Skip to content
← BlogStrategyAnalysis7 min read

Three companies control 84% of the AI agent market, says France's competition watchdog

France's competition watchdog found OpenAI, Google, and Anthropic hold over 84% of the AI agent market — what that concentration means for buyers and builders.

By David Lara, Founder

Founder-reviewed ·How we research and correct articles

France’s competition authority spent six and a half months building its own AI agents. It gave them 550 shopping-related questions to answer and logged, query by query, which websites they visited and which sources they cited. The result, published 17 July 2026 as Opinion No. 26-A-05, runs to more than 3,700 pages with annexes, and its headline finding is blunt: OpenAI, Google, and Anthropic together hold more than 84% of the AI agent market, based on Sensor Tower usage data gathered in May 2026.

Three companies. Eighty-four percent. That’s not “a competitive market with a couple of leaders” — it’s the kind of concentration regulators usually see in mature, capital-intensive infrastructure markets, not one that’s barely two years past its consumer-facing launch.

What the Autorité is actually worried about

The opinion isn’t primarily about market share as a number — it’s about what happens next if that number holds. Competition-law analysis of the opinion frames the risk in three parts: platformization (agents becoming the layer users interact with instead of the underlying services, giving the agent provider disproportionate influence over which services get seen at all), disintermediation (the agent provider inserting itself between a business and its own customers), and algorithmic collusion (concentrated infrastructure making coordinated pricing or ranking behavior easier to sustain, even without explicit agreement). The Autorité’s own reporting on its shopping-agent tests supports the first two directly: which sites an agent chooses to visit and cite is now effectively a distribution decision made by three companies’ models, not by search rankings or direct traffic the way it used to be.

Why concentration this extreme matters if you’re a buyer

You don’t need to be a regulator to feel the downstream version of this problem. Three vendors controlling 84% of a market you depend on translates into concrete business risk:

Lock-in. Build meaningfully on one vendor’s proprietary agent framework — its tool format, its orchestration layer, its specific behaviors — and switching later means rebuilding, not reconfiguring. The fewer credible alternatives exist, the more that lock-in costs you at renewal time.

Pricing power. A market with three players controlling 84% of usage has far less competitive pressure holding prices down than a fragmented one. When the Autorité warns about the digital economy concentrating around “a small number of vertically integrated firms,” pricing leverage over everyone building downstream of them is the practical version of that warning.

Fewer places to point the finger. If an agent from one of three providers makes a decision that costs you — a wrong recommendation, a biased ranking, a service outage — your recourse is whatever that one vendor’s terms allow. Concentration doesn’t just raise switching costs; it narrows your options when something goes wrong.

The Autorité’s own prescription: interoperability, not less AI

Notably, the opinion doesn’t call for less AI agent adoption. It calls for full enforcement of existing competition rules plus interoperability and open standards — the regulatory version of “don’t let the winners lock the door behind them.” That’s the same instinct worth applying at the level of an individual buying decision, well before any regulation actually bites: prefer building on interfaces more than one vendor supports over building exclusively inside one vendor’s proprietary agent framework.

Single-vendor agent frameworkOpen standard / public API
Switching cost if the vendor changes termsHigh — logic tied to one proprietary shapeLower — the interface itself doesn't change
Exposed to one vendor's pricing power
Works with tools from other vendors without a rewrite
Your integration outlives a single product's roadmap

This isn’t a claim that any one open standard has “won” or that proprietary agent frameworks are never the right tool for a given job — it’s a claim about where the risk sits when a market gets this concentrated. Betting your integration work on interfaces that stay usable regardless of which three companies are ahead this year is a hedge, not a prediction about who wins.

Software markets have concentrated like this before

This isn’t the first time a new layer of computing consolidated around a handful of vendors this quickly. The browser market went through a similar arc: dozens of competitors through the 1990s and 2000s, a long middle period where two or three browsers split most usage, and today a single rendering engine underpins the browser most people actually use, whatever brand name is on the toolbar. Cloud infrastructure followed a comparable pattern — from a wide field of hosting providers to a market where a small number of hyperscale platforms account for most enterprise workloads. In both cases the concentration didn’t stay theoretical for long: it showed up as pricing power, as switching costs that grew every year a customer stayed, and as smaller competitors either getting acquired or exiting the category entirely.

The AI agent market looks like it’s following the same curve, just compressed. What took the browser market roughly a decade and cloud infrastructure closer to fifteen years appears to be happening to AI agents in about two. Read that way, the Autorité’s opinion is less a verdict on today’s market share and more an early warning aimed at compressing that timeline before the same downstream effects — pricing power, vendor-specific lock-in, fewer places to appeal when something goes wrong — become as entrenched as they did in those earlier markets.

A few things worth clarifying

Does an 84% share mean this market is operating illegally?

No. An advisory opinion from a competition authority is a warning and a set of recommendations, not a finding of wrongdoing or a binding ruling. What it signals is that regulators are now paying close attention to how agent providers behave — and that today's concentration is the baseline any future enforcement action would be measured against.

Why did this market concentrate so fast, compared to past software categories?

The Autorité's own testing points to the mechanism: a foundation-model provider controls both the agent's reasoning and, increasingly, which outside sources it trusts enough to cite or transact with. That combination — the intelligence layer and the distribution layer owned by the same three companies — compounds faster than concentration in a market where those two things were built by different players.

What should a smaller company actually do about this today, practically speaking?

Nothing dramatic, and nothing that requires waiting for regulators to act. The practical version is to prefer documented, open interfaces over proprietary ones wherever you have the choice — for AI agent tooling and for everything else you build on. That doesn't mean avoiding the three leading providers; it means not writing so much business logic against one vendor's proprietary shape that switching later becomes a full rebuild instead of a reconfiguration. The interoperability the Autorité is asking regulators to require is the same thing you can simply choose to prefer on your own, years before any rule forces the issue.

Where this lands for your own stack

The same logic that applies to picking which AI agent providers to build around applies to picking the tools underneath your outreach program. This is exactly why Norbelys is built on a public REST API that any client, script, or agent framework can call — rather than a platform that only works inside one vendor’s proprietary agent shell. That’s the same hedge the Autorité is recommending at the market level, applied to your own vendor selection: API-first tooling and a Norbelys MCP server built on that same public interface both mean your integration work doesn’t have to bet on which three companies still control 84% of the agent market next year. Whichever agent framework ends up ahead, it can still reach your sending program through the same documented API — nothing about switching agent vendors requires switching outreach platforms too.