Editorial
AI agents, which complete tasks on users’ behalf, are developing rapidly. So rapidly that leading AI labs have called for slowing frontier AI development over safety concerns, including the risk of autonomous hacking without human oversight. In response, several AI firms, with President Trump’s approval, have agreed to voluntary safety standards. While AI safety is a serious concern, such agreements may also harm competition by raising barriers to entry for competing firms.
Yet competition in AI has never been fiercer. Leading AI labs continue to release more capable, cheaper, and safer models, while deploying AI agents to consumers at an accelerating pace. Meta’s Muse, a consumer-facing general-purpose AI agent launched in the United States, reportedly surpassed 5 million downloads in less than a month, outpacing other AI assistants.
This rapid adoption is fueling concerns about disintermediation on major platforms such as Booking.com and Amazon. Amazon has blocked Meta’s Muse over safety concerns, but some observers see a deeper threat: AI agents could weaken Amazon’s control over the customer journey and undermine its multibillion-dollar advertising business (see the picture below). The shift towards AI agents may become one of the most significant forces reshaping competition in digital markets, even as it creates new competition concerns.
Over the coming weeks, I will explore competition and AI safety, the changing dynamics between AI agents and intermediaries, cloud and digital regulation, and AI infrastructure and competitiveness. Feel free to connect to work with me on these topics.
Enjoy!
My Work
Opinion: Accelerating AI Safety, Not Slowing AI Development
AI labs and governments should accelerate AI safety, not slow AI development.
The AI agents and Competition Training delves into the competition implications of AI agents, both upstream and downstream, to help tech firms, law firms, and economic consulting firms build the expertise needed to navigate these emerging challenges.
Conference: Digital Competition Retreat
I organised a closed-door digital competition retreat in Mougins, France, sponsored by AWS and CRA, on rethinking digital competition policy. The conference gathered 24 leading competition experts from competition authorities, law firms, economic consulting firms, tech firms, and academia to discuss the evolving goals and instruments of digital competition policy and strengthen the competition community.
The Crossroads
AI labs agree to slow AI development
AI development raises serious safety concerns. Yet, an agreement to slow it down also limits technological progress. A lawsuit in the United States alleges that such an agreement is anticompetitive and harms consumers.
The agreement addresses a collective-action problem: firms may be reluctant to slow down individually for fear of losing ground to competitors. But collective action also reduces competitive pressure to develop and improve AI. While some jurisdictions, including the European Union, may exempt such agreements under certain conditions, a fundamental question remains: is a collective slowdown necessary to ensure AI safety?
AI labs disagree. Some consider it necessary; others argue that the priority should be more AI safety research, not slower development. Just as competition and privacy dominated the antitrust debate of the past decade, the intersection of competition and AI safety may define the next.
AI agents disrupt intermediaries
AI agents, such as Meta Muse, Google Spark and OpenAI Dots, can perform many tasks traditionally handled by online intermediaries, from product discovery to purchasing. Consumers may no longer need to navigate platforms like Booking.com to connect directly with hotels.
This shift challenges intermediaries whose business models rely on website traffic to generate revenue through commissions and advertising. AI agents may bypass these platforms and contract directly with business users. Even when they visit intermediary websites, agents may respond to advertising differently from human users.
The deeper challenge is that AI agents could undermine the very rationale for some intermediaries. They can also mediate directly with business users at a much lower cost. The Verge, for instance, tested Bites, an AI food-delivery service, on ChatGPT and found that an order cost $56.73, compared with $70.29 for the same order on DoorDash, a food-delivery platform. The difference reflected DoorDash’s higher meal price (due to its commission fee) and service fee.
AI agents could therefore intensify competition over product discovery and prices, putting pressure on intermediaries’ business models.
Yet agents still need access to websites and intermediaries to perform their tasks. Platforms and large websites face a difficult choice. Granting access may secure their place in the agentic economy, but could also mean losing control over the customer journey and traffic-based revenue. Blocking access may protect their existing business models in the short term, but prove counterproductive if AI agents become the dominant way consumers interact online. Depending on the circumstances, restricting access could also raise competition concerns.
AI services form part of search engines
Are AI services distinct from Google Search? Recent regulatory and judicial developments in the United States, the United Kingdom and the European Union point towards a more integrated view of search and AI.
In the United States, Judge Mehta found that “Google Search and AI Overviews have the same use case: providing answers to users’ search queries.” On this reasoning, Google is not tying a distinct AI service to its search engine.
In the United Kingdom, the Competition and Markets Authority has included AI assistants in its approach to browser choice screens, aiming to “ensure [the criteria to be part of the choice screen] are technology neutral given the continued evolution in the way services are incorporating AI and how users are searching for information online, including through AI assistants”. This recognises that search is evolving to answer engines.
In the European Union, the Commission has designated ChatGPT as a Very Large Online Search Engine (VLOSE), recognising its growing role in providing search functionality.
Taken together, these developments support the view that AI features can be integral to search rather than distinct services. This distinction matters for the Digital Markets Act (DMA): if AI features form part of Google Search, their integration does not, by itself, establish tying or self-preferencing. It may instead reflect the evolution of search into answer engines. The Commission’s approach to defining AI services will therefore have important implications for the DMA’s application to AI.