The Autorite de la concurrence did something unusual to prepare for this report. It wrote 550 questions about buying things (350 purchase-intent questions and 200 follow-ups across electronics, household appliances, hotel bookings and transport tickets) and put them to ChatGPT and Gemini through their APIs, with web search switched on. It logged which websites each agent consulted, then compared them with the sources cited in the answers. The experiment ran from May 20 to May 30, 2026. The results are in Annex 2 of Opinion 26-A-05.
That methodological choice matters. France's competition authority did not take vendor self-reporting at face value and did not rely on general AI anxiety. It ran its own test. The resulting opinion (97 pages in the English version) is credible in a way most regulatory documents about AI are not.
The headline finding: OpenAI, Google and Anthropic together held more than 84 percent of the global AI agent market in May 2026, according to Sensor Tower data the Authority cited. Three companies. 84 percent. In a market that barely existed two years ago.
TL;DR: France's Autorite de la concurrence published Opinion 26-A-05 on July 17, 2026, finding OpenAI, Google and Anthropic hold over 84% of the global AI agent market. The report flags default integration into existing platforms, barriers to data portability, and how agentic commerce protocols such as Google's UCP and OpenAI's ACP are governed. It is advisory, not punitive, but enterprise teams buying AI agent tools now are making vendor decisions that will be harder to reverse in 18 months than they appear today.
The opinion is advisory. It accuses no one of breaking the law and notes explicitly it is not assessing conduct against Articles 101 and 102 of the Treaty on the Functioning of the European Union. What it does instead is describe how a market built as a tool becomes a gateway, and which structural features make that transition hard to reverse.
What changed between a chatbot and an agent
The Authority draws a clear line between the category it reviewed in Opinion 24-A-05 from June 2024, which looked upstream at model training and infrastructure, and what it examined here. Opinion 26-A-05 moves downstream to the layer where AI tools plan, reason, and take actions across multiple steps without moment-to-moment human direction.
The European Commission's own definition, quoted in the report, describes AI agents as "software applications designed to perceive and interact with the virtual environment" that "operate autonomously, meaning they are not directly controlled by a human." This is not a semantic distinction. The lock-in implications are different when a tool is executing a workflow rather than answering a question.
A chatbot you can replace with a different chatbot. The prompts carry over, the context resets, the outputs are roughly comparable. An AI agent that has been running your procurement process for six months, or scheduling your team's calendar for a quarter, has accumulated something closer to institutional memory. Switching means rebuilding accumulated context, reintegrating with connected tools, and accepting some loss of workflow continuity. The Authority found that switching is held back by migration costs and by the loss of accumulated data such as conversation histories and project spaces, and it recommended effective data portability between agents.
Why entry is easy but growth is not
One of the opinion's more careful arguments distinguishes between market entry and market scale. Building an AI agent has become comparatively accessible. Capable models are available through APIs. Open-weight releases lower the technical floor. Mistral AI, a French company, is the one European name that keeps coming up: most respondents to the public consultation called Mistral's Le Chat "the only European alternative," and others said French solutions like it "do not have the same investment capacity, infrastructure or large-scale adoption as the global US operators."
The gap between building something and reaching users is where the 84 percent figure starts to make sense. Google's Gemini may come pre-installed on Android. Microsoft's Copilot ships inside Office 365. Meta AI can be reached inside WhatsApp conversations. The French Authority was watching what happened in January 2026, when new WhatsApp Business Solution terms, introduced in October 2025, took full effect and excluded Meta AI's competitors from the platform. Italy's competition authority ordered Meta to suspend those terms in December 2025, and the European Commission imposed interim measures on June 9, 2026 to restore free access to WhatsApp for rival AI assistants. The French opinion cites these cases as the lock-in pattern to watch.
Two structural advantages compound for the incumbents beyond distribution. Access to usage data at scale improves model quality in ways smaller providers cannot replicate from a standing start. And the total cost of inference, the compute required each time a model is used, can substantially exceed the one-off cost of training, because it recurs with every use. The Authority cites an EY study putting the cost of a response at four US cents in 2023 and the average cost of an "orchestration" by an AI agent (organising a complex task) at an expected 1.20 US dollars in 2026. At that cost level, a new entrant needs revenue to fund the compute required to serve users at quality, but needs quality to attract the users who generate the revenue.
The Authority's own experiment captured one dimension of this concretely. ChatGPT, in the Authority's testing, consulted Reddit for 87.4 percent of its discoverability questions, yet cited Reddit directly in only 1.0 percent of its responses. The sources an agent visits most are not necessarily the sources it tells users it used. Gemini was less transparent still: it returned its answers without citations, and the sources it consulted were visible only in the API metadata.
Three sets of concerns the Authority documented
The opinion closes with three sets of recommendations rather than binding orders.
The first set is mostly about using what already exists. The AI Act, the Digital Markets Act, the Digital Services Act, GDPR and the Data Act may all apply to the sector, and the Authority recommends making "full and swift use" of that framework, with soft law such as guidelines before any new rules. Within this set it asks for particular vigilance on three points. First, equity investments by major digital companies in competing AI agent developers, and partnership agreements between them. Second, default placement: whether users can actually choose and use an AI agent other than the one integrated by default into a platform they already depend on. Third, the key parameters that influence how an agent selects, ranks or recommends services. It also wants competition authorities to watch AI agent distribution channels, particularly "Model as a Service" (MaaS) platforms, the cloud marketplaces through which businesses reach third-party AI models, and it asks the European Commission to consider, in its Digital Markets Act work, whether marketplaces that distribute AI models should be designated as core platform services.
