Summary

  • France's Autorité de la concurrence says OpenAI, Google and Anthropic accounted for more than 84% of the global AI-agent sector in May 2026, based on Sensor Tower data cited in Opinion 26-A-05.
  • The authority identifies distribution, rankings, data portability and unequal technical integration as potential sources of market power, and recommends using existing EU rules before reaching for new legislation.
  • The opinion is advisory: it imposes no remedy, makes no infringement finding and expressly does not prejudge any later antitrust assessment or enforcement action.

France's competition authority has moved the AI-agent debate downstream. Opinion 26-A-05, published on July 17 after about 40 responses to a public consultation, is less concerned with who trains the largest model than with who controls the routes through which an agent reaches users, services and transactions.

That shift matters because distribution can harden a technical lead into an economic one. The opinion says OpenAI, Google and Anthropic together represented more than 84% of the global sector in May 2026, citing Sensor Tower. Some competitors must buy model access or inference capacity while established digital groups can place their own agents inside operating systems, office suites, cloud marketplaces and other services with large existing user bases.

The document is a market study and set of recommendations, not a case decision. It does not identify unlawful conduct, order a company to change its practices or establish liability. The authority says the risks it describes will be monitored, but expressly reserves any future assessment of the facts under competition law.

Traffic is small, but allocation power is already visible

The authority's central concern is that an agent can become the interface through which a user reaches the rest of the digital economy. A conventional search page offers a list of links. An AI agent may instead present one answer and only a handful of products or services. The operator's choices about sources, ranking and presentation can therefore allocate demand even before it executes a transaction.

The opinion cites estimates that traffic sent directly from AI agents to French merchant sites remains below 5%, but could approach 25% by 2030. Those figures are a current estimate and a projection, not observed proof that one quarter of online commerce will move through agents. They nevertheless explain why ranking criteria and commercial partnerships are becoming competition issues before fully autonomous purchasing is widely available in France.

The authority also tested 550 shopping-related questions across electronics, appliances, transport and accommodation using two general-purpose agents with internet-search tools. The providers used markedly different mixes of sources, and source diversity varied by sector. The authority cautions that its API-based experiment was not statistically significant and captured conversational commerce in May 2026, rather than a mature market for automated transactions. Its narrower point is still useful: the choice of agent can change which intermediaries, merchants and information sources are visible.

That makes partnerships ambiguous. A data, distribution or commerce agreement can help a smaller provider improve its service or reach customers. The same agreement can also steer demand towards a preferred partner, especially when users cannot see how recommendations were selected. Merchants and publishers then bear the cost of adapting content, negotiating access or losing traffic; smaller firms may be least able to pay those costs.

Native integration shifts the cost to challengers and users

Opinion 26-A-05 distinguishes formal access from equivalent access. A third-party agent may be able to connect to an incumbent's services through documented interfaces, yet still face more complex permissions, data formats and implementation work than the agent built into the same ecosystem. Native integration can therefore lower friction for users while raising the cost of matching that experience for a rival.

Switching costs fall on customers too. For companies, changing providers can require fresh integrations, staff training and a temporary productivity loss. Individuals may lose conversation history, stored projects and accumulated preferences. Using several agents at once limits lock-in today, but the authority argues that weak portability could matter more as one interface absorbs a wider range of tasks.

The recommendations follow that diagnosis. The authority urges scrutiny of investments and partnerships between large digital groups and rival AI providers, the ability to choose an alternative to a default agent, and the parameters that determine selection, ranking and recommendations. It also asks competition authorities to watch model-as-a-service marketplaces and says the European Commission should consider whether marketplaces distributing AI models warrant treatment as core platform services under the Digital Markets Act.

For interoperability, the authority calls on companies to adopt technical and contractual conditions that let third-party agents connect to vertically integrated services, supported by accessible documentation and software specifications. It also advocates effective portability between agents and transparent, collaborative governance of open standards. These are recommendations, not immediately binding duties.

The practical test will be whether access becomes comparable in depth and ease, rather than merely available on paper. Other watchpoints are whether users can move histories and project context without a material loss, whether ranking and sponsorship become intelligible, and whether standards remain open as agents begin to initiate payments and manage after-sales tasks. If those frictions persist, the cost of compliance and integration will favour incumbents long before a regulator needs to prove an exclusionary act.

Sources