Anthropic merges chat and Cowork in Enterprise: the agent stops being a separate product and becomes a mode of conversation

By PYMNTS · October 25, 2026.

Anthropic has unified Claude chat and Cowork for its Enterprise customers, the company announced on Friday, October 23. The combined offering works on web, desktop and mobile, and users no longer have to choose one service or the other before getting started.

The company describes how it works on its support page: “every conversation starts as chat and, when a task needs to run code or work with files, Claude continues the same conversation in an isolated cloud sandbox managed by Anthropic, and returns the finished work for review.” A sandbox is a closed environment where the agent can take actions without directly touching the user’s systems. When a task does not require agentic work, the session goes back to being a simple chat.

The rollout is gradual. According to the announcement, the feature appears in admin settings as “Chat and Cowork unified.” It is off by default for existing organizations and will be switched on by default on November 16. That gives administrators about three weeks to decide and prepare policies before the change reaches their teams.

The PYMNTS article provides no usage figures, pricing or further technical details about the sandbox. It is best not to read more into the change than what Anthropic has said.

**Our take.** The thesis is simple: the agent stops being a product and becomes a mode of the conversation. As long as users have to decide up front “is this a question or a task?”, they bear friction that shouldn’t be theirs. By merging the two, Anthropic hands that decision to the model, which is best placed to know whether it needs to run code or handle a file. It is the same direction we see across the industry: the interface gets simpler and the complexity is hidden behind it.

That shift comes with a governance cost. If any chat can turn into a session that runs code and handles files, the boundaries of what the AI “does” are no longer obvious to the user. That is why it matters that activation is opt-in at first and sits in the admin console. In corporate environments, the useful question is not what the agent can do, but what it is allowed to do, with what permissions and with what traceability.

In the same piece, PYMNTS links the news to its own report on incidents involving agents. According to its study, 57.1% of companies with at least $1 billion in annual revenue reported incidents or losses involving adversarial bots and agents, compared with 32.4% of those with $50 million to $250 million. These are PYMNTS figures on fraud and identity verification, not on Claude, and should not be read as a problem with Anthropic’s product. But they do illustrate the context: the larger the company and the more transactions it handles, the greater its exposure to malicious agents, and the greater the need to verify who, or what, is acting and with what authority.

In general, running tasks in a provider-managed sandbox is a reasonable response to near-term risk: it isolates the agent’s work and returns the result for human review. That final review step is, in our view, the key piece. As long as a human keeps checking what is delivered, automation boosts productivity without giving up control. The danger comes when review becomes routine and nobody really looks.

In the long run, this convergence points in the right direction: assistants that don’t force users to learn “modes” and that absorb more and more mechanical work, leaving judgment and decisions to people. The transition, however, will be uneven. Companies with good governance will benefit quickly. Those that switch on the feature on November 16 without reviewing permissions will find its limits the hard way.