Opinion 26-A-05 of July 17, 2026

Initially designed as simple conversational agents, generative AI tools have gradually evolved into AI agents. These agents no longer simply perform isolated tasks, but are now capable of reasoning, planning and orchestrating multiple tasks by coordinating with other AI agents with a degree of autonomy. The terms “agentic AI” and “AI agents” have been used to describe these new capabilities. According to the European Commission, AI agents are “software applications designed to perceive and interact with the virtual environment. [They] operate autonomously, meaning they are not directly controlled by a human”.

AI agents operate downstream in the generative AI value chain, during the deployment and marketing of generative AI services (the upstream segment of the value chain was analysed by the Autorité in Opinion 24-A-05). Their use cases are numerous: text generation (e.g. emails, social media posts and user interactions), image creation or editing, and software code generation are just a few examples, already widely deployed across the economy.

Agentic commerce is one of the most recent developments in agentic AI. It refers to the application of agentic technologies to the e-commerce sector, with the potential to fully or partially automate the purchasing journey.

While such services were not yet available on the French market as of the date of publication of this opinion, users can already receive product recommendations through AI agents, a development that some stakeholders refer to as “conversational commerce”. To demonstrate these new functionalities in practice, the Autorité conducted an experiment involving two AI agents, examining how conversational commerce works within each agent and, in particular, how sources are referenced. The methodology and results of this experiment are presented in Annex 2 to this opinion.

To make an AI agent available to users, three main resources are required: (i) access to a generative model, which constitutes the technical foundation of the AI agent; (ii) access to high‑quality data, necessary to improve its performance; and (iii) inference capabilities, which enable the effective use of the AI agent.

The main operators in the AI agents sector are OpenAI, Google and Anthropic, which together accounted for more than 84% of the global market as of May 2026. Other operators include vertically integrated companies already active in the sector, such as Amazon, Microsoft and Nvidia, as well as newer AI-native companies, such as Mistral AI, Perplexity AI and xAI. Operators in the agentic commerce sector include both these companies and e‑commerce companies that may develop specialised AI agents.

Some AI agents are no longer limited to simple question-and-answer interfaces. They are increasingly positioning themselves as entry points for users to browse the internet and are now evolving into fully-fledged platforms, enabling users to access information and third‑party services directly without leaving the agent’s environment.

The main AI agents are offered to individual users under a freemium model, with basic functionalities available free of charge and several tiers of paid subscriptions providing access to more powerful models, higher usage limits and additional functionalities. Specific subscription plans are available for business customers, providing access to shared workspaces and more advanced administrative capabilities. Access to AI agents via an API is charged based on usage, with prices generally depending on the number of tokens used.

Several pieces of EU legislation may apply to the sector, including the AI Act, the Digital Markets Act (DMA), the Digital Services Act (DSA), the General Data Protection Regulation (GDPR) and the Data Act.

Barriers to entry and expansion

Compared with the training of generative models, barriers to entry are lower downstream because AI agent developers can use third-party models via APIs or open-weight models. However, AI agent developers relying on such third-party solutions may face specific barriers, particularly when adapting these models for specific use cases. More importantly, this strategy may create dependencies on third-party model providers, depending on how the third-party generative model is accessed, with access to models via APIs resulting in the greatest degree of dependence.

Scaling remains heavily constrained by the need to access users, which determines AI agent developers’ ability to deploy and sustain their solutions. Beyond the vertical integration of certain operators, several large, established digital companies benefit from advantages arising from the direct integration of their AI agents into their ecosystems. Examples include the integration of Copilot into Microsoft’s Office 365 suite, Gemini into Google’s Android operating system, and Meta AI into WhatsApp. While players such as OpenAI and Anthropic have managed to establish a significant position in the sector without controlling the entire value chain, the ability of their competitors to integrate the user experience into the ecosystems of large, established digital companies makes market entry and expansion both economically and technologically costly.

