Claude Duet: What a Proactive AI Agent Means for Swiss SMEs

On October 21, 2026, Anthropic opened a research preview of Claude Duet, a standalone app in which Claude works as a proactive agent. Instead of waiting for a prompt, it watches your connected apps, sends a morning briefing and asks you before it takes important actions. For a small company, that sounds like an assistant. For IT and compliance, it is a new kind of employee with access to your mailbox and calendar. Here is what we can confirm, what we cannot, and what to decide before anyone on your team switches it on.

What Is Reported So Far

Coverage is based on Anthropic's announcement as relayed by news outlets. We could not verify every detail against Anthropic's own pages, so treat specifics as reported:

• Access: a waitlist for US Pro and Max subscribers on iOS, Android and claude.ai. It is not announced for Switzerland or the EU.
• Behaviour: the agent uses connected apps and a cloud-based browser and keeps working while you are away. It runs in the cloud, not on your device.
• Rhythm: a daily briefing, periodic checks of connected apps through the day, and a message only when a decision is yours.
• Approvals: reportedly it asks before sending messages, paying, signing or deleting.
• Model and price: reportedly built on Claude Opus 5.5, with no separate price beyond the existing Pro and Max plans.
• Competition: coverage frames it against Meta's Muse and OpenAI's dots.

Why "Proactive" Changes the Risk

A chat assistant only sees what you paste in. A proactive agent reads continuously and acts on its own schedule. Three things follow:

1. Standing access. The agent holds credentials to email, calendar and other tools, so a mistake or a compromise has a much larger blast radius. We described this pattern in Shadow AI agents and the governance gap.
2. Untrusted input. Every email the agent reads is text written by someone else. A malicious message can try to steer the agent (prompt injection). Approval prompts help, but only if people read them rather than click through.
3. Approval fatigue. An agent that asks ten times a day teaches users to approve without looking. The safeguard is only as strong as the attention behind it.

None of this is unique to Duet. It applies to any agent that acts across your tools, and it is a core reason many enterprise agents fail in production.

Data Protection Questions for Swiss SMEs

Before connecting a mailbox, clarify:

• Whose data is in there? Customer, employee and supplier data in email falls under the Swiss FADP and, for EU contacts, the GDPR.
• Where is it processed? A US-first research preview gives you little to base a transfer assessment on. Ask for processing locations and retention terms.
• Who owns the account? A personal Pro subscription connected to a business mailbox is shadow IT. Use company-managed accounts or keep it off business data.
• What do regulators expect? See our overview of the Swiss AI regulation consultation draft and the EU AI Act transparency obligations.

A Safe Way to Pilot a Proactive Agent

1. Start read-only. Grant access to a test mailbox or a non-sensitive calendar first.
2. Define actions that always need approval: sending externally, payments, signatures, deletions, changing access rights.
3. Keep a log. You should be able to answer "what did the agent do last Tuesday?"
4. Name an owner per agent who reviews its activity weekly.
5. Plan for portability. Agent products change fast, and vendor decisions can change faster, as in the Pentagon and Anthropic case.
6. Measure value. Track time saved and errors caught against the effort of supervising it.

The Bottom Line

Claude Duet is a signal of where the market is going: assistants that act, not just answer. It is not yet available to most Swiss companies and it is a research preview, so there is no need to rush. There is a need to prepare. Companies that already have an agent policy, clean access rights and a review habit will adopt these tools safely; the others will meet them as shadow IT. If you want to build that foundation, talk to our team.

Sources: Let's Data Science: Anthropic opens Claude Duet research preview · BetaNews: Proactive AI agent Claude Duet · VKTR: Claude Duet previews an AI agent that messages you first · Braindetox: Claude Duet approval rules and access