AI agents: France’s competition authority flags an 84% lock-in risk — what it means for your SMB

On Friday, July 17, 2026, the French competition authority (Autorité de la concurrence) published opinion no. 26-A-05 on the competitive functioning of the AI agents sector — a self-initiated inquiry opened on January 8, 2026, the third the institution has devoted to AI after its earlier work on cloud computing and generative AI, and a study on the sector’s energy impact published in December 2025. To ground its findings in practice rather than theory, the Authority built its own AI agents and ran them through 550 shopping-related questions, logging exactly which websites they visited and cited in their answers. The headline figure: OpenAI, Google and Anthropic together control more than 84% of the AI agents sector.

The risk the Authority calls “platformization”

The opinion’s central concern is not the concentration figure itself but what it enables. The Authority describes a “platformization” process — AI agents becoming the layer users go through to search, compare and eventually buy — that raises three specific competitive risks: disintermediation of the businesses and services an agent used to simply point to, discrimination in how an agent ranks or excludes options, and opaque visibility conditions that determine which products or providers even get mentioned. Today, “agentic commerce” — an AI agent handling all or part of a purchase, from product search to payment — remains marginal in France: traffic redirected from AI agents to e-commerce sites is still under 5%. But the Authority’s own projection puts that share at 20% to 25% by 2030, which is exactly the kind of trajectory a business only notices once it has already lost the channel.

A real opening for specialized vendors — in the regulator’s own words

The opinion does not stop at the warning. It also notes that scaling up is heavily constrained for AI agent publishers in general — by the need to reach users, and by barriers including data access, technical migration, interoperability and inference costs. But it specifically flags that the data-quality barrier is far less significant for specialized AI agents than for generalist ones. In plain terms: a narrow, sector-specific agent built around a real business process does not need the scale of a ChatGPT or a Gemini to compete on quality — it needs the right data for its one job. The Authority’s main recommendation follows from that: it calls on the sector’s vertically integrated players to open up interoperability — accessible, complete and up-to-date documentation, technical specifications enabling third-party integrations, open standards, and data portability — rather than trapping users inside a single default agent.

What it means for your SMB

This concerns an SMB in two distinct ways. As a buyer of AI agent tools, the lock-in risk the Authority describes at the market level plays out identically at contract level: an agent wired into your CRM, your inbox or your invoicing that cannot export its data, its logs or its configuration if you switch providers leaves you exactly as dependent as the Authority describes for the sector as a whole. As a business trying to be found, the same opinion is a signal, not yet an alarm: agentic commerce sits under 5% of traffic today, but the products and service pages that get selected, compared and cited by an AI agent tomorrow are being built — or not — right now, the same discipline that already matters for getting cited by ChatGPT or Perplexity — a watch-and-visibility effort a regulatory watch agent can run on your behalf.

Before signing with an AI agent vendor: four checks

• →Before committing to any AI agent — generalist or specialized — ask for the documentation on data export and integration with other systems; a provider unable to answer clearly is the lock-in risk described above, in miniature.
• →Do not assume a big-name generalist agent is automatically the safer choice: the Authority’s own opinion notes that a narrow, well-built specialized agent competes on the quality of its data, not on scale.
• →If you sell online, test today how AI agents (ChatGPT, Gemini, Copilot) describe and recommend your products or services — the current under-5% figure is a floor, not a ceiling, on how fast this channel can grow.
• →Track this kind of institutional opinion with a regulatory watch agent rather than occasional trade-press reading — it is non-binding today, but it is exactly the sort of signal that shapes tomorrow’s obligations and vendor contracts.

None of this calls for giving up on AI agents — quite the opposite: the Authority’s own opinion suggests that a well-built, narrowly scoped agent has more room to compete than the 84% concentration figure might suggest. It is a reminder to read the contract, not just the demo, before an agent becomes the layer between your business and your customers.