On July 17, France’s competition regulator put a hard number on a market most people cannot yet see. In a 3,700-page opinion, the Autorité de la concurrence found that three companies — OpenAI, Google, and Anthropic — already hold more than 84 percent of the market for AI agents. The figure is striking on its own. What turns it into a warning is the regulator’s account of how that lead compounds: agents are becoming the layer through which people search, shop, and act online, and the routes a challenger would use to reach those people are narrowing while adoption is still early.
The document is an “avis” — an advisory opinion, not a fine or a binding order. But it is the most detailed public map yet of how the agent market is actually forming, and it arrives just as regulators on three continents begin circling the same question. For anyone choosing an agent to build on, sell through, or lean on every day, Opinion 26-A-05 reads less like a policy update than an early read on where lock-in is going to hurt.
This is the Autorité’s third pass at the AI stack, following earlier reviews of cloud computing and generative AI. It is the first to move downstream, past model training and into the deployment layer — where “agentic” tools reason, plan, and carry out multi-step tasks with limited human oversight. The inquiry opened on January 8, 2026, drew on roughly 40 stakeholders, and put the 84 percent headline figure on Sensor Tower usage data from May 2026.
What sets this opinion apart from the usual regulatory white paper is that the authority did not only survey the market. It built its own AI agents and ran them through 550 shopping-related questions across the ChatGPT and Gemini APIs between May 20 and 30, logging exactly which websites each agent visited and which ones it cited back to the user. That experiment is what lets the regulator talk about agent behavior as evidence rather than assertion, and it is the same method that maps cleanly onto how the broader agent-platform market is taking shape.
The opinion’s core argument is not that three firms are large. It is that their position is structurally hard to challenge, for three reinforcing reasons.
The first is default placement. Agents are riding into billions of devices as the pre-installed option — Copilot inside Microsoft 365, Gemini across Android, Meta AI inside WhatsApp. A user rarely chooses an agent the way they once chose a browser; it is simply there, and the incumbent controls the surface. Google’s tightening grip on how its Gemini models are accessed is a small preview of how much leverage that default position confers.
The second is data. Incumbents pair proprietary usage data with publisher partnerships, and the search operators among them can rank and retrieve documents no rival can see in full. Every query an agent answers sharpens the model that answered it, and that feedback loop is not available to a newcomer starting from zero.
The third is raw economics. The authority estimates that running an AI agent at scale now costs on the order of $1.20 per orchestrated task in 2026, against roughly $0.04 for a single model response back in 2023. Agents chain many calls together, so the per-task cost climbs fast — and only players who can subsidize that burn, or bundle it into an existing product, can afford to give it away while the market forms.
Nowhere is the lock-in risk clearer than in shopping. Two incompatible standards have already emerged for how a store lets an agent transact on a customer’s behalf: Google’s UCP, launched in January 2026, which asks merchants to publish a capability file, and the OpenAI-and-Stripe ACP, launched in September 2025, which requires a merchant application and dedicated APIs. They do not interoperate. A merchant that wants to be reachable by both has to build twice or pick a side.
Why fight over shopping at all? Because that is where the money and the traffic are heading. The authority found that AI agents currently redirect under 5 percent of e-commerce traffic in France, but it projects that share could approach 25 percent by 2030. Whoever owns the agent that sits between a shopper and a store at that point owns an enormous slice of retail demand — and the monetization models are still being decided. Individual agent subscriptions already run as high as $100 a month; OpenAI began advertising trials in the United States in January 2026, Perplexity abandoned its own ad tests early in the year, and Anthropic committed in February to staying ad-free so that, in its words, responses stay “guided by the principle of helping users.” Those are three different bets on how agents pay for themselves, and each one bends the incentive to favor the platform’s own answers.
