Three vendors - OpenAI, Google, and Anthropic - together hold more than 84% of the AI agent market. That number did not come from an analyst deck. On 17 July 2026, France's Autorité de la concurrence adopted Opinion 26-A-05 on the competitive functioning of the AI agents sector. The regulator went further than reading vendor documentation: it built its own AI agents, asked them 550 shopping-related questions, and logged exactly which websites they visited and cited in response. If your team is mid-evaluation on which agent platform to standardize on, this opinion is the clearest outside evidence yet that the choice carries real lock-in risk.
What 84% market share actually means for teams picking tools
Market share figures in software usually describe revenue. This one describes users, sourced from Sensor Tower data as of May 2026, and it describes the deployment layer specifically - not model training. The opinion is the first to move downstream from model training to the deployment layer, where "agentic" AI tools reason, plan and execute multi-step tasks with limited human intervention.
That distinction matters. You can fine-tune a Mistral model on your own hardware and escape the training-layer concentration. But when your agents need to act - browse, call APIs, route tasks, coordinate with other agents - making an AI agent available to users depends on three main resources: access to a generative model, access to quality data, and inference capabilities. All three of those chokepoints sit inside the same vertically integrated firms the report names.
Initially designed as simple conversational agents, AI tools have gradually evolved into agents capable of reasoning, planning and orchestrating multiple tasks by coordinating with other AI agents. That evolution is why the report treats this as an urgent problem: once your internal workflows are wired to an agent platform's task orchestration, switching costs are not just API migration - they are re-wiring every automation your team depends on.
How France's regulators actually tested this
The inquiry started in January 2026 and took an unusually hands-on approach - regulators didn't just read reports, they built and deployed their own AI agents, running them through 550 shopping-related prompts to see how these systems actually behave in the wild.
What they were watching for was default bias: the tendency of an agent to surface its own ecosystem's products, data sources, or downstream services when a neutral answer was possible. What they found was a market that has quietly shifted from simple chatbots to autonomous agents capable of reasoning through problems and coordinating real-world actions - and the authority's concern is that this shift, left unchecked, could pull the broader digital economy into the orbit of a small number of vertically integrated tech giants.
The report calls for regulatory intervention on three fronts: data access, interoperability between AI systems, and how default AI placements are handled.
Teams should read "default AI placement" as a procurement warning. When you deploy an agent that lives inside, say, Microsoft 365 Copilot or Google Workspace, the default retrieval sources, default model routing, and default tool integrations are not neutral. They reflect the vendor's ecosystem, not yours.
The interoperability gap that actually bites teams
Barriers to entry are lower than for the training of generative models. However, AI agent developers' ability to scale is significantly constrained by the need to reach users, as well as by numerous barriers to expansion including data access and technical means.
This is the part most vendor evaluations miss. A competing agent builder can reach frontier model quality relatively cheaply right now - the open-weight ecosystem (Kimi K3, Inkling, Llama 3) has closed a lot of capability gap at the model layer. But reaching users through the interfaces they already use - Slack, Teams, browser, mobile - means going through distribution channels that the big three largely control.
For a practical team, this shows up as two specific risks:
- Schema lock-in: Each major platform has its own tool-calling schema and memory API. Work you build on Claude's tool-use format does not port to the OpenAI Assistants API without rewriting every tool definition. MCP helps at the connection layer, but orchestration logic above it is still platform-specific.
- Data gravity: Agents that learn from your team's Slack history, your Notion docs, or your CRM records accumulate indexed representations of that data inside the vendor's infrastructure. Exporting that learned context to a new platform is not a solved problem.
Beagle in action#tools-eval, team evaluating agent platformsWhat teams can actually do about AI agent lock-in right now
The France report is advisory - no enforcement yet. But it telegraphs where regulatory pressure lands next, and it gives procurement teams something concrete to point at when pushing back on vendor contracts.
Three practical moves that reduce exposure:
- Prefer MCP-native integrations over proprietary SDKs wherever you can. MCP (Model Context Protocol) is the closest thing the industry has to a neutral connection standard. Beagle's integrations use it precisely because the alternative is gluing your workflows to one vendor's API surface.
- Keep orchestration logic in your own codebase. The model call can live in OpenAI or Anthropic's infrastructure. The task routing, memory management, and handoff rules should live somewhere you control.
- Ask every agent vendor for a data export SLA before you sign. Specifically: can you export indexed representations of your documents and conversation history in a portable format, on demand, without vendor support tickets?
AI agents are offered either through a freemium model, with paid subscriptions providing access to advanced features, or on a pay-as-you-use basis. The pricing model matters here too: freemium onboarding is specifically how the big three build the usage depth that makes switching painful. The first six months are cheap. The migration cost, eighteen months later, is not.
The France opinion is not the last word, and it will not stop teams from using OpenAI or Google agents. The three dominant providers are dominant partly because they are genuinely good. But "genuinely good" and "safe to build deep dependency on" are different questions, and the 3,700-page answer to the second one just landed.
AI agent vendor lock-in: common questions
What is AI agent vendor lock-in?
AI agent vendor lock-in occurs when your team's workflows, data, and orchestration logic become embedded in a single provider's platform in ways that make migration expensive. The main vectors are proprietary tool-calling schemas, indexed data stored in vendor infrastructure, and distribution through the vendor's own user interfaces.
Which companies control the AI agent market?
According to France's Autorité de la concurrence, OpenAI, Google, and Anthropic together held more than 84% of the global AI agent market as of May 2026, measured by user share. Amazon, Microsoft, and Nvidia are also named as vertically integrated operators present across multiple layers of the stack.
Does MCP solve AI agent interoperability?
MCP addresses the connection layer - how agents discover and call external tools - but it does not standardize orchestration logic, memory formats, or task-handoff schemas above that layer. It reduces one lock-in vector meaningfully while leaving others open. Treating MCP support as a complete answer to portability overstates what the spec covers.
Should teams avoid the big three AI agent platforms?
Not necessarily. Capability and ecosystem maturity are real advantages. The practical move is separating which parts of your stack sit on vendor infrastructure (the model call, the retrieval index) from which parts you own (the orchestration rules, the task routing logic). The less proprietary state you accumulate inside a vendor's platform, the lower your switching cost if pricing or capability changes.
What did France's regulator actually find when it tested AI agents?
The Autorité de la concurrence deployed its own AI agents and ran 550 shopping-related prompts, logging exactly which sites the agents visited and cited. The exercise was designed to surface default placement bias - whether agents systematically favored sources within their own vendor ecosystems. The resulting opinion called for regulatory action on data access, interoperability, and default placement practices.