Anthropic’s new desktop and agent apps are changing how knowledge work gets done at the workplace. At the center is Cowork, Anthropic’s agentic desktop app that doesn’t just answer but actually works: reading, creating and modifying files, synthesizing research, preparing documents. For German and European enterprises, one question follows immediately: can these client apps be run in a GDPR-compliant way, without company data leaking uncontrolled?
The answer is yes, and the lever is more precise than most expect. Cowork and the Claude desktop apps can be pointed at data-sovereign, GDPR-compliant model endpoints instead of Anthropic’s consumer endpoint. Combined with CompanyGPT as the governed org platform, this creates a setup that unites individual productivity with organization-wide control.
What Is Claude Cowork?
Cowork is Anthropic’s agentic desktop app for knowledge work. Unlike classic chat, Cowork executes multi-step tasks on its own. You give Claude access to a specific folder on your machine and phrase a goal in plain language. Cowork plans the steps and works through them: transferring receipts from photos into an expense report, writing a summary from interview notes, cleaning up a chaotic directory.
Anthropic launched Cowork as a research preview in January 2026, first on macOS, soon after with feature parity on Windows. The app is available on all paid plans through the Claude desktop app. In practice, Cowork works like a digital coworker with access to your laptop: instead of pasting text into a chat window, you point Cowork at a folder and set a high-level goal.
This is exactly where the tension lies for enterprises. An agent that accesses local files and works in multiple steps is enormously productive. At the same time, it needs a clear answer to where inference goes and which data reaches which endpoint.
The Decisive Point: Endpoint, Not App
In an earlier article we explained that Claude via AWS Bedrock or Vertex AI gives you the model, but not the app you know from claude.ai. With Cowork, that picture shifts. Cowork is now the app, and it can be coupled to a sovereign endpoint.
Technically this runs through Claude Desktop’s third-party inference. In developer mode you configure a provider with region, model ID and authentication. This configuration is not set by every user individually but distributed centrally as a configuration file through device management: Microsoft Intune, Jamf or Group Policy. The file defines the model ID, inference profile, authentication method and organizational policies.
For EU enterprises, three sovereign paths emerge:
- AWS Bedrock in EU regions. Authentication via AWS IAM or a Bedrock API key. The inference region is determined by the inference profile. Bedrock offers in-Region, geo cross-Region and global profiles, so you can choose the right level of data residency. Importantly, Bedrock stores neither prompts nor files, tool inputs and outputs, or model responses, and does not use them for training.
- Google Vertex AI in Frankfurt. Claude runs as a model on Vertex AI in the EU. The region is determined by the endpoint, keeping data within the chosen geography.
- Your own innFactory API gateway. The client does not point directly at the hyperscaler but at an intermediary gateway. This enables central logging, policy enforcement, model routing and egress control in one place.
In all three cases, the setup inherits the existing governance tooling of the respective cloud: network isolation via VPC endpoints, audit via CloudTrail, granular cost attribution and optional telemetry. Company data stays in controlled EU infrastructure while employees keep the full Cowork experience.
How the Combination with CompanyGPT Works
Cowork solves a different problem than CompanyGPT. Confusing the two leaves value on the table. The strongest effect comes from understanding them as two layers.
Cowork is the power-user layer. Individual employees who work heavily with files gain noticeable speed from the agentic desktop. Cowork sits close to the operating system, works on local folders and handles tasks that would otherwise require manual clicking.
CompanyGPT is the governed org layer. The platform runs in the company’s own Azure tenant as a dedicated Kubernetes environment. It provides central model routing with token consumption per user and model via companyDASHBOARD, document creation via companyFILES, organization-wide knowledge access via companyRAG, and automation via MCP servers and n8n. What is the single workstation in Cowork is the auditable, organization-wide platform in CompanyGPT.
The following comparison makes the division of roles clear:
| Dimension | Claude Cowork | CompanyGPT |
|---|---|---|
| Primary benefit | Individual productivity, agentic file work | Governed, org-wide AI platform |
| Reach | Power users at the desktop | All employees, browser-based |
| Data handling | EU endpoint via MDM configuration | Own Azure tenant, data never leaves it |
| Logging and cost | Cloud-native (CloudTrail, telemetry) | companyDASHBOARD per user and model |
| Documents | Local files in the access folder | companyFILES creates Excel, Word, PowerPoint, PDF |
| Knowledge access | Local folder context | companyRAG with SharePoint integration |
| Model routing | One configured endpoint | Central routing across OpenAI, Gemini, Claude, Llama |
The combination is strong because both layers feed into the same compliance logic. Cowork points at a sovereign endpoint, CompanyGPT keeps data in the own tenant anyway. There is no second tool landscape with uncontrolled data flow, but one end-to-end governed setup.
