks are apparent in the United States, and extend throughout the world to the UK, EU, and beyond. Practitioners should also be aware of the interplay between fair competition and the interoperability of AI agents so that they can talk to each other. While interoperability is highly desirable, with interoperability comes risks that certain technical standards emerge that have IP, standard essential patent, or software licensing obligations that might have positive or negative effects on fair competition.
While the antitrust community continues to evaluate the market dynamics and regulatory challenges surrounding generative AI, the advent of agentic AI ushers in a new frontier of complexity at the intersection of technology, innovation, and competition policy.
To start with, what is "agentic AI?" Agentic AI refers to "an artificial intelligence system that can accomplish a specific goal with limited supervision. It consists of AI agents — machine learning models that mimic human decisionmaking to solve problems in real time."2 The promise of agentic AI is that it does not just use AI to make predictions or generate text or other content in response to algorithms or manual prompting, but that it accomplishes more complex tasks autonomously or semi-autonomously. Agentic AI can be used to create and purchase groceries for an Italian dinner,3 book hotel reservations,4 or plan an entire vacation with only high-level human inputs.5 To complete these tasks and others, AI agents are adept at collecting data to make decisions, as an "agent" of a human for whom it is trying to carry out the tasks. AI agents can thus be thought of as somewhat of a cross between a webcrawler, a data aggregator, and a search engine that can carry out particularly defined tasks. Enterprises are increasingly utilizing embedded agentic AI both to perform customer-facing tasks such as handling customer inquiries, resolving issues with online orders, and providing personalized sales suggestions and support, but also more complex tasks around demand forecasting and inventory or supply chain optimization, dynamic pricing, and cybersecurity threat detection and response. The technology unlocks exciting opportunities. At the same time, agentic AI creates potential risks for businesses in navigating competition rules.
01
RISK OF A DOMINANT AI AGENT PROVIDER
Currently, there are numerous players in the market that provide AI agents. The agentic AI market remains in its nascent stages, and numerous vendors and platforms have emerged, with no single firm currently holding a dominant share. Nevertheless, observers are closely monitoring the development of this market to determine whether it will replicate certain anticompetitive trends seen in the traditional search market.
The Department of Justice's seminal case against Google Search illustrates how an emerging digital market can evolve into a highly concentrated one. The DOJ secured a landmark victory against Google when Judge Mehta ruled that the company had unlawfully maintained its monopoly over the search engine market. The court found that Google engaged in anticompetitive conduct by spending billions annually on agreements with device manufacturers to ensure that Google remained the default search engine on new mobile devices.6
The remedies phase of the trial concluded on May 30, 2025, during which the DOJ expanded its scope to include AI. The DOJ argued that "[w]e are at an inflection point . . . it is [] critical to extend remedies to GenAI to allow this new technology to rise or fall outside the shadow of Google's search monopoly." The DOJ also made specific mentions of the future of agentic AI. Citing Google's internal documents the DOJ warned, "Google plans to make Gemini the 'primary agent' in Chrome, prioritizing its integration over Gemini's rivals and building on Google's history of making switching the default difficult in Android."7 However, Judge Mehta seemed skeptical of the argument that the search engine case's remedies might extend to Google's generative AI or AI agents, stating that "[i]t seems to me you now want to kind of bring this other technology into the definition of general search engine markets that I am not sure quite fits." Judge Mehta's remedies decision is expected in August 2025.
In the context of agentic AI, there are not yet any known examples where a company has anticompetitively "bundled" an AI agent with other products or required a particular agent as a "default." Nonetheless, one could imagine scenarios where "AI agents" are subject to anticompetitive agreements that solely direct the human users to particular companies' products or services, perhaps excluding competitors from consideration, and do not seek to objectively determine what might be best for the person who is trying to find or utilize a tool or resource within a particular industry. If such agreements are used pervasively enough, one particular AI agent could become dominant, reducing competition that can spur innovation, control pricing, or create optionality
02
RISK OF COLLUSIVE DATA COLLECTION AND COORDINATION
Aside from monopolization concerns, the agentic AI market may also be susceptible to collusion under a hub-andspoke theory. In a hub and spoke conspiracy, there is a central conspirator (the hub) that reaches separate agreements with competitors (the spokes), which facilitate an explicit or implied agreement among the competitors (the rim connecting the spokes) to agree to some anticompetitive terms through the hub, such as fixing prices, limiting supply, allocating markets, or exchanging competitively sensitive information. Traditionally, a "rimless" wheel – one without an agreement among the competitor-spokes – does not violate antitrust laws because there is no agreement among the competitors. However, as the agentic AI market evolves, complex legal questions likely will emerge concerning the interpretation of "intent" and "agreement" in contexts where AI agents operate autonomously.
