France Flags Lock-In Risk as OpenAI, Google and Anthropic Hold 84% of the AI Agent Market

Sophie Laurent Legal & Compliance Lead, YuSMP Group · Tracking EU AI Act, DMA and competition rules for US and EU software teams

The short answer

What the Autorite actually said

The Autorite de la concurrence opened its inquiry into the AI agents sector on 8 January 2026 and published the resulting opinion, No. 26-A-05, on 17 July 2026. It is the regulator's third contribution to an ongoing review of the AI value chain and the first to move downstream from model training to the deployment layer — the point where "agentic" tools reason, plan and execute multi-step tasks with limited human intervention. The headline finding is stark: OpenAI, Google and Anthropic collectively control more than 84% of the market, with Amazon, Microsoft and Nvidia present as integrated operators and Mistral AI, Perplexity AI and xAI as sector-native challengers.

The methodology is worth noting because it is unusually hands-on for a competition authority. The document runs to more than 3,700 pages with annexes, and the regulator built its own AI agents, asked them 550 shopping-related questions, and logged exactly which websites each agent visited and cited. That is how it reached its concern about disintermediation: as agents become the interface people use to shop, book and decide, they could redirect an estimated 20–25% of consumer traffic by 2030, concentrating the choice of which merchants and services a user ever sees inside a few companies.

Why concentration at the agent layer is different

Concentration in foundation models is already familiar. What the French regulator is describing is one layer up: not who trains the models, but who owns the agent — the runtime that holds memory, calls tools, keeps context across sessions and decides which action to take. That layer is stickier than a model API. A model can be swapped behind an interface in an afternoon; an agent that stores your users' history, integrations and preferences is far harder to leave, and that is precisely the switching cost the Autorite is worried about.

The same integration that makes agents useful makes them concentrating. An agent that is pre-installed in an operating system, wired into a productivity suite and holds the user's data has three compounding advantages a challenger cannot easily match: default placement, distribution and accumulated context. The regulator's point is not that these firms have done anything unlawful, but that the structure of the market rewards vertical integration so strongly that, without interoperability and portability, the 84% could harden rather than erode.

The lock-in mechanics regulators are watching

The opinion is useful precisely because it names the specific mechanisms rather than gesturing at "big tech." Six recur, and each maps to a decision a product team can control.

• Lock-in and switching costs. User data retention and the absence of export create barriers to moving between agents. If a user's memory, history and connected tools cannot leave, the platform does not have to compete to keep them.
• Interoperability barriers. Agents act through common standards; if the governance of those standards sits with a single operator, or if integration is deliberately hard, the ecosystem fragments in the incumbent's favour.
• Default placement. Pre-installation in operating systems and productivity suites hands integrated operators disproportionate reach before anyone has chosen anything.
• Self-preferencing. Because the agent decides what it surfaces and cites, it can steer demand toward the operator's own services — the same discrimination concern that shaped a decade of search and marketplace cases.
• Disintermediation. As agents replace the browsing step, the merchants and services a user never sees effectively cease to exist for that user.
• Algorithmic collusion. If agents negotiate or transact on users' behalf, their pricing behaviour could align in ways that look like coordination without anyone agreeing to it.

The Autorite's remedies are equally concrete: standards developed through open, transparent and collaborative processes; the ability to switch between agents "without significant loss of information"; technical and contractual terms that let third parties integrate; and full enforcement of the tools Europe already has, including the Digital Markets Act and the AI Act.

What it means for US & EU software teams

For most teams the immediate takeaway is not regulatory, it is architectural. You cannot control whether OpenAI, Google and Anthropic hold 84% of the market, but you can control how tightly your product is bound to any one of them. The concentration the Autorite describes is exactly the dependency risk that bites when a provider changes pricing, deprecates a capability, or restricts an integration — and a market with three dominant players gives you little leverage when it happens.

The defensive pattern is the same one that has protected teams through every platform shift: an abstraction layer over agent and model providers, portable data formats, and a tested fallback path. Confirm you can actually export a user's conversation and memory state, not just in principle but in a script you have run. Avoid wiring business-critical workflows directly to one proprietary agent runtime when an open protocol would do. For regulated FinTech and HealthTech teams the stakes are higher, because a supervisor will ask how you would continue operating if a core provider changed its terms — and "we assumed they wouldn't" is not an answer that survives an audit.

There is a strategic read too. If European enforcement pushes toward interoperability and data portability, the teams that already built portable will move fastest when open standards arrive, while those locked into a single stack will pay a migration tax to catch up. Betting on portability is not just compliance hedging; in a concentrating market it is how you keep your own optionality.

What to do now

You do not need to react to an advisory opinion with a compliance project. You need a short, honest inventory of where your product would hurt if a dominant agent platform changed the rules.

1. Name your agent-layer dependencies. List every place your product relies on a specific agent runtime, memory store or orchestration platform, and mark which are business-critical.
2. Test your exit. Actually export a user's conversation, memory and connected-tool state and confirm you could reload it elsewhere — treat "export exists" as unproven until you have run it.
3. Put an abstraction layer in. Route agent and model calls through an internal interface so swapping a provider is a config change, not a rewrite.
4. Prefer open standards. Where an open, multi-vendor protocol covers your need, use it instead of a proprietary one that only one platform speaks.
5. Keep a fallback warm. Maintain at least one alternative agent or model path you have tested, so a change in terms is a migration you have rehearsed.
6. Document it for auditors. In regulated sectors, write down the dependency, the switching plan and the fallback — that record is what a supervisor or enterprise customer will ask to see.

None of this is legal advice, and an advisory opinion is not enforcement. But the direction is clear: European regulators now see the agent layer, not just the model, as where competition will be won or lost. The teams that treated portability as a design principle rather than a rainy-day plan will meet whatever comes next with options instead of a scramble.

Frequently asked questions

Sources

Autorite de la concurrence — AI agents: the Autorite issues its opinion on the competitive functioning of the sector (Opinion 26-A-05), 17 July 2026 (primary source)
MLex — Emergence of AI agents needs close scrutiny, French antitrust enforcers say
Concurrences — The French Competition Authority issues an opinion on the competitive functioning of the AI agents sector