Claude for Business: The Complete Guide 2026

• Chat, Cowork or Code: If the result is an answer, use Chat. If it is a file or a completed task, use Cowork. If it is code in a repository, use Claude Code.
• Context window: The context window is Claude's working memory for a request. The current models hold 1 million tokens, roughly 555,000 words, but accuracy declines as length grows.
• Licences: Free, Pro and Max run under consumer terms without a data processing agreement and are unsuitable for company data. Claude Team is the first plan for businesses: no training on your data, DPA included, central administration.
• EU operation: No plan on claude.ai guarantees processing in the EU. Claude runs exclusively in the EU via Google Vertex AI (multi-region "eu") or AWS Bedrock ("eu." profile called from an EU region).
• New in October 2026: Claude Design is available on all plans, Claude Motion and Dashboards launch in beta, and Chat and Cowork are being merged into a single Claude.

In 2026, Claude is no longer a single tool but a family: Chat, Cowork and Claude Code, plus Design, Motion, Dashboards, Claude in Chrome and the add-ins for Microsoft 365. This guide puts all of this into context for businesses: which tool suits which task, how the context window works, which licence is compatible with the GDPR and how Claude runs exclusively in the EU.

The guide was written by Thomas Knorr, Managing Director of datadice GmbH in Coburg. He uses Claude daily in client projects and teaches the use of AI in certificate courses at the IHK zu Coburg (Chamber of Commerce and Industry) and Coburg University of Applied Sciences.

We deliberately leave out installation instructions. Most people in your organisation work in the app, and anyone setting up Claude Code in the terminal will find the steps in the official documentation.

As of: 13 October 2026. Prices, features and regions at Anthropic change within weeks. All information is taken from the websites of Anthropic, Google, AWS, Microsoft and GitHub, retrieved on that date.

Chat, Cowork and Claude Code differ in who carries out the work steps and what comes out at the end: in Chat you direct every step, in Cowork Claude plans the steps itself and delivers a file, and in Claude Code Claude works inside a software project.

Chat: you lead, Claude answers

You ask a question, Claude answers, and you decide on the next step. Suitable for drafts, summaries, translations, thinking an idea through or analysing an uploaded file. Every step goes through you, so you keep full control and also carry the entire work of steering.
Typical: drafting a reply to a customer, summarising a contract, having a spreadsheet explained.

Cowork: you hand over a task

You describe a result, Claude plans the steps itself and delivers a file: a report, an Excel analysis, a presentation, a tidied-up folder. Cowork keeps working in the cloud even when your laptop is closed, accesses folders you share via Claude Desktop and takes on scheduled tasks, for example every Monday morning.
Typical: "Evaluate the 40 quotes in this folder and create a comparison table with price, delivery time and deviations from the requirements specification."

Claude Code: for everything that is software

Claude Code reads a repository, changes files, runs commands and tests and prepares pull requests. It runs in the terminal, in VS Code and JetBrains, in the desktop app, in the browser at claude.ai/code, in Slack and in CI pipelines such as GitHub Actions.
Typical: building a feature, narrowing down a bug, migrating a database, developing an internal application further.

The rule of thumb

• The result is an answer: Chat.
• The result is a file or a completed task: Cowork.
• The result is code in a repository: Claude Code.

What is changing right now

Anthropic is merging Chat and Cowork into a single Claude, with the rollout starting on Pro and Max. The toggle disappears, and Claude decides itself whether a request needs an agent. The distinction remains useful nonetheless: it determines how you phrase a task, which folders and systems you share and how you check the result.

One difference also remains: only Claude Code can be operated in the EU via Google Vertex AI or AWS Bedrock. Chat, Cowork, Design and Motion are available exclusively on claude.ai.

The context window is the amount of text Claude takes into account at once for a single request, measured in tokens. With this concept you understand why Claude works precisely in one conversation and forgets things in the next.