Access to data also plays a key role. Established AI agent developers can improve their generative models using proprietary data and enhance the quality of their responses using usage data, complemented by partnerships with third parties such as press publishers. These major operators have an additional advantage when they have access to technologies underpinning a search engine, which can rank and identify the documents most relevant to a user’s search. This barrier relating to data access also varies depending on the use case: it is particularly significant for general-purpose and news agents, but may potentially be mitigated in professional use cases by access to internal customer data.

Although both individual and business users often use several AI agents in parallel, which may limit lock-in effects, this practice is subject to significant technical limitations in terms of portability and interoperability, despite the tools developed by some operators to overcome these difficulties.

The downstream segment of the value chain may also involve significant costs. The total cost of inference can substantially exceed the cost of training, as inference costs are incurred repeatedly and increase in proportion to the use of the AI agent. Accordingly, the more an AI agent is used, the higher these costs become, particularly due to its energy consumption. Moreover, agentic tasks, which require multi-step processing, significantly increase the total cost of a user query. Several operators also indicated that high regulatory compliance costs may favour incumbent players.

Lastly, the development of agentic commerce faces specific technological barriers, including a lack of standardised data, established standards and shared infrastructure, as well as behavioural barriers.

Competition risks associated with AI agents as new intermediaries in the digital economy

In a context where the underlying generative model is sometimes no longer a key differentiating factor between operators, competition may shift to other criteria, particularly the quality of the AI agent’s responses. The quality of these responses depends not only on the performance of the underlying model but also on the agent’s associated functionalities.

The quality of the AI agent’s responses is therefore an important competitive parameter. When a user submits a query, they are seeking a relevant response, i.e. one that addresses their request appropriately and is context-sensitive. AI agent developers are therefore adopting various strategies to overcome the aforementioned barriers and improve the quality of the services they provide. For example, they may choose to specialise in specific segments, such as the legal sector, offer enhanced data sovereignty guarantees, or enter into partnerships to quickly reach the level of quality achieved by established operators.

By offering users access to an ever-expanding range of services and information through a single interface, AI agents are emerging as increasingly important gateways to digital services. The Autorité notes that the “platformisation” of AI agents raises competition risks that could undermine the proper functioning of the sector and the diversity of services offered.

These risks stem primarily from the disintermediation of parts of the digital economy, i.e. the replacement or disappearance of economic operators within a value chain. A large proportion of user queries are likely to be satisfied by a single response from an AI agent, which could undermine the business models of many websites and digital services whose revenues depend on the display of advertising or subscription fees, thereby reducing the variety or quality of online content available. While traffic currently directed to e-commerce websites from AI agents remains minimal (less than 5% in France), it could become a significant channel for reaching consumers in the medium term (almost 25% by 2030).

Risks of discrimination or self-preferencing could arise in relation to access to digital services and the terms imposed on e-commerce websites to make their services available to users through an AI agent’s interface. Competition could also be distorted by the influence of partnerships or advertising on the responses provided to users.

Control over the conditions governing the visibility of content in AI agent responses may also raise competition concerns. These concerns relate in particular to the clarity, objectivity and non‑discriminatory nature of the visibility criteria applied by the AI agent, as well as the limited number of sources and offers presented to users in the response displayed to them. Control over visibility within their service could therefore give AI agent developers greater ability to steer consumer demand.

These risks may be heightened when AI agents are integrated into the existing services of vertically integrated companies. While such strategies may benefit users, certain practices could undermine merit-based competition in the sector. Operators could, for example, restrict access to their key services through self-preferencing, bundling or tying, or technical or contractual lock-in. The marketing of AI agents could also potentially give rise to price‑based exclusionary practices.

Risks associated with the automation of actions by AI agents

The rapid pace of innovation in the AI sector has fuelled the development of agentic capabilities. This ability to perform actions autonomously relies on common standards that enable interoperability between digital services and AI services. The adoption of common standards may raise competition concerns, including the risk of centralised governance of those standards by a single operator and the potential emergence of barriers to interoperability, which could in turn create a risk of fragmentation of the internet.

Moreover, under certain conditions, agentic AI could cause significant disruption across the digital economy, which could materialise very rapidly given the capabilities of established companies, including their wide range of services and large user bases. To provide their services, AI agents collect and process ever-increasing volumes of data. This large-scale data processing, combined with the retention of user histories, enables increasingly personalised services but also makes switching from one AI agent to another costly and difficult. Such practices could therefore give rise to lock-in effects.

