Claude Cowork and Chat Become One: What the Enterprise Interface Means for Your Document Work

Imagine asking, in the afternoon, for a report that would normally take your team several weeks: pulling figures together from dozens of documents, comparing them, putting them in context, and shaping them into a readable paper. You phrase the task, close your laptop and drive home. The next morning the finished report is ready, and you can open it on your phone. That is exactly the picture painted by an announcement from Anthropic dated October 23, 2026.

On that day, Anthropic merged chat and the agentic "Cowork" into a single interface for Claude Enterprise customers. For managing directors and decision-makers in mid-sized companies, this is a useful case study, because it makes visible a pattern that will show up in many tools over the coming years: a task is started, an agent keeps working in the background, and a finished document comes back at the end. This article describes what the source confirms, what it leaves open, what that means for a company in Germany, Austria or Switzerland, and how you can prepare before introducing something like this.

What happened on October 23, 2026

The primary source is a post by Anthropic titled "Claude Cowork and chat come together for Enterprise", published on October 23, 2026 in the Announcements category. We read it in full. All statements in this section are the vendor's claims about its own product; we have not verified them independently.

The core point in plain words: until now, users had to decide whether to chat with Claude or hand a task to the agentic working environment "Cowork". According to the vendor, both modes were already merged for Pro and Max plans in September. The same now applies to Claude Enterprise plans. According to the post, chat and Cowork are a single Claude, on web, desktop and mobile. Chats, projects, connections, so-called skills and plug-ins carry over.

• How it works: A conversation starts as a normal chat. If a task needs code or files, Claude keeps working in the cloud, according to the vendor, and returns the finished result.
• Keep working offline: According to the post, the work continues after the user has closed the laptop. It can be resumed on the web, on the desktop or on mobile.
• Rollout: Administrators control the rollout in the organization settings. According to the post, existing Enterprise customers receive an email about it.

At first glance this is a question of usability. What lies behind it matters more: the line between "I ask something" and "I hand off a task" disappears from the interface. Users no longer need to know in advance whether a task is small or large. The tool decides, and where needed it brings in an environment in which longer work is possible.

The customer example: weeks become ten minutes of prompting

The post contains a single customer quote, and it is the reason the topic is getting attention in the press and in videos. Paul Simpson, CFO and Head of Technology at LifeMoves, a non-profit organization in the US, is quoted as saying: "With under ten minutes of prompting, Claude worked through stacks of county, state and federal budgets." According to the post, he describes the resulting report as comprehensive and says it would have cost his team "weeks".

How should this statement be read? First, as what it is: the statement of a user, relayed by the vendor, who uses it in an announcement for its own product. It is not evidence of a general time saving. We do not know how long the task actually ran in the background, how much rework was needed, who reviewed the report in the end, or how many attempts came before. The figure "under ten minutes" also refers explicitly to phrasing the task, not to the total duration.

The example is still useful, as a mental model. It shows where the leverage lies: human effort shifts from compiling to describing and reviewing. You should not transfer the sample calculation to your company, but run it with your own figures. How many hours per week do your employees currently spend pulling together information from contracts, quotes, reports and spreadsheets? How many of those are tasks that can be described cleanly? The answer to that is the real gain from this news item, not the figure from the US.

The pattern is: describe the task, the agent keeps working, the finished document comes back. The human stays where judgment is needed: at the start with the assignment and at the end with the review.

Security and data: what the vendor describes

For a company, the question of security is at least as important as the question of time savings. The post describes several points on this, which we relay without having checked them technically:

• Isolated environment: The agentic work runs in cloud sessions, each in its own sealed-off virtual machine.
• Outbound network access: For Claude Enterprise it is switched off by default ("Network egress is off by default for Claude Enterprise").
• Credentials: The credentials of the connections are kept outside the sealed-off environment.
• Encryption and training: According to the vendor, data is encrypted in transit and at rest and is not used for model training by default.
• Existing controls: The post names single sign-on and a compliance interface as the foundation. Organizations with Zero Data Retention for Claude Code, a HIPAA-ready configuration or self-managed keys stay on their current configuration.
• Administration: Anthropic operates the environment and, according to the post, offers security teams central visibility and control.

That is a solid list, and the underlying idea is plausible: anyone who lets an agent work alone for a longer time should put it in a room it cannot leave at will. What matters, though, is what the post does not say.

The open question for DACH: data residency

The post names neither a data location nor any data residency options. Where the data from your documents is processed during a cloud session cannot be seen from the source. We searched the source specifically for this and found nothing. No statement can be derived from that, neither that processing takes place in the EU nor that it does not. The question is open, and for a company in Germany, Austria or Switzerland it is not a side issue.

The reason lies in what is in your documents. Contracts contain the names, prices and terms of business partners. HR records contain personal data. Quotes contain calculations you would not want to see in other hands. Whether and under what conditions such content may be processed with a service depends on the individual case and belongs in the hands of your data protection officer or legal counsel. We cannot and do not want to assess that here.

In practice this means: before you feed such a tool with real documents, clarify in writing with the vendor where processing takes place, which data processing agreements apply, and which data you will allow in a cloud session at all. The answer may in the end be that certain document types stay out.

Availability: generally available, but only for Enterprise plans

To put it in context, here is who can use this today at all:

• Status: Generally available for Claude Enterprise plans since October 23, 2026 (according to the vendor). This is not a trial phase and not a waitlist.
• Who: Only organizations with a Claude Enterprise contract. For Pro and Max, the merger was already completed in September, according to the post.
• Region: Not stated in the source.
• DACH: unknown. An Enterprise contract is in principle also conceivable for companies in German-speaking countries, but the data residency question remains unanswered.

