Claude Duet catches missed details but can overreact
Anthropic’s Claude Duet spent a week helping one reviewer manage a busy work and travel schedule, spotting overlooked deadlines and linking information across email, Slack and shared documents. It also misread a flexible deadline as urgent, showing how an AI assistant can be useful without understanding workplace nuance.
Duet is a consumer-facing AI agent aimed at “professionals and small business owners,” rather than a tool for ordering food or writing code. Anthropic calls it a “research preview”; access is limited to Pro, Max and Ultra subscribers in the US, with a waitlist. Users connect services such as email, calendars, Google Docs and Slack, then review suggested tasks through swipeable cards in Duet’s app.
The reviewer connected all of those services. Duet used an automatically generated summary of a staff meeting, an email and a Slack message to connect related information. It also inferred who the reviewer’s boss was from a shared document used for weekly one-on-ones. Each morning, it offered a short daily briefing and reminders, including a prompt to pack an item needed for a video shoot.
Some suggestions were less well judged. After the reviewer had told an interviewee an article would likely appear Tuesday, Duet flagged early Tuesday that the draft was still awaiting an edit and urged action. It missed that the date was not a firm deadline and that the editor was in a different time zone. In another instance, it helped plan around upcoming leave by suggesting weekly notes for an editor and drafting a Slack message. The message imitated the reviewer’s lowercase style but used an abbreviation they would not normally use; it was also labelled “sent with Claude.”
The reviewer found Duet’s cross-checking useful as a backstop for personal organizational systems, but did not want it to take over writing or review its own work. Giving the agent access to emails, drafts and messages also felt uncanny, despite being an informed choice. Anthropic says it is still testing Duet’s judgment about when to contact users. The reviewer’s broader concern is that AI can miss tone and nuance that a human assistant would recognize.
Comments (0)
No comments yet. Be the first!