E-commerce operators consulted by the Autorité indicated that such risks could result in the sector becoming concentrated around a small number of operators, purchasing processes becoming less transparent, and end consumers being deprived of the ability to make informed choices. Risks of disintermediation and algorithmic collusion also exist. Initiatives to develop agentic commerce are multiplying across the sector and, once agentic commerce becomes fully available to users, the changes it brings about are likely to be particularly far‑reaching.

Recommendations

The Autorité makes three sets of recommendations.

First, the Autorité recommends making full and swift use of the existing regulatory framework, including to promote user awareness and understanding of AI agents.

Recommendation 1: The existing regulatory framework should be implemented fully and effectively. While particular innovations may warrant the adoption of specific measures, the Autorité recommends giving priority, in the first instance, to the use of soft law instruments, such as guidelines or stakeholder consultation mechanisms modelled after the specification procedures provided for in the Digital Markets Act (DMA), in order to enable rapid and targeted action. Any regulatory developments should, in any event, be pursued at European Union level.

Recommendation 2: The Autorité calls for particular vigilance with regard to competition law, specifically concerning:

• equity investments made by major digital companies in competing AI agent developers and partnership agreements between these operators;
• whether users are able, in practice, to choose and use AI agents that compete with those integrated by default into the ecosystems of major digital companies;
• the definition and implementation of the key parameters that may influence how AI agents select, rank or recommend the services proposed by AI agents.

Recommendation 3: The Autorité identifies three priority areas for action to support fair access to AI agent distribution channels for all operators across the sector:

• competition authorities should pay particular attention to the role of AI agent distribution channels, particularly MaaS platforms;
• the European Commission’s work under the Digital Markets Act (DMA) Review and its ongoing market investigations into the cloud sector should continue, building on the Commission’s preliminary position and considering whether marketplaces that distribute AI models should be designated as core platform services;
• if additional companies are designated as such platforms, the Commission should ensure, as part of the ongoing market investigation into whether the obligations currently set out in the DMA are effective in addressing practices that restrict contestability or are unfair in the cloud sector, that AI agent and generative model developers have guaranteed access to such marketplaces and benefit from fair, reasonable and non‑discriminatory visibility.

Recommendation 4: A competitive sector depends, in particular, on users being able to compare available offers and make informed choices. This requires:

• greater efforts by public authorities and professionals to improve user awareness and understanding, particularly among younger users;
• proactive engagement by regulators, in particular the Directorate General for Consumer Affairs, Competition and Fraud Prevention (DGCCRF), to ensure the effective enforcement or, where necessary, adaptation of existing regulatory frameworks, particularly those relating to restrictive competition practices, consumer protection law and the Digital Services Act (DSA);
• greater attention from the relevant public authorities, such as the French Center of Expertise for Digital Platform Regulation (Pôle d’expertise de la régulation numérique – PEReN), to the technical analysis of AI agents, including through pilot projects, in order to assess the impact of this technological innovation on public policy;
• closer coordination between all the relevant public authorities.

Second, the Autorité invites operators in the sector to promote interoperability and portability in the sector.

Recommendation 5: The Autorité urges operators in the sector to establish technical and contractual terms of use that promote interoperability between the services of vertically integrated companies and those of third-party AI agent developers. This could include, in particular, providing accessible, comprehensive and up-to-date documentation, together with software specifications enabling the development and maintenance of such third‑party integrations.

Users should also be able to switch from one AI agent to another without significant loss of information or functionality. This requires, in particular, ensuring effective data portability between AI agents.

Third, the Autorité stresses the importance of ensuring the establishment and maintenance of open technical standards.

Recommendation 6: The standards governing agentic commerce should be developed and maintained through transparent, open and collaborative processes, in order to prevent any situation in which a dominant operator exercises excessive control or influence. These standards should also ensure the highest possible level of openness consistent with security and reliability requirements, including through the adoption of open-source solutions and fully interoperable standards.