Back to the journal

Actualités du marché

Claude merges chat and agent: decide before November 16

Anthropic is merging chat and Cowork for Claude Enterprise customers. Off by default for existing organizations, the new experience turns on by default on November 16, 2026. Roles, network, connectors, autonomy, budget and data: the decisions to make first, and how to turn it into real usage.

Author
Tomorrow Solutions
Published
October 26, 2026
Reading
6 min
Claude merges chat and agent: decide before November 16

On October 23, Anthropic announced a change for Claude Enterprise customers: chat and Cowork, its agent that works on files, are now one (PYMNTS, 10/25/2026).

For users, the change is simple. They ask their question. If the task requires producing a file or running code, Claude moves on its own into an isolated environment, does the work, then hands back the result for review.

For IT, it's a different story. This mode is off for existing organizations, and it turns on by default on November 16, 2026 (Claude Help Center). That leaves three weeks to decide who gets it, and how.

What really changes

Until now, you had to choose. On one side, chat, to ask questions and draft. On the other, Cowork, to delegate a full task: a spreadsheet, a case summary, a presentation.

Now every conversation starts as a chat. When the work calls for it, Claude opens a cloud session, on a virtual machine dedicated to that conversation and managed by Anthropic. Employees no longer switch modes: the tool decides when it becomes an agent.

Microsoft is taking the same path. Its new Copilot home page is presented as the place "where Chat and Cowork come together", and Cowork is billed on usage (Microsoft, 09/25/2026).

Two vendors, same direction. The line between "I ask a question" and "I hand over a job" disappears from the interface. So it has to live somewhere else: in your rules.

The default-setting trap

Anthropic's admin guide is clear on one point worth reading twice. Once the setting is on at the organization level, the standard roles (user, admin, owner) get access. To exclude someone, you have to move them into a custom role before it's enabled.

In other words, if nobody touches anything, on November 16 the agent reaches everyone who meets the prerequisites.

There are two prerequisites: chat enabled, and cloud code execution with file creation. If either is off, nothing turns on. That's a first lever, a blunt one.

The finer approach is to choose. Organization owners can enable it now, keep it off, or open it by role and in waves, using directory group sync.

Five decisions to make before November 16

1. Who gets access. A pilot group per business function beats a general rollout. Anthropic itself recommends piloting with one group, monitoring for two to four weeks, then expanding in waves.

2. What the agent can reach on the network. For Enterprise, outbound access from the environment is off by default. It can be opened to package managers only, to a list of domains, or to the whole internet. That choice belongs with security, not to chance.

3. Connectors. Email, storage, business tools: each connector can be blocked, require approval or be limited to certain roles. Anthropic enforces these controls on every call. The right reflex: start from the use case, and open only what it needs.

4. The level of autonomy. On Enterprise, the default mode asks the user to approve each action (Claude Help Center). Admins can disable the modes that skip these approvals, including for scheduled tasks. For a legal or finance department, it's the first setting to lock.

5. Budget. Agent work is billed on usage, with spend limits possible per organization, group or member. Without a limit, IT discovers consumption after the fact. With a limit set too low, the pilot stops halfway through a file. You need a number, and an owner.

What to check with legal

Three topics deserve a review by the legal department or the DPO.

Data retention. Cloud sessions expire 30 days after creation. Sessions started from a conversation are deleted with it. However, custom retention doesn't cover everything: Cowork sessions started directly and scheduled tasks, for example, follow their own rules.

Excluded configurations. The feature isn't available for organizations using zero data retention, the HIPAA-ready configuration or customer-managed encryption keys. If your group negotiated one of these options, the question doesn't arise yet. But you need to know before promising the tool to business teams.

Traceability. The compliance, analytics and admin APIs cover the new experience with existing keys. Good news for teams that already track usage. Watch your dashboards, though: usage appears on a separate line, "Chat and Cowork unified". If you add up Chat and Cowork without it, you underestimate real usage.

What I've seen in the field

This fall, in the legal department of a large group, the use case was ready before the tool. An annual campaign of corporate documents, existing templates, a willing lawyer who knew exactly where she was losing time.

What took weeks was another question. Who authorizes an agent that handles entire files? With what data? And who pays when the bill depends on usage? The answer wasn't written down anywhere. Meanwhile, enthusiasm faded.

The lesson is simple. When the tool makes the switch to agent invisible to users, the decision has to be made upstream, in writing, with a named owner. Otherwise it gets made by default, the day the setting changes.

From setting to real usage

On November 16, thousands of employees may have an agent one click away. Usage still has to be built.

Real usage means a precise business case, people trained on their own files, and a measure that goes beyond logins. Companies that succeed do both at once: they frame the tool, and they support the first teams all the way to results.

A good question for your leadership meeting this week: if the agent is open to everyone on November 16, who will be able to say, a month later, what it changed?

Prepare your decision

I help organizations frame this kind of rollout: choosing the pilot group, settings to validate with security and legal, business use cases, support for the first teams and real-usage indicators.

Want to talk before November 16? Get in touch.


Sources

Continue reading

Other perspectives on enterprise AI adoption

  1. 01

    Generative AI: Everyone's got their own!

    Generative AI, represented by tools like ChatGPT and Claude, specializes in various uses, with ChatGPT serving as a versatile advisor for daily tasks, while Claude acts as an automaton for executing business tasks. This evolution highlights the need for a strategic approach in choosing AI tools, while also emphasizing issues of privacy, reliability, and accountability. The future lies in learning to use these technologies in combination to optimize their effectiveness.

    October 14, 2025
  2. 02

    Will AI really destroy jobs? Yes...

    Artificial Intelligence will lead to the disappearance of certain jobs, particularly those related to repetitive tasks, while creating new roles focused on leveraging AI, such as "Prompt Engineers" and "AI Strategists." To adapt to this evolution, it is crucial to develop skills in prompt engineering, refine critical thinking, and value soft skills. By anticipating these changes, professionals can position themselves as "information architects" and take advantage of the opportunities offered by AI.

    May 27, 2025
  3. 03

    AI in French SMEs

    The article emphasizes that Artificial Intelligence (AI) has become essential for French SMEs, with an expected adoption rate of 55% by the end of 2025, enabling productivity gains and improved customer relations. Despite obstacles such as cost and lack of knowledge, accessible tools and government initiatives, such as subsidies and training, are facilitating this transition. SMEs that adopt AI are seeing tangible benefits, and those that commit now are positioning themselves favorably for the future.

    May 6, 2025

Turn this analysis into a decision.

A 30-minute diagnostic identifies the main friction point in your AI program and the next useful checkpoint.

Book a diagnostic