What the context window is

Everything Claude has in front of it at a given moment is in the context window. It is not the model's training knowledge, and it is not the memory feature. It is the working memory for exactly this one request. Whatever is not in it does not exist for Claude at that moment.

Tokens: the unit in which Claude reads

Claude does not read words but tokens, that is, parts of words. According to Anthropic, 1 million tokens correspond to roughly 555,000 words or 2.5 million characters for the current models. The new tokenizer that Anthropic introduced with Opus 4.7 produces around 30 per cent more tokens for the same text than the old one. German texts tend to need more tokens than English ones because of their long compound words; Anthropic does not publish an official figure for this.

What counts towards the window

• the system instructions, including those you do not see
• your project instructions, skills and the description of every active connector
• the entire conversation history so far, including uploaded files and images
• the results of every tool: every web search, every file read, every database query
• the response itself, including the reasoning steps

With every new message, Claude sends the complete history to the model again. A conversation therefore becomes heavier with every round, not just longer.

How large the window is

The current models Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5 have a window of 1 million tokens via the API. The same applies to the new models in the chat app; older models such as Opus 4.6 to 4.8 have 500,000 tokens there, and even older ones 200,000. In Cowork the window is partly smaller, depending on the model.

More is not automatically better

Anthropic says so itself: as length grows, accuracy and recall decline. The technical term for this is context rot. A model has a limited attention budget; the more there is in the window, the more thinly it is spread. Research also describes the "Lost in the Middle" effect (Liu et al., 2023): language models make worse use of information in the middle of very long texts than of information at the beginning or end.

What happens when the window is full

• In Chat, on paid plans Claude summarises older messages to make room. This requires code execution to be enabled. The summary loses details that may have mattered to you.
• In Claude Code, automatic compaction (auto-compaction) kicks in at around 967,000 tokens for models with 1 million tokens. With /compact and a focus ("keep the database decision") you compact selectively; with /clear you start afresh.

Projects and RAG: knowledge that does not sit permanently in the window

In a project you store instructions and knowledge files. As long as the knowledge is small, Claude loads it completely into the window. If it becomes too large, Claude automatically switches to RAG (retrieval augmented generation) on paid plans: a search tool fetches only the relevant sections. According to Anthropic, this increases capacity up to tenfold. The price: Claude no longer sees everything at once, only what the search finds.

Persistent context without ballast

• Project instructions apply to every chat in the project.
• CLAUDE.md and AGENTS.md are the project memory in Claude Code: architecture, commands, rules. Claude reads them at the start of every session.
• Skills are folders with instructions and scripts that Claude loads only when a task needs them. They therefore only take up space in the window at that point.
• Memory retains context across conversations. On Free, Pro and Max the feature is on by default, on Team and Enterprise it is off.

What the context window costs

Via the API you pay per token, and because every round rereads the entire history, long sessions cost disproportionately more. Prompt caching softens this: content that has already been read costs only 5 per cent of the normal input price when reread with Opus 5.5 and Sonnet 5.5, and 10 per cent with most other models. Since the 4.6 generation of models, a token at the end of a 1-million window costs the same as one at the beginning. Claude Code caches automatically.

Five rules for practice

• New task, new chat. A conversation about three topics gets worse for all three.
• Give Claude the files it needs, not the whole folder.
• Whatever should always apply belongs in the project instructions or in CLAUDE.md, not in the tenth message.
• Only activate the connectors you need for the task. Each one takes up space.
• Have long work sessions deliberately summarised and start afresh with the summary.

For businesses, Claude Team is the first suitable plan, because only from Team upwards do the commercial terms apply: no training on your data, a data processing agreement and central administration. The decisive line does not run between cheap and expensive, but between consumer terms and commercial terms. All prices in US dollars plus tax, as of 13 October 2026; Anthropic shows the euro price depending on your billing address.