That is why we classify the topic as "observe and prepare" and not as "immediately usable without restrictions". The feature is there. Whether it is suitable for your documents under your legal situation is something you have not yet clarified, as long as the answer to the data question is missing.

What it costs: usage-based, with controls for administrators

The post also comments on costs, and does so cautiously. According to the vendor, prices for Claude Enterprise remain usage-based. It names two figures: Claude Opus 5.5 costs 40 percent less to run than Opus 5 on typical workloads, and Claude Sonnet 5.5 up to 30 percent less than Sonnet 5. These are manufacturer figures for typical workloads; how they play out in your case depends on your usage.

Worth noting is the remark that agentic tasks can consume more tokens than a short chat. Anyone who lets an agent work in the background for hours pays more for it under usage-based billing than for a question-and-answer round. The vendor therefore names four levers for administrators:

• Efficient mode: keeps the work on the Sonnet and Haiku models.
• Model approvals: determine which models may be used for which role.
• Spending limits: can be set for the organization, for groups and for individual members.
• Shared group budgets: give a group a shared monthly amount.

For a mid-sized company this is a practical insight: an agent that is allowed to work for a long time needs an upper limit. That holds regardless of the product. Anyone introducing an agent decides in advance what it may cost at most, just as they decide what it may do at most.

What this means for a mid-sized company

Even if you are not a Claude Enterprise customer and perhaps never will be, the pattern is worth a look. We consider three conclusions robust, because they do not depend on this one product.

First: the task becomes an assignment. Those who work with agents phrase less "Write me a sentence" and more "Create this report from these documents for me". That requires a skill that many companies underestimate: describing an assignment so that it can be completed without follow-up questions. Which sources, which format, which questions should the result answer, what must not be in it? Anyone who can already write that down cleanly today has an advantage with any agent.

Second: the review moves to the end. When an agent works in the background for hours, you do not see what happens in between. You see the result. What matters then is who reviews it and against what. A finished report looks convincing even if it contains an error. Anyone who passes it on unchecked relies on a machine without knowing the path that led there.

Third: an assignment needs limits. An agent that can read, create and possibly change files needs clear rules. Which folders may it see? May it delete anything? May it send anything? The isolation in a separate environment, as the vendor describes it, is one technical part of the answer. The other part is organizational and rests with you.

Which agent at Amazing Agents would be responsible: the Document Brain

On amazing-agents.com, this kind of task belongs to the Document Brain. The website describes it like this: create quotes, review contracts, summarize reports. At its core, that is the same type of work the LifeMoves example describes: turning many documents into a readable result.

We are describing an approach here, not a result from a customer project. In an audit, we would define with you which document types the Document Brain may process, which it may not, in what format results come back, and who reads them before they are passed on. We do not name a concrete time saving as long as we do not know your processes.

Agent with limits and an approval step: the difference

The case shows well what distinguishes an agent without an approval step from an agent with one.

Without an approval step: An agent works in the background for hours, changes files in the process, stores results or possibly forwards them right away. Nobody has seen the finished report before it reaches its recipient. An error in a figure then sits in a paper that is already in circulation.

With an approval step: The agent works in a bounded environment and presents the result for review. A person reads it, compares the key figures with the sources, and either approves it or sends it back. Only after that does anything go to customers, banks or management. That costs time, and this time is the price for knowing what is passed on in your name. An agent with limits and an approval step works fast on the preparation and leaves the decision with you.

On amazing-agents.com, this framework applies to all agents: approval steps for critical actions, and emails are created as drafts, not sent directly. What applies there to emails can be transferred to documents: a report is a draft until someone approves it.

What we do not know

• Data residency: Where your documents are processed in a cloud session is not stated in the source.
• Quality of the results: We have not tested the product and cannot say anything about accuracy or error rate.
• Transferability of the example: The LifeMoves quote comes from a US organization and reflects a single experience.
• Total costs: The source names price changes for typical workloads, but no amounts for your usage.
• Legal assessment: Whether and how you may process contracts or HR data with such a service, we have not reviewed.

What you should prepare this week: three steps

Because region and data residency are open, the sensible path consists of preparation, not introduction. For this you need neither an Enterprise account nor prior knowledge.

1. Describe two tasks: Look in your business for two recurring tasks in which people pull together information from several documents, such as quote comparisons or contract overviews. Write down which sources belong to them, what the result should look like, and how many hours per week it costs today. That is your own hours calculation.
2. Sort documents by sensitivity: Divide your document types into three groups: uncritical, internal, confidential or containing personal data. Decide which groups an agent may see at all. Ask your data protection officer before you put anything into a cloud environment.
3. Define a reviewer and an upper limit: Determine in advance who reads an agent's result, what they check figures against, and what happens if they reject it. Also set a monthly budget that such an agent may consume at most.

Conclusion

For October 23, 2026, the source confirms: Anthropic is merging chat and Cowork into one interface for Claude Enterprise plans, agentic tasks keep running in the cloud where needed even when the user closes the laptop, and the feature is generally available for Enterprise plans. It does not confirm where the data is processed, and it does not confirm that the time saving from the US example transfers to your company.

For you this means: the pattern "start a task, the agent keeps working, the finished document comes back" is a usable target picture, regardless of which product ends up delivering it. Clarify the data question beforehand, decide who reviews the result, and set an upper limit for costs and access. An agent with limits and an approval step is a better foundation than handing off an assignment and only noticing at the result everything that happened in between.

Sources: Anthropic, "Claude Cowork and chat come together for Enterprise", claude.com/resources/articles, October 23, 2026 (retrieved October 24, 2026; statements on availability, security, prices and the LifeMoves quote are the vendor's claims). Assignment to the agent: amazing-agents.com, home page (Document Brain; approval steps for critical actions, emails as drafts).