The second set asks operators to make agents interoperable and portable: users should be able to "switch from one AI agent to another without significant loss of information or functionality."
The third set is about open technical standards. The standards governing agentic commerce "should be developed and maintained through transparent, open and collaborative processes, in order to prevent any situation in which a dominant operator exercises excessive control or influence."
That last recommendation lands on a split the market has not resolved. Google's Universal Commerce Protocol launched in early January 2026, developed with a coalition that includes Shopify, Walmart, Carrefour, Visa and Stripe. OpenAI's Agentic Commerce Protocol, developed with Stripe, has been live since September 2025. Both govern how AI agents complete purchases on behalf of users, and their architectures differ. An enterprise that integrates its procurement workflow with one and wants to switch AI vendors later faces a protocol problem alongside the data problem.
The commerce projection is the number to track
Traffic from AI agents to e-commerce websites in France is still below 5 percent, the Authority says, and some respondents to its consultation estimate it could approach 25 percent by 2030. If that trajectory is even roughly accurate, the vendor an enterprise has integrated into its purchasing workflows by late 2026 will be embedded deeply enough by 2028 that switching becomes a significant project, not a checkbox item.
ChatGPT reached over 900 million weekly active users and 50 million subscribers worldwide as of March 2026, according to OpenAI's own figures. Google reported in May 2026 it was processing more than 3.2 quadrillion tokens per month, seven times the volume of a year earlier. These numbers are growing faster than enterprise procurement cycles typically move. The lock-in question is not theoretical. It is a function of how quickly your current integration becomes load-bearing before your next vendor review date.
The opinion also documented that advertising is starting to appear inside AI agents. OpenAI began testing ads in ChatGPT in January 2026. Perplexity dropped its own advertising tests early in 2026. Anthropic announced in February 2026 its intention to keep Claude ad-free. The divergence in monetisation strategy matters for enterprise use cases: an agent that serves ads may not have the same response priorities as one funded entirely by subscriptions.
What to ask before signing an AI agent contract
The Authority's findings describe a structural situation. The procurement question is what to do about it before you are inside it.
Before integrating an AI agent into any workflow where switching would be disruptive, ask the vendor three things in writing.
First: what data can be exported, in what format, and on what timeline? That includes historical conversation records, usage preferences, any fine-tuned context the agent has built from your team's patterns, and integration configurations. If the vendor cannot specify a standard export format, the implicit answer is that the data stays with the vendor.
Second: which agentic commerce protocols does the product support? UCP, ACP, both, or neither? If neither, and the vendor says it plans to support one eventually, get the commitment in the contract with a specific timeline and a termination right if the timeline slips. Two parallel, non-interoperable standards mean your protocol choice today is also a vendor choice.
Third: what is the actual exit path? The cost of vendor migration is rarely the subscription fee. It is rebuilding the connections to your other tools and reconstructing the process context the agent has accumulated. Ask the vendor to estimate this explicitly. If they cannot or will not, that is a useful signal about how they think about retention.
The Autorite de la concurrence did not conclude that any company is doing something illegal. The opinion is an early warning about market structure, issued by a regulator that went to the trouble of testing the leading agents itself to see what was actually happening. That level of care is worth taking seriously by the enterprise teams making these vendor decisions now.
Correction, October 22, 2026: an earlier version said the Authority built its own AI agents for its test (it put its questions to ChatGPT and Gemini through their APIs), called the opinion 3,700 pages long (the English text runs 97 pages), said inference costs now exceed training costs (the opinion says they can), presented the 25 percent 2030 traffic figure as the Authority's projection (it is an estimate from some consultation respondents), said UCP and ACP are not interoperable (the opinion says their architectures differ), grouped the recommendations differently from the opinion and paraphrased several quotes; those passages now follow the official text (Opinion 26-A-05, English version).
Related Reading
- Vetting AI Tools: Fake Apps, Malware, and Typosquatting Risks
- AI Vendor Contract Red Flags: What to Check Before Signing
- Anthropic Export Ban: What the 17-Day Fable 5 Shutdown Means for Your AI Vendor Policy
- MCP Server Security Governance Checklist 2026
- FTC AI Enforcement Actions 2026: What the Pattern Means for Small Teams
- AI Spend Governance: Token Budget Controls and Cost Management 2026
- CNIL's Agentic AI Note: Three GDPR Risks Your Team Needs to Address
- Starbuck v. Google: AI Defamation Survives Dismissal, Heads to Trial
- China's Moonshot Accused of AI IP Theft. Treasury Threatens Sanctions.
Sources: Autorite de la concurrence, Opinion 26-A-05 of July 17, 2026, English full text (published September 28, 2026), Autorite de la concurrence press release, July 17, 2026, French press release, ppc.land coverage of Opinion 26-A-05, Concurrences summary.