Free ($0): ruled out for company data

The consumer terms apply. There is no data processing agreement (DPA), no administration and no control over who enters what. Since the change to the consumer terms on 28 August 2025, Anthropic uses chats to train new models if the person consents. According to reports by TechCrunch and Tom's Guide, the toggle for this was preset to "On" in the notice window. If you consent, Anthropic stores your chats for up to five years; without consent, for 30 days. Personal data of customers or employees has no place here.

Pro ($20 per month, $17 annually) and Max ($100 or $200): more capability, same legal position

Pro and Max bring Claude Code, Cowork, projects, Claude in Chrome and the Microsoft 365 add-ins. Legally, nothing changes compared with Free: consumer terms, no DPA, storage in the US, training dependent on a toggle per person. The company cannot verify how each individual has set this toggle. This explicitly also applies to Claude Code when it runs via a Pro or Max account. A private subscription on a company credit card is therefore not a solution for company data, but shadow AI with an invoice.

Team (from $25 per person per month, $20 annually): the entry point for businesses

Team is the first plan we recommend to businesses. The commercial terms apply here:
- Anthropic does not train on your data.
- The DPA (Data Processing Addendum) is automatically part of the terms; your company is the controller, Anthropic the processor.
- An admin console manages users, roles and settings centrally, with single sign-on and domain capture.
- Skills and connectors can be approved centrally, and memory is off by default.
- Team Standard costs $25 ($20 annually), Team Premium with five times the usage $125 ($100 annually). Between 2 and 150 people are possible.

What Team cannot do: the data is stored in the US, processing runs worldwide by default, and there is no automatic user provisioning via SCIM and no audit logs.

Enterprise (from $20 per person per month annually, plus usage at API prices)

Enterprise adds what larger organisations and auditors require: SCIM, audit logs, a compliance API, custom retention periods, customer-managed keys and, for Claude Code, the option of zero data retention after review by Anthropic. Processing can be restricted to the US. The entry point is 20 people via self-service and 50 via sales.

The point no plan solves

None of the plans on claude.ai guarantees that your data is stored or processed in the EU. Anthropic stores data in the US and by default processes it in data centres in the US, Europe, Asia and Australia. For data that must not leave the EU, you need the route via a cloud provider.

Claude runs exclusively in the EU only via a cloud provider: via Google Vertex AI with the multi-region endpoint "eu" or via AWS Bedrock with the inference profile "eu.", called from a region in an EU member state. Anthropic itself offers only "global" or "us" for processing and only "us" for storage. There is no EU option directly from Anthropic, neither for the API nor for Enterprise. EU operation is delivered by the major cloud providers.

Google Cloud Vertex AI

Vertex AI offers Claude via three types of endpoint. The global endpoint distributes requests worldwide and gives no guarantee about location. According to Google, the multi-region endpoint "eu" keeps processing strictly within EU member states; the United Kingdom and Switzerland are excluded. Regional endpoints such as europe-west1 exist only for Sonnet 4.6 and older models. Opus 5.5 is available in the "eu" multi-region; for other models, check availability in the Model Garden. Multi-region and regional endpoints carry a 10 per cent surcharge.

AWS Bedrock

Bedrock works with geographic inference profiles. The eu. profile for Opus 5.5 distributes requests coming from an EU region such as Frankfurt only to Frankfurt, Stockholm, Milan, Spain, Ireland and Paris. Pay attention to the source region: if you call the same profile from Zurich or London, Switzerland or the United Kingdom respectively are added as destinations. With Claude in Bedrock, Anthropic employees have no access to the inference infrastructure.

Microsoft Foundry

Claude is available in Microsoft Foundry, but only globally or in the US data zone. There is no EU data zone for Claude; Anthropic lists it as "Coming soon", without a date.

Other routes

• GitHub Copilot has offered EU data residency since April 2026, including for Claude models. The policy is off by default and costs 10 per cent more.
• Snowflake Cortex processes several Claude models via the AWS_EU region, though not every current model.
• Databricks offers Claude in European regions; its "Europe Geo", however, also includes the United Kingdom and Switzerland.
• IONOS and STACKIT do not offer Claude. If you want a European cloud without a US provider, you work with open models there.

