Claude Duet starts the conversation. The next test is closing the loop.

Initiative has an approval gate

Duet carries over account memory, skills and connected apps. Anthropic describes a morning brief, periodic checks and work such as preparing forecasts or drafting HubSpot follow-ups. Its original announcement introduces that workflow; it is a demonstration of the product’s intended behaviour, rather than evidence that every suggested task will be completed correctly.

The company says Duet can read, draft and prepare work on its own, but waits for approval before sending messages, paying, signing documents or deleting anything. Its browser runs remotely. A laptop can be closed while work continues; checking accounts every few hours also means an urgent item can arrive between checks.

Disconnecting an app does not make it forget

Update, 24 October: Anthropic’s new privacy explanation separates two controls that are easy to confuse. Disconnecting an app stops new access, but information Duet already learned from it remains until its memory is reset. The current help page confirms that past information is kept. Removing the connection is therefore not the same as removing the context.

The company also describes a separate check on outgoing messages. It is designed to block the most sensitive details, such as credentials, and require consent for other private information. An explicit instruction to share can itself count as consent, so a second approval prompt is not promised in every case. Anthropic acknowledges that this check can miss sensitive content or flag a harmless message.

For someone connecting a client inbox, the practical question extends beyond permission to open it: what should the assistant remember, and where may that information go? These are documented safeguards, not an independent test of their reliability.

A useful suggestion is only the beginning

Jing Shi, whose profile identifies him as MotusAI’s founder and a former Meta and Tencent employee, described a few hours of use. Duet checked his connectors and proposed tasks; clicking Start moved work into separate task threads. He also found channels useful for separating longer-running work. That is a concrete early account of getting started, without a published completion-rate test.

Claire Vo, who builds ChatPRD and CXO and hosts How I AI, praised the onboarding, connected tools and mobile artifacts. Her written summary also describes creating a sales channel that then asked what the channel was for. In her linked hands-on review, she reports email and Slack triage, ticket shopping and sales-deck work. She liked the design but wanted better prioritisation and follow-through.

Vo explicitly says she was unsure whether notifications were enabled when discussing the lack of proactive nudges. Neither account supplies a matched workload, repeat runs or a controlled comparison. Her frustration and Shi’s positive first impression can both inform the product questions without becoming a verdict about every user.

The preview boundaries matter

The product FAQ says Team and Enterprise plans are not included. European subscribers are outside this US preview, and a waitlist entry does not promise an invitation date. Background work shares the same plan limits as ordinary Claude use; an included assistant is not unlimited background capacity.

The help documentation is more specific than the broad connector pitch: Microsoft 365 is not available yet, some connectors such as Slack have limits, and users cannot watch or take over Duet’s browser. It also says Duet cannot see all projects or local computer files. Existing local skills should not be assumed to appear automatically in the cloud assistant.

The next test is the handover

For a marketing team, a drafted follow-up or campaign brief becomes useful when the sources, proposed action and remaining decisions are clear. The early reports make that handover worth watching: can the assistant keep context, surface the right task and return it at a useful moment?

A fair follow-up needs the task instructions, connected services, notification settings, elapsed time and finished output. It should count corrections and abandoned tasks alongside successes. For now, Duet offers a consequential new workflow with visible preview limits; the evidence supports examining those limits before promising saved hours.