Adoption is still tiny — the authority found just 26 of more than three million websites it scanned had implemented UCP four months after launch — which is precisely why the moment matters. Standards harden early, and the firms writing them are the same ones that dominate the agents doing the buying. The Autorité’s remedy is blunt: users should be able to move between agents “without any significant loss of information or functionality,” and the plumbing of agentic payments should stay open rather than fold into one company’s rails. That is the same tension now playing out in how agent-driven payments get settled on the merchant side.
The experiment holds the finding every reader should sit with. When the authority’s test agents went looking for products, ChatGPT consulted Reddit in 87.4 percent of its discovery queries — but cited Reddit in only 1.0 percent of the answers it returned. Gemini leaned heavily on the price-comparison site Idealo (87.0 percent) and the retailer Cdiscount (62.5 percent) to complete purchases, while its single most-visited source was YouTube. Across both agents, the top ten sites drew about 24 percent of all visits but only 20 to 22 percent of the citations.
Strip out the percentages and the point is simple: the sources an agent actually relies on are largely invisible to the person reading its answer, and the handful of links it does show are a thin, unrepresentative slice of what shaped the recommendation. For a casual query that is a curiosity. For a purchase, a health claim, or a financial decision, it is a real gap between what you can check and what the machine used. The opinion pairs that with a field experiment showing Google’s AI Overviews cut outbound organic clicks by 39.8 percent and pushed zero-click searches up 34.5 percent — a measure of how quickly the agent layer is absorbing traffic the open web used to carry.
France is not acting alone, and the opinion is careful to say so. It points to the European Commission’s June 2026 interim measures, which forced Meta to restore third-party AI-assistant access inside WhatsApp after the company had cut competitors off in January. That is enforcement, not commentary — a signal that default-placement complaints can already draw a binding response.
The Autorité’s own recommendations lean on tools that exist rather than new law. It urges full enforcement of competition rules, the AI Act, and the Digital Markets Act; floats designating Model-as-a-Service cloud marketplaces as regulated gatekeeper services under a DMA review; and insists that agent standards be developed “in a transparent, open and collaborative manner” to avoid capture by a dominant player. Readers tracking the compliance calendar can line this up against the EU AI Act’s shifting deadlines, because the two regimes are about to intersect directly over agents.
The regulator’s timing argument cuts for buyers and builders too: the cheapest moment to avoid lock-in is before it sets. A few practical moves follow directly from the opinion.
If you buy agents for a team, treat portability as a purchasing term, not an afterthought. Ask exactly how you export memory, history, and workflows, and resist wiring a critical process to a single vendor’s closed hooks — the same discipline a good enterprise agent buyer’s checklist already applies to security and cost.
If you sell through agents, do not stake a storefront on one commerce protocol while UCP and ACP are unsettled; favor the open, interoperable path even if it is more work today. If you build on agents, design for switching from the start and keep a clear eye on any default-placement dependency you do not control. And if you simply use them, remember the citation gap: for anything that carries real consequences, verify the source the agent chose not to show you.
It is worth being honest about what this opinion is not. It levies no fine and binds no one. The 84 percent figure depends on where you draw the market’s edges, and a lead this early can loosen as challengers scale and open-weight models keep improving. Concentration is a snapshot, not a verdict.
But the mechanisms the Autorité describes — defaults, data, cost, and incompatible standards — are the exact forces that hardened the platform markets that came before this one. The value of counting now, while the numbers are still moving, is that the doors get described while they are open. Whether anyone walks through them before they close is the part no opinion can decide.
Article Brief
Key Takeaways
5 Points30s Read
Bloomberg reported on July 20 that Z.ai — the Chinese lab that used to call…
If you remember Jibo, you’ll probably also remember its emotional farewell. The social robot, once…
Samsung has expanded its Galaxy lineup with the launch of five new devices at its…
Samsung has expanded its Galaxy lineup with the launch of five new devices at its…
If you remember Jibo, you’ll probably also remember its emotional farewell. The social robot, once…
At 15:59 UTC today, two of the most-used model names in open-weight AI simply stop…