What you give up with EU operation

Via Vertex AI and Bedrock you get the model, not the app. Chat, Cowork, Design and Motion are not available there. You need your own interface, your own application or Claude Code. Billing is per token via your cloud invoice.

Claude Code with an EU backend

Claude Code officially supports Vertex AI, Bedrock and Foundry. With CLAUDE_CODE_USE_VERTEX=1 and CLOUD_ML_REGION=eu, processing runs via Google's EU multi-region. Telemetry and error reporting are off by default with these providers. Two connections to Anthropic remain nonetheless: before every web page fetch, Claude Code sends the hostname to Anthropic for a security check, which can be switched off with the skipWebFetchPreflight setting. You switch off the short rating surveys with CLAUDE_CODE_DISABLE_FEEDBACK_SURVEY=1.

The residual risk: the US CLOUD Act

An EU region determines where your data is processed. It does not change the fact that AWS, Google, Microsoft and Anthropic are US companies that US authorities can compel to hand over data. In 2019, the European Data Protection Board stated that such a transfer under the GDPR requires an international agreement. The General Court of the EU upheld the EU-US Data Privacy Framework on 3 September 2025; an appeal before the Court of Justice of the European Union is pending. Data protection experts regard the CLOUD Act as a residual risk that you address with a transfer impact assessment, a DPA and standard contractual clauses, as well as technical measures such as customer-managed keys. This section does not replace legal advice.

Our recommendation in three tiers

• Everyday work without personal data: Claude Team with a clear policy on which data may be entered.
• Applications with customer, HR or business data: Claude via Vertex AI or Bedrock in the EU, integrated into your own application with permissions and logging.
• Highly sensitive data: open models in your own data centre or in a European cloud, for tasks where their quality is sufficient.

What this looks like in an application with permissions is shown on our pages on AI Implementation and Knowledge Management.

Claude Design

Generally available on all plans since 8 October 2026. Claude Design creates prototypes, slides, one-pagers and mockups in dialogue, derives a design system from existing code and design files and exports to Canva, PDF, PowerPoint and HTML or hands over directly to Claude Code. The separate site claude.ai/design closes on 14 December 2026; the feature lives on in the app.

Claude Motion

In beta since 8 October 2026, only for Team and Enterprise; on Enterprise, an admin enables it. Motion creates explanatory animations from your own material, for example from reports, charts or product workflows, and exports them as MP4. The animation is code, so text, figures and timing remain editable. According to Anthropic, Motion does not use a video generation model; Claude therefore does not produce realistic videos or avatars.

Dashboards

Also in beta since 8 October 2026, on all paid plans: Claude builds live dashboards on BigQuery, Databricks, Snowflake or Salesforce, writes the queries and keeps them up to date. The quality depends on the data foundation: a dashboard built on unclarified metrics shows unclarified figures faster.

Claude in Chrome

Generally available since 26 August 2026. Claude reads and operates web pages in your browser, in Cowork and in Claude Code. Only give the agent access to the pages and accounts the task needs: web pages can contain instructions aimed at the AI (prompt injection).

Claude for Microsoft 365

Add-ins for Excel, PowerPoint and Word are available on all paid plans, Outlook in beta. Claude reads from one open file and writes into another.

Skills and connectors

Skills are an open standard: folders with instructions, scripts and templates that Claude loads when needed. Team and Enterprise distribute them centrally, for example the quote template or the rules for customer communication. Connectors based on the Model Context Protocol (MCP) connect Google Workspace, Microsoft 365, Slack, GitHub or your own systems. Every connector is a gateway to your data and therefore needs approval by IT.