For example, with the use of "AI agents," competitors using the same AI agent might end up exchanging competitive pricing information and using that information to set prices, without even knowing it. For example, if five widget manufacturers all ask the same AI agent to "consider my costs and capacity, survey the market for widget prices, demands, and other information and recommend the most profitable price for my widget," and the AI agent uses information provided by each widget manufacturer to recommend prices, the widget competitors may rely on their AI agents' price recommendations without knowing that the AI agent is using the competitors' information to make that recommendation, or even that the agent is instructing each competitor to charge the same price.
This is perhaps what motivated the DOJ and several attorneys general when they banded together to sue RealPage. RealPage's product allegedly draws nonpublic pricing information from competing landlords and uses that information in connection with AI-based pricing algorithms to recommend rental prices to those same landlords.8 The DOJ asserts that, "[b]y feeding data into a sophisticated algorithm powered by artificial intelligence, RealPage has found a modern way to violate a century-old law through systemic coordination of rental housing prices – undermining competition and fairness for consumers in the process. Training a machine to break the law is still breaking the law."9
While the RealPage litigation was only filed in 2024 and motions to dismiss are pending, one could imagine other scenarios involving agentic AI that trigger similar concerns. For example, one could imagine a service that uses AI agents' work on behalf of industry competitors to login and/or collect data from various websites to gather pricing or other information that is behind a paywall or login portal so as to aggregate that information across retail sectors in an anticompetitive way. Increased use of agentic AI will only incentivize and encourage the collection of more data — both for purposes of training the AI models themselves, and for purposes of carrying out the relevant agentic tasks. Indeed, some have said that agentic AI will transform the very way that databases are designed and organized, and that "[w] ith the rise of AI agents, the database has become an even more critical application layer"10 — the very structure of those databases may become subject to antitrust scrutiny in the future.
03
AGENTIC AI THAT ITSELF FACILITATES ANTICOMPETITIVE AGREEMENTS OR ARRANGEMENTS
In the above scenarios, AI agents might be used to anticompetitively gather information at a hub, or be subject to certain human-based agreements that anticompetitively constrain the AI agent's outputs or actions. But what if agentic AI is used to facilitate future contracting or agreements in an anticompetitive manner? At a recent technology event, AI agents were on both sides of a transaction — an AI agent called into a hotel to book a reservation, and an AI agent from the hotel took the call and facilitated the booking of a reservation.11 Therefore, scenarios involving direct communications between AI agents to bind human activity are not just hypothetical — they are real. If agentic AI is left to its own devices, one can imagine future, more complex scenarios where AI agents are used to negotiate and bind businesses into contracts. Perhaps those contracts or a series of contracts AI agents negotiate include anticompetitive exclusive dealing or tying arrangements, and could themselves be the basis for antitrust or competition concerns. This underscores that, quite apart from considering how AI agents collect data and use it to perform tasks, organizations that deploy agentic AI will need to consider how to ensure that AI agents do not run amok and commit potential antitrust violations through the transactions or agreements that they undertake or facilitate. Such issues might need to be addressed using both technical controls and legal measures. These sorts of concerns are why many agentic AI systems being deployed today require a human in the loop to ensure that there are checkpoints before a human adopts any decision or transaction that the AI agent recommends.12
04
THE ABOVE RISKS EXTEND BEYOND THE UNITED STATES
There are also lessons from the UK and EU when assessing the risks from an antitrust perspective of advancements in agentic AI functionality, where the underlying models rely on competitively sensitive data inputted from competing businesses, and the outputs contribute to decisions or recommendations on competitive parameters to be adopted by those businesses, including on pricing or other factors. Such scenarios are increasingly common where industry-wide data is already collated to some degree and many software providers are now looking to augment existing datasets with additional AI-powered tools which help companies make strategic commercial decisions, including in relation to dynamic and discriminatory pricing.
