Writing

Claude Duet in Europe? Why Always-On AI Agents Stop at the Border — and Dots Skips Consumers

10 min read

Glowing abstract AI assistant paused at a line of light dividing a lit city from a dim northern town at night
An agent that works while you sleep waits at the border: Claude Duet is US-only, and Dots skips Europe’s consumers. Illustration: AI-generated.

No – per October 2026, Claude Duet is not available in Europe. Anthropic’s always-on agent, launched as a research preview on 21 October, is open only to Pro and Max subscribers in the US, via a waitlist. OpenAI’s Dots draws a similar line: it skips Pro users in the EEA, Switzerland and the UK, but reaches European business workspaces.

My thesis: an agent that works while you sleep is not a feature but a processing operation – with a purpose, a legal basis, logs, retention and an approval trail someone must be able to show an auditor. That is why it reaches European consumers last. I build clinical AI in Norway under exactly those constraints, and the border these launches drew says more than the launches themselves.

  • Claude Duet is Anthropic’s always-on personal agent: Opus 5.5, its own cloud computer, more than 6,000 connectors and plugins.
  • Availability: US Pro and Max plans only, via a waitlist; Team and Enterprise “not included yet”.
  • Claude itself is supported in all 27 EU countries plus Iceland, Liechtenstein and Norway – the geofence is product-specific.
  • OpenAI Dots (29 September 2026): GPT-6 Astra, one cloud computer per agent, more than 4,000 apps.
  • Dots in Europe: no Pro access in the EEA, Switzerland or the UK; Business Premium workspaces get it.
  • Reasons given: none, from either company.

Is Claude Duet available in Europe?

Not yet, and no date is announced. The Claude Duet product page says it plainly: “Available in Pro and Max plans in the US.” The FAQ calls it “a research preview on a waitlist” and says Team and Enterprise plans are not included yet.

Claude itself is not missing from Europe. Anthropic’s list of supported countries (read 25 October 2026) includes every EU member state, Iceland, Liechtenstein and Norway. A Norwegian Max subscriber can use Claude today – just not the one product that acts on their behalf while they are away.

Dan Shipper at Every, who had two weeks of early access, called Duet “the first one we’d call a pro tool” in his launch-day review on 21 October 2026, while noting that its scheduled runs “can’t reach my custom connectors”.

Claude Duet vs OpenAI Dots: where does each one stop?

Architecturally they are near twins: a dedicated cloud computer, broad app access and an approval step before consequential actions. They differ in model, reach and where the line is drawn.

Claude Duet

Duet “asks you before sensitive operations like sending messages or paying”, and its product page says approvals, logins and safety checks “run in your own cloud computer, so sensitive details stay private”. Users “can see what Claude Duet has done and is doing for you, and steer it.” Anthropic is candid about the limits: Duet “may misread a thread or miscalibrate what’s urgent, and it can miss something that arrives between checking in with you.” That admission is the reliability gap I wrote about earlier this year, now in a consumer product.

OpenAI Dots

According to SiliconANGLE’s DevDay report (29 September 2026), Dots has a more explicit control model. When an agent does proactive research without a task, “its connections to the owner’s apps are read-only during that time, so it cannot send messages or change anything.” Actions touching accounts or sharing information “must first clear a check called auto-review”.

The line on the map

The detail most coverage missed: Dots is not US-only. SiliconANGLE reports that Pro subscribers in the EEA, Switzerland and the UK are excluded, while enterprise customers “including those on Edu and Healthcare plans, can try a beta once a workspace administrator switches it on”. A European hospital can, in principle, pilot Dots; a European Pro consumer cannot.

Why do always-on agents skip Europe?

Neither company has said why. What follows is my interpretation.

The Dots split is the clue. The line does not run between the technology and the law; it runs between who is the controller and who has signed a data processing agreement. In a business workspace, the customer is the controller, a contract makes the vendor its processor, and an administrator decides whether the agent runs. In a consumer account, the vendor sits much closer to being controller for an agent that reads the user’s inbox and documents – and the personal data of everyone who ever emailed them.

That is what separates always-on agents from chat. A chatbot processes what you paste into it; an always-on agent continuously processes what arrives, about people who never agreed to anything.

One caveat weakens a pure “EU regulation” story: the UK and Switzerland are outside the EEA, yet Dots skips them too. I read that as consistent – UK GDPR and the Swiss data protection act impose similar duties – but it means the gate is about data-protection regimes in general, not Brussels in particular.

What does GDPR require before you deploy an always-on agent?

The General Data Protection Regulation (EU) 2016/679, which applies in Norway through the Personal Data Act, never mentions agents. It does not need to.

Purpose limitation and storage limitation – Article 5(1)(b) and (e)

Personal data must be “collected for specified, explicit and legitimate purposes”, and kept “for no longer than is necessary”. An agent told to “find ways to help me” has no specified purpose; one that keeps state indefinitely has no retention period. Both are design problems, not paperwork.

Accountability – Article 5(2)

The controller “shall be responsible for, and be able to demonstrate compliance”. This turns an approval button into an evidence requirement: if you cannot show what the agent did, when, under whose approval and on which data, you cannot demonstrate compliance – however well it behaved.

Data protection by design and by default – Article 25

By default only necessary data may be processed – covering “the amount of personal data collected, the extent of their processing, the period of their storage and their accessibility”. With 6,000 connectors, “by default” means most are off.