A Concrete Mid-Market Scenario
A mechanical engineering company with 400 employees wants to roll out AI broadly but is subject to strict requirements from customer contracts and export control. Personal data and technical design documents must not leave the EU.
IT sets up two layers. For the roughly 30 power users in engineering, sales and controlling, Claude Cowork is rolled out via Intune. The configuration points at AWS Bedrock in an EU region, authenticated via IAM. A design engineer has Cowork transfer a bill of materials from several supplier PDFs into a consolidated Excel spreadsheet, locally, in the access folder, without any data going to a consumer endpoint.
For the remaining roughly 370 employees, CompanyGPT runs in the own Azure tenant. Sales creates quotes via companyFILES, controlling queries metrics from SharePoint via companyRAG, and management sees in companyDASHBOARD, to the day, which department consumes how many tokens. Both layers are documented in the company-wide AI policy under the EU AI Act.
The result: power users get the sharpest available tool, the rest of the organization a controlled platform, and data stays in EU infrastructure throughout.
Why Model Agnosticism Is Not a Luxury
A sovereign setup is not only a data privacy question but a matter of operational resilience. June 2026 made that clear. On 12 June 2026, Anthropic had to disable Claude Fable 5 and Mythos 5 for all customers worldwide on a directive from the US government, three days after launch. Anyone who had built their processes on one of these models was left without a solution overnight.
The lesson is not to avoid Claude. On the contrary: the strongest currently available Claude model is Claude Opus 4.8, followed by Sonnet 4.6 and Haiku 4.5. The lesson is not to depend on a single model. A sovereign gateway, through which both Cowork and CompanyGPT route, lets you switch to another available model within minutes if needed. CompanyGPT is model-agnostic by design and connects OpenAI, Google Gemini, Claude and Llama in parallel. In the current regulatory environment, this flexibility is not a comfort feature but risk mitigation.
Frequently Asked Questions
What is Claude Cowork?
Cowork is Anthropic’s agentic desktop app. You give Claude access to a specific folder and a goal in plain language, and Cowork autonomously executes multi-step knowledge work: reading, creating and modifying files, synthesizing research, preparing documents. Cowork is available on all paid plans through the Claude desktop app.
Can I run Claude Cowork in a GDPR-compliant way?
Yes. Cowork can be configured via device management (Intune, Jamf, Group Policy) so that inference does not run through the consumer endpoint but through a sovereign endpoint: AWS Bedrock in EU regions, Google Vertex AI in Frankfurt, or your own API gateway. The configuration centrally defines the model ID, inference profile, authentication and policies.
How do I connect Claude Cowork to AWS Bedrock or Vertex AI?
Through Claude Desktop’s third-party inference. In developer mode you configure a provider with region, model ID and authentication (AWS IAM or a Bedrock API key) and distribute it as a configuration file via MDM. Alternatively, traffic can be routed centrally through an LLM gateway.
Do I still need CompanyGPT then?
Cowork and CompanyGPT solve different problems. Cowork is individual productivity for power users. CompanyGPT is the governed, auditable org platform with central model routing, logging via companyDASHBOARD, RAG, document creation and automation. Together they form an end-to-end controlled setup.
Why does a model-agnostic setup matter?
Because models can disappear. On 12 June 2026, Anthropic had to disable Claude Fable 5 and Mythos 5 for all customers on a government directive. Anyone locked into a single model was left without a solution. A sovereign, model-agnostic gateway lets you switch to another available model such as Claude Opus 4.8 at any time.
Conclusion
Cowork and the Claude desktop apps are a real productivity leap, and they do not conflict with data privacy. The key is separating app from endpoint: the app runs at the workplace, inference runs through a sovereign EU endpoint. Combined with CompanyGPT as the governed platform, this creates a setup that unites individual speed with organization-wide control, without compromises on GDPR and the EU AI Act.
innFactory sets up both layers: the MDM configuration for Cowork on Bedrock, Vertex AI or your own gateway, and CompanyGPT in your own tenant. If you want to know what that looks like in your environment, talk to us.
Tobias Jonas is the founder of innFactory AI Consulting GmbH and innFactory GmbH, a Microsoft Cloud Solution Provider and Google Cloud Partner. He advises companies on the secure and sovereign adoption of AI. LinkedIn profile