The United Kingdom has, like the United States, scrutinized so-called "hub-and-spoke" arrangements, or "A to B to C" infringements, whereby businesses will provide commercially sensitive information (for example in relation to costs and future pricing intentions) to a third party (whether in the context of a distribution system or for other purposes such as data aggregation) and that third party passes on that information to other competing businesses. In cases relating to the supply of replica football kits, children's toys, and dairy products (Cases CA98/06/2003, CA98/8/2003, and CA/03/2011) the UK Competition & Markets Authority (and its predecessor organizations) found that information sharing had facilitated a price-fixing cartel implicating both the first business, the intermediary, and the recipient of that information as guilty of anticompetitive conduct. However, a defense in such scenarios is where there was no knowledge by Company A that Company B would pass on its information, in such a scenario, only Company B and C are liable. Thus, agentic models which act as modern-day hub-and-spoke arrangements (passing information from one competitor to another) could lead to antitrust scrutiny.
In a related but separate vein, in the European Union, there has long been an acknowledgement that service providers and parties that enable cartel activity (but do not directly participate in it) can also be liable under Article 101(1) of the Treaty on the Functioning of the European Union. In Case C-194/14 P AC Treuhand v. Commission, the European Courts examined and ultimately upheld a Commission fine in relation to a cartel involving heat stabilizers, which had found a third party not directly involved in the cartel to be nonetheless liable with other members of the cartel on account of its role as a "cartel facilitator." Similarly, in Case C-39/18P, Commission v. Icap Management Services, the Commission fined ICAP $7.9 million for taking part in several yen interest rate derivatives cartels, despite not itself being active on the affected markets. These cases serve as a warning to service providers that have a supporting role in enabling cartel conduct that they may face hefty fines and should take steps to mitigate the risk of – perhaps inadvertently – being party to anticompetitive information sharing and collusive conduct. Indeed the subsequent case law involving AC Treuhand confirmed the possibility that enforcers could presume that third parties involved in data aggregation could have knowledge of the illegal use of their data (a scenario of close if not equivalent application to providers of Agentic AI models).
To view the full article clickhere
Footnotes
2 Cole Stryker, What is agentic AI?, IBM Think, https://www.ibm.com/think/topics/agentic-ai (last visited June 5, 2025).
3 Steven Melendez, Instacart users can now plan meals using AI, Fast Company, Tech (May 31, 2023), https://www.fastcompany. com/90903040/instacart-users-can-now-plan-meals-using-ai.
4 Lance Ulanoff, Two AI chatbots speaking to each other in their own special language is the last thing we need, MSN TechRadar (Feb. 25, 2025), https://www.techradar.com/computing/artificial-intelligence/two-ai-chatbots-speaking-to-each-other-in-their-own-special-language-is-the-last-thing-we-need.
5 Trevor Laurence Jockims, From Google to Expedia, AI travel agents planning future trip far beyond 'assistant' status, Technology Executive Council (May 16, 2025), https://www.cnbc.com/2025/05/16/ai-travel-agents-planning-future-trip-far-beyond-assistant-status.html.
6 United States v. Google LLC, No. 1:20-cv-03010 (D.D.C. 2020). This liability finding is subject to appeal.
7 Pls.' Remedies Proposed Findings of Fact, 133, United States v. Google LLC, No. 1:20-cv-03010 (D.D.C. 2020), ECF No. 1370.
8 Press Release, Dept. of Justice, Justice Department Sues RealPage for Algorithmic Pricing Scheme that Harms Millions of American Renters (Aug. 23, 2024), https://www.justice.gov/archives/opa/pr/justice-department-sues-realpage-algorithmic-pricing-scheme-harms-millions-american-renters.
9 Ibid.
10 Andrew Davidson, The Rise Of AI Agents And The Future Of Data, Forbes Technology Council (Mar. 6, 2025), https://www.forbes.com/ councils/forbestechcouncil/2025/03/06/the-rise-of-ai-agents-and-the-future-of-data/.
11 Lance Ulanoff, Two AI chatbots speaking to each other in their own special language is the last thing we need, MSN TechRadar (Feb. 25, 2025), https://www.techradar.com/computing/artificial-intelligence/two-ai-chatbots-speaking-to-each-other-in-their-own-special-language-is-the-last-thing-we-need.
12 Cobus Greyling, AI Agents With Human In The Loop, Medium (Aug 15, 2024), https://cobusgreyling.medium.com/ai-agents-with-human-in-the-loop-f910d0c0384b.
Originally published by Competition Policy International
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