Impact assessment – Article 35

A data protection impact assessment (DPIA) is required before processing “using new technologies” that is likely to result in a high risk. An agent continuously reading third-party communications is a textbook case.

Automated decisions – Article 22

People may not be subjected to solely automated decisions with legal or similarly significant effects. An agent that sorts email usually falls outside this; one that pays, declines, cancels or signs may not.

What should “asks you before sensitive operations” mean in an audit?

An approval only counts if it leaves a record answering five questions: who approved, what exactly (the payload, not a summary), when, under which policy version, and what the agent did next.

The Duet product page promises visibility and steering. It does not describe an audit log, a retention period or a deletion mechanism. They may exist in the preview – but for a European deployer, undescribed is unavailable until a contract says otherwise.

Anthropic’s own usage policy update (8 October 2026, effective 12 November) points in the same direction. For uses that can affect someone’s “health, legal rights, finances, livelihood, or access to essential services”, it requires a qualified “human in the loop” – “someone who has the authority to review and, if necessary, change Claude’s recommendations” –. A sleepy tap at 07:00 is not that.

How we run agents in regulated healthcare

At MediVox and Eir Tec every one of these questions comes up in procurement, long before anyone asks about model quality. I described the pattern in why agentic AI hits a wall in regulated environments. Our principles for anything agent-shaped:

  1. Approval queues, not pop-ups. Consequential actions wait in a queue a qualified person works through; the queue is the log.
  2. Write locks by default. Agents read; writing is a separate, narrower permission. Dots’ read-only proactive mode is the same idea.
  3. A sandbox with an egress policy. The agent should not be able to reach hosts nobody approved.
  4. Credentials outside the agent. A model that can be prompt-injected should never hold a secret – the lesson of the Semantic Kernel prompt-injection RCE.
  5. Logging we own. A vendor dashboard is not evidence.

NVIDIA OpenShell, an Apache-2.0 runtime currently at version 0.1.x, runs each agent in an isolated sandbox where “every network connection passes through a policy check before it leaves the sandbox”, and agents “never see real credentials”. The 1Password integration announced on 28 September 2026 (beta) gives the agent placeholders while the gateway injects real secrets only for approved hosts, for a defined time to live. Its controls “are enforced outside the agent, so the agent cannot change the rules that govern its own access.”

Neither source describes an audit log of agent actions. That is the layer you add yourself – nobody should call this stack “audit-ready” out of the box.

Is an always-on agent a high-risk AI system under the AI Act?

Not by itself – a personal assistant is not an Annex III use case – and the high-risk rules have moved anyway: according to the Norwegian government’s EEA note on the AI omnibus (3 June 2026), Annex III obligations apply from 2 December 2027 and Annex I from 2 August 2028. Transparency duties already apply, as I covered in what the AI Act requires in October 2026. The constraint on always-on agents today is GDPR, not the AI Act.

For Norwegian organisations, the closest public experience is Datatilsynet’s regulatory sandbox for AI, a controlled test environment where NTNU and the regulator ran a project on Copilot, an assistant that reads mail and documents inside an organisation. That is a better starting point for an agent DPIA than a vendor’s waitlist.

What I would take from this

The border is not a rollout bug. It shows where the compliance work lives: in the deployer’s contract, logging and approval design, not the model. Three predictions to check me on:

  • By 30 June 2027, Claude Duet will be available to Team or Enterprise customers in the EEA before it is available to Pro or Max consumers there.
  • The first always-on agent offered to EEA consumers by Anthropic or OpenAI will ship with an exportable activity or approval log documented on its product page.
  • By the end of 2027, at least one European data protection authority will publish guidance or a sandbox report that names always-on agents specifically.

If you run a regulated organisation, do not wait for the waitlist: write the purpose, retention rule and approval-log design now, and run the DPIA on the business-tier products that already reach Europe.

Frequently asked questions

Is Claude Duet available in Europe?

No. Per October 2026, Claude Duet is a research preview for Pro and Max subscribers in the US only, via a waitlist, and Anthropic has not announced a European date.

Why are always-on AI agents like Claude Duet and OpenAI Dots not available to European consumers?

Neither company has given a reason. My reading is that an agent that continuously processes other people’s personal data needs a defined controller, purpose, retention and audit trail – which a business workspace with a processing agreement provides and a consumer account does not.

Can I use OpenAI Dots in Norway?

Not on a ChatGPT Pro plan, which excludes the EEA, Switzerland and the UK. Business Premium workspaces get Dots in all supported ChatGPT regions, according to OpenAI’s statements to the press.

What does GDPR require before deploying an always-on AI agent?

A specified purpose and storage limit (Article 5), the ability to demonstrate compliance (Article 5(2)), data protection by design and by default (Article 25), and usually a DPIA before processing starts (Article 35).

What is the difference between Claude Duet and OpenAI Dots?

Duet runs on Opus 5.5 with more than 6,000 connectors and asks before sensitive operations; Dots runs on GPT-6 Astra with more than 4,000 apps, a read-only proactive mode and an auto-review check. Duet is US-only for now; Dots reaches European business workspaces.

Portrait of Håkon Berntsen

Written by

Håkon Berntsen

CTO and co-founder of Skjld Labs AS. Builds production AI agents, Microsoft 365 systems and the courses around them, from Sandefjord.

About LinkedIn GitHub