Claude Duet’s real test is whether delegation can pay for itself

Claude Duet is not primarily a test of whether Anthropic can make Claude do more things. It is a test of whether people will pay every month to have an AI decide what deserves their attention. That is a much harder product problem.

Anthropic has opened Claude Duet as a research preview for Pro and Max users. Available through a standalone mobile app and Claude on the web, it monitors connected applications, sends a morning briefing, checks for developments during the day and works in a cloud-based virtual machine even after the user closes the app. The Verge describes the product as an early version aimed at professionals and small-business owners.

The commercial backdrop is clear enough. Pro is listed at $20 monthly or $200 annually, while Max starts at $100 monthly, according to VentureBeat. Duet does not appear to have a separately disclosed price. Its background checks also use the limits of the existing plan. That makes the subscription, rather than the app itself, the economic unit being tested.

The product sells attention, not automation

Most software promises to help users complete tasks faster. Duet makes a more ambitious promise: users should not have to notice the task first.

That distinction matters. Claude already offered connected apps, memory, skills and cloud-based work through related features. Duet’s differentiator is initiative. It decides that an email, calendar conflict or unpaid invoice may deserve an interruption, then brings the issue to the user. The appeal is not simply that it can draft a response or inspect a record. The appeal is that it might find the problem before the owner does.

Anthropic says Duet can work across more than 6,000 connectors and plugins, with examples including Gmail, Slack, QuickBooks, Stripe, Gusto, LinkedIn and Shopify. That breadth makes the product look less like a chatbot and more like a layer over business operations. In our view, that is the right direction for an agent business: the value of delegation rises when the work crosses several systems and would otherwise require repeated manual handoffs.

But attention is a scarce resource too. An agent that interrupts too often becomes another inbox. One that stays quiet may miss the reason the customer subscribed. Anthropic acknowledges that Duet can misread a thread or misjudge what is urgent. That is not a minor preview-stage imperfection; it is the central product risk.

The subscription only works if the economics stay invisible

An always-on agent has a different cost structure from a conventional chat assistant. It must periodically inspect connected accounts, interpret changes, maintain context and sometimes run tasks in a remote computer. The user may experience one short notification, while the service has performed recurring background work.

That creates a difficult pricing question. If a subscriber pays $20 monthly, how much autonomous monitoring can Anthropic provide before the plan becomes unattractive to operate? If usage limits are tightened, the agent may feel unreliable. If they are generous, the service may be expensive to run, particularly when Duet is built around Opus 5.5 rather than a smaller model.

The company has not presented a separate meter for “delegation,” which is probably the correct consumer design. People do not want to calculate the cost of every background check. Yet hidden consumption does not mean nonexistent consumption. The business still has to decide which users generate sustainable value and which users create costly activity without retaining their subscriptions.

That is why small businesses may be the more promising customer than casual consumers. A missed invoice, an overlooked customer message or a scheduling conflict can have a visible operational cost. If Duet prevents even a few meaningful failures, a recurring fee becomes easier to justify. For a personal user, by contrast, the value may be intermittent while the privacy and notification burden remain constant.

Approval gates reduce risk, but they also limit the magic

Anthropic has drawn a sensible line around consequential actions. Duet is designed to ask before sending a message, making a payment, signing a document or deleting a file. Securities trading and transfers between financial accounts remain prohibited even with approval. Users can disconnect an application, and Anthropic says the agent does not directly see passwords or login tokens.

Those controls are necessary for trust, but they expose the tension at the heart of the category. The more important the task, the more likely the user must return to review it. That makes the agent safer, but also makes it less autonomous. A system that prepares a payroll action, invoice reminder or agreement for approval may still save time, but it is not replacing judgment. It is reorganizing judgment into checkpoints.

Privacy adds another layer. Anthropic says a dedicated classifier is designed to prevent the sharing of highly sensitive information, including credit card numbers, while health and personal finance information require explicit instruction. Those safeguards sound meaningful, but the underlying bargain remains substantial: the more useful Duet becomes, the more connected data it must inspect. Users are being asked to trade a wider view of their digital lives for less administrative work.

Anthropic should sell reliability before autonomy

The launch is surrounded by competitive pressure from Meta’s Muse and OpenAI’s Dots, but copying the category is not the same as winning it. Duet’s strongest positioning is not that it can perform spectacular tasks. It is that it might become a dependable professional filter for routine decisions.

That suggests a better commercial strategy than maximizing apparent autonomy. Anthropic should prioritize controllable scopes, transparent activity histories, quiet hours, clear explanations for interruptions and easy permission revocation. The best agent may not be the one that acts most independently. It may be the one that makes its boundaries easiest to understand.

Our verdict is cautious but favorable: Claude Duet is a serious subscription experiment, not yet proof of a mass-market business. Professionals and small-business owners have a plausible reason to pay if the agent catches valuable work without creating a new stream of supervision. Everyone else should treat the research preview as exactly that—a test of whether delegation is genuinely useful, not a guarantee that an always-on assistant will earn its place in the monthly budget.