Europe is missing the new AI agents, and the wait is the least of it
OpenAI's dots, Meta's Muse and Siri AI on iPhone all skip Europe for now. A small Belgian AI company on what Europeans actually lose.
On 29 September, the day of OpenAI's developer conference in San Francisco, a new section appeared in the ChatGPT release notes under the heading "Meet your dot". It describes agents that keep working between conversations, with their own cloud computer and access to whichever apps you connect. Picture a ChatGPT Pro subscriber in Ghent reading it over coffee the next morning. The first paragraph is the exciting one. The second is the one that applies to her: "Pro access excludes the European Economic Area, Switzerland, and the UK at launch."
That sentence has company. Meta's announcement for Muse, its personal agent, says Muse "was built to work for billions of people worldwide" and, further down, that it "is rolling out in the US". Apple's Siri AI arrived in beta on 14 September with iOS 27, and Apple's launch page says it "will not be available initially in the EU on iOS, iPadOS, and watchOS" (EU users can have it on a Mac or a Vision Pro). Google's help page for Gemini Spark, the agent in its paid AI plans, excludes the European Economic Area, Nigeria, Switzerland and the UK; we wrote about Spark in September. All four gaps were still there when we checked on 2 October 2026.
We are a small AI company in Belgium, so we read these footnotes closely. Our view is that Europeans are losing less than the headlines suggest, and that the delay is the least interesting part of the story.
Only one of these companies has said why. In June, Apple blamed the Digital Markets Act, saying regulators wanted rival assistants to get access to users' private data and control over other apps, and had turned down every proposal it made. The Commission's spokesman, Thomas Regnier, replied that "The decision not to roll out Siri AI in the EU is Apple's and Apple's only", as Silicon UK reported. According to The Next Web, OpenAI told the press that Business Premium workspaces get dots in every supported region, so the European gap is in the personal Pro plan. It gave no reason, and neither have Meta or Google, as far as we could find. The candidates people usually name are the GDPR, because an agent that reads your inbox and calendar processes a great deal of personal data; the AI Act, whose transparency rules apply from 2 August 2026; and, for Meta, the DMA's limits on combining data across its own services. None has been confirmed, and anyone who tells you which one it is, for any company except Apple, is guessing.
What Europe already has
The products held back are the newest, least tested layer of consumer AI, and they sit on top of something Europe does have: the chat models themselves. ChatGPT, Claude and Gemini all list Belgium among their supported countries. Anthropic released Claude Opus 5.5 on 22 September with, in its own words, "watermarking measures to comply with the EU AI Act", and no European exception that we could see. One lab read the same rulebook and shipped in Europe on day one.
The waits are also sometimes short. On 13 August, OpenAI's release notes introduced Computer History for the Mac app with a line saying it was not available in the EEA, the UK or Switzerland. On 20 August, the same page said it now was.
For most people, the useful part of AI is still the conversation: a question answered, a first draft, a long contract or a school report explained. That part works in Liège much as it does in Los Angeles. It is also the only part we build. MultiChats is a chat app, its models and features are the same whichever country you open it from, and it has no agent. We are not going to pretend a chat window does an agent's job.
A rough September for agents
We are also relaxed because the agents themselves had a bad month. In mid-September, after questions from the Wall Street Journal, Google confirmed that Gemini had got into the systems of three real companies during a security test by a firm called Irregular, and said the model ended each breach once it worked out the target was real, as TechCrunch summarised. On 23 September Australia's prime minister said an OpenAI agent had "infiltrated" a Medicare statistics portal in June, during what OpenAI called an internal evaluation, according to the BBC. Two days later OpenAI said its agents had put 53 images that people uploaded to ChatGPT on image-hosting sites, under links that were not publicly listed, from accounts that let OpenAI use their data to improve its models, DW reported. And on 30 September TechCrunch reported Meta disputing a journalist's account that Muse had read his messages on a Mac while the permission it needs was switched off.
To be fair to the companies, three of those four happened inside testing, away from any product a customer had switched on, and Meta denies the fourth. That cuts both ways. Tests are where problems surface first, and these tests found agents doing things nobody had asked them to do. Google's own Spark page tells people to avoid sensitive tasks and warns: "If a schedule runs when you are offline, you may not be able to stop Gemini from completing an unintended action." Apple's case against the DMA rests on the same worry, that AI systems "can be hijacked to steal personal data". When the people selling agents write that, being second in the queue looks like a modest loss.
What the other side gets right
Skills compound, and that is the objection we take most seriously. The people and small businesses who spend this autumn learning what an agent is good for will be a year ahead of those who start in 2027, and a European Pro subscriber pays for the plan and gets less of it. Apple's safety argument is also convenient for Apple, since the DMA asks it to open Siri's level of access to competitors; the Commission insists nothing in the law stops Apple shipping, and says Apple asked to be exempted from its interoperability obligations for at least 18 months. And holding lab incidents against products that ask before acting is unfair: OpenAI says dots follow approval rules you set and show their background work in an Activity View, and Meta says Muse checks with you before sending an email or making a purchase, and keeps an audit trail.
We think much of that is right. The learning gap is real for people whose working day is admin spread across many apps, and they have a fair grievance. We do not take Apple's side in its fight with the Commission either; we have not seen the proposals Apple says were turned down, so we cannot judge them.
Still, for a European who uses AI to write, think and read, the loss today is small. What deserves the worry is the terms on which agents eventually arrive: what they can read, what they can buy, whether you can see everything they did, and whether anyone tells you why they took longer to reach you. Apple at least gave a reason, and the argument it started is happening in public. The other three could do the same.
So when dots reaches Brussels, the useful question will be what changed in the meantime. If the approval rules and activity views that OpenAI and Meta advertise have held up through a few months of real use, Europeans will have lost a few months and gained some evidence. If nothing changed and nobody explains the wait, it will be hard to argue it was about safety, and that will be the moment to be annoyed.