• Claude Fable 5.1 for demanding reasoning and long agentic tasks. $10 per million input tokens, $50 per million output tokens.
• Claude Opus 5.5 is Anthropic's default recommendation for most tasks. $4 and $20.
• Claude Sonnet 5.5 combines speed and quality. $2 and $10.
• Claude Haiku 5.5 for high-volume tasks such as classification and extraction. From $0.10 and $0.50.

All four have a context window of 1 million tokens, up to 128,000 output tokens and a knowledge cutoff of June 2026. Batch processing costs half. If you use Claude via Vertex AI or Bedrock in an EU region, you pay a 10 per cent surcharge.

Maintaining the project memory

A good CLAUDE.md describes in a few paragraphs what the project does, how it is built and tested and which rules apply. It is the difference between an agent that guesses afresh every time and one that knows the team's conventions.

Plan first, then build

In plan mode, Claude Code describes what it intends to change before it touches any files. For anything beyond a small fix, this step pays off: mistakes in the plan cost minutes, mistakes in finished code cost days.

Set permissions deliberately

Claude Code asks before commands and changes unless you configure it otherwise. For teams, the configuration should specify what may run without asking and what never may: tests yes, production deployments no.

Specification, tests, review

AI-generated code without tests and without review is technical debt. Claude Code writes tests, runs them and checks its own changes; responsibility for the result remains with the people who approve the pull request.

Keeping an eye on costs

As a guideline for enterprise teams, Anthropic cites an average of around $13 per developer per active day, roughly $150 to $250 per month. Long sessions cost more because every request rereads the entire history.

• Rule out private accounts for company data. Free, Pro and Max run under consumer terms.
• Define data classes. What may go into Claude Team, what only into an EU application, and what stays in-house?
• Choose the right plan. Team as the standard, Enterprise if you need SCIM, audit logs and custom retention periods, Vertex AI or Bedrock for EU operation.
• Approve connectors and skills centrally. Every connector is a gateway to your systems.
• Involve the works council early. Tools that log work results touch on co-determination rights.
• Train your people. Since February 2025, Article 4 of the EU AI Act has required companies to ensure sufficient AI literacy among their staff. The datadice Academy trains teams on their own tasks.
• Start with a real task. A recurring task with a clear result shows faster than any strategy where Claude delivers value in your organisation.

Which task is suitable for this in your organisation is what we work out in the strategy workshop: one day on site with you, ending with a plan for the first 90 days.

All sources retrieved on 13 October 2026.
- Anthropic, prices and plans: claude.com/pricing
- Anthropic, change to the consumer terms of 28 August 2025: anthropic.com/news/updates-to-our-consumer-terms
- Anthropic Privacy Center, training, retention, DPA and server locations: privacy.claude.com
- Anthropic, Regional Compliance: claude.com/regional-compliance
- Anthropic, API data residency: platform.claude.com/docs/en/manage-claude/data-residency
- Anthropic, models, prices and context windows: platform.claude.com/docs/en/about-claude/models/overview
- Anthropic, Effective context engineering for AI agents (29 September 2025): anthropic.com/engineering
- Anthropic, Dashboards and Motion (8 October 2026): claude.com/resources/articles/dashboards-and-motion
- Claude Code, third-party providers, Vertex AI and data usage: code.claude.com/docs
- Google Cloud, data residency in Vertex AI: docs.cloud.google.com/vertex-ai/generative-ai/docs/learn/data-residency
- AWS, model card Claude Opus 5.5 and geographic inference: docs.aws.amazon.com/bedrock
- Microsoft Learn, Claude models in Foundry: learn.microsoft.com/azure/foundry
- GitHub Changelog, Copilot data residency (13 April 2026): github.blog/changelog
- EDPB and EDPS, statement on the US CLOUD Act (2019): edpb.europa.eu
- General Court of the EU, judgment T-553/23 on the Data Privacy Framework (3 September 2025): curia.europa.eu
- Liu et al., Lost in the Middle (2023): arxiv.org/abs/2307.03172