Claude Duet Arrives: The AI Assistant That Speaks First
AI competition is shifting from answer accuracy to judging when to step in. Anthropic's Claude Duet is a new test of that idea.
AI no longer waits for your question
AI assistants have long shared one trait. However smart, they move only after a user asks or gives a command.
Claude Duet, which Anthropic unveiled on Oct. 21, tries to change that relationship.
Each morning it scans email and the calendar and sorts out what matters. When a connected app shows something that needs a check, it sends a message first. Tasks you hand off keep running on a cloud computer even when you step away.
The difference is clear. Earlier chatbots looked for answers to questions. Duet tries to judge which question to ask now and which job to handle first.
What changes in real work
Picture opening your phone in the morning to find a briefing like this.
Claude Duet morning briefing (hypothetical use case)
- Unpaid invoices: Found two overdue invoices in Stripe. A reminder draft for the customers is ready.
- Customer response: Saw a contract request in Slack. Wrote a reply using the customer history in HubSpot.
- Schedule conflict: Two meetings tomorrow overlap. Here are alternative times.
On the commute, the user answers by voice: "Send the invoice reminders, and soften the contract reply a bit."
Duet makes the edits and goes through any required approval steps. In the afternoon, it can also build a revenue forecast dashboard from Stripe payment data and the HubSpot sales pipeline.
This example is not a record of real use. It combines officially announced features. Still, it shows how interaction with AI could change: instead of a person finding and directing every task, the AI prepares first and the person focuses on the key decisions.
The real technical challenge: when to speak
Duet is built on Claude Opus 5.5. It uses the memory, skills and connected apps of an existing Claude account, and it can reach more than 6,000 connectors and plugins in the Claude Marketplace. Not every integration is supported equally well.
The number of connected apps is not what matters. What matters is the ability to decide which of the information found across 6,000 services to ignore and which to pass on to the user.
If it speaks too often, the AI becomes one more source of notification fatigue. If it misses something important, it loses trust.
Anthropic itself acknowledges that Duet can misjudge conversational context or urgency. Account checks also run at set intervals, not as real-time monitoring.
Clear limits on automation
- Prepared in advance: searching materials, drafting, reviewing schedules, building dashboards
- Needs user approval: sending messages, payments, signing documents, deleting data
- Not possible even with approval: stock trades, moving money between accounts, accessing other people's accounts
Duet is designed so the model never sees passwords or login tokens directly. It also has separate safeguards to prevent leaks of card numbers and government ID details.
In the current research preview, however, users cannot watch the AI's browser work in real time or take control of it. When handing off sensitive work, checking the results and approval requests remains important.
Free-looking, but not costless
During the research preview, Duet is included in Pro and Max plans at no extra charge. But the morning briefing, background checks and complex tasks all share Claude's existing five-hour session and weekly usage limits.
The longer Duet works on your behalf, the less room may be left for using Claude or Claude Code directly.
Too many approval requests also erode the benefit of automation. Time spent supervising the AI has to be less than the work time it saves for the gain to count as real productivity.
Insight Times Editorial Desk
