Claude for Business: Plans, Data Protection and Where It Pays Off

Claude is Anthropic's AI assistant. A company meets it in four forms: a chat workspace for teams, an enterprise platform with extra controls, an API for custom applications, and a model offered by cloud providers. This guide sorts out the options, lists the data-protection settings that matter and shows where deployment tends to pay off. Status: October 2026.

What is Claude, and what do companies use it for?

Claude is a family of language models delivered through a chat interface, a coding tool and an API. Companies use it to write, analyse, summarise and program, and as the engine inside agents that work through tasks on their own with access to tools.

In practice there are three ways of using it:

• Assistance at the desk: staff use Claude in the browser or desktop app for drafting, research, spreadsheet analysis and summaries.
• Software development: Claude Code supports programming. According to Anthropic, it is included in the Team plans.
• Your own processes and agents: through the API, Claude is wired into ticketing, ERP reporting or document workflows. This is where the value sits that a chat licence alone cannot deliver.

The models come in tiers. Anthropic's pricing page currently lists Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5. Larger models cost more and suit hard tasks; smaller ones are faster and cheaper. Which tier fits which job is something you find by testing on your own data, not by reading vendor descriptions.

Which plans exist, and how do they differ?

Organisations can choose between Team, Enterprise and the API. Team bundles administration and collaboration for small and mid-sized groups. Enterprise adds security and governance features. The API bills by usage and serves custom applications. Free, Pro and Max are individual plans.

The table summarises the pricing page and the Help Center article on Team. Prices are US list prices excluding tax and may change.

For Team, Anthropic lists standard seats at 20 US dollars per seat and month on annual billing, or 25 US dollars billed monthly. Premium seats with a higher usage allowance cost 100 or 125 US dollars. Standard and premium seats can be mixed.

The pricing page lists Enterprise at 20 US dollars per seat and month on annual billing, plus usage billed at API rates. That points to the key difference: with Team, a usage allowance is part of the seat; with Enterprise, consumption is billed separately. For budgeting, Enterprise costs therefore track actual usage more closely.

What does Team offer, and when is it enough?

Team is enough as long as you need central administration, SSO and shared projects, but not SCIM or audit logs. According to Anthropic it includes role and spend controls, single sign-on with domain verification, connectors and Claude Code.

The Help Center also mentions spend limits at organisation and user level, connectors for Google Drive, Gmail, GitHub, Microsoft 365 and Slack, and monthly API credits. The pricing page adds shared projects, memory, skills and usage analytics.

Team fits well when:

• a department or pilot group of two to a few dozen people starts out,
• identity management through SSO is sufficient,
• there is no requirement for tamper-proof logs,
• you want experience before fixing contracts and processes.

A common mistake is letting Team seats coexist with private accounts. If staff copy company data into personal Free or Pro accounts, those accounts follow different data rules than Team. A clear instruction on which accounts are allowed matters more than any single setting.

When is Enterprise worth it?

Enterprise is worth it once evidence and control requirements outgrow Team. According to Anthropic, it adds role-based access, SCIM, audit logs, a Compliance API and custom data retention on top of what Team offers.

The Help Center article on Enterprise also names an Analytics API, customer-managed encryption keys, US-only inference as an option and a HIPAA-ready configuration. The last one concerns US health data and is irrelevant for most European firms.

For mid-sized companies three features usually decide:

1. SCIM: users are created and disabled automatically from your directory, which prevents orphaned accounts after people leave.
2. Audit logs: you can see who did what and when, which helps with internal investigations and with evidence for auditors.
3. Retention settings: you decide how long conversations are kept.

Enterprise comes in two routes. The pricing page describes a self-serve version and a sales-assisted version with invoicing and payment terms. Which suits you depends on your procurement process, not on technology.

When is the API the right route?

The API is right when Claude becomes part of a process or product rather than a conversation partner in a browser. You pay per token, control the model and requests yourself, and connect Claude to your own tools and data sources.

According to the pricing page, list prices per million tokens currently read:

For Haiku 5.5 the page lists two price levels depending on context length. The API documentation describes a batch interface for asynchronous processing with a 50% lower cost. That suits work that need not finish immediately, such as classifying a large document archive overnight.

Two points matter for cost planning. Token costs arise not only from your question but from all the context you send along and from every tool round of an agent. An agent with ten steps consumes many times what a single question does. So plan measurements with real data before fixing budgets. For agent projects, the overview of AI agents by department helps to think about process and cost model together.

How are privacy and data handling covered?

For Team, Enterprise and the API, Anthropic states that inputs and outputs are not used for model training by default. Exceptions are feedback and explicit opt-ins. This does not replace your own assessment, but it is an important precondition for business use.

The Privacy Center spells out the details:

• Commercial products do not use content for training by default.
• Submitting feedback, for example a thumbs up or down, releases the related conversation for use. Anthropic says it is stored for up to five years.
• Owners of Team and Enterprise organisations can switch feedback off in the settings.

On retention, the Privacy Center says that for the API, inputs and outputs are deleted from backend systems within 30 days by default, unless an exception or a separate agreement applies. For flagged policy violations, retention can be longer; Anthropic names up to two years for inputs and outputs.

Contractually, the Commercial Terms state that Anthropic may not train models on customer content from the services and that a data processing addendum forms part of the terms. This is orientation, not legal advice. Whether a particular processing activity is lawful depends on your data, your legal basis and any transfers to third countries. Involve your data protection officer in the decision.

A simple classification of data helps:

1. Uncritical: public texts, product descriptions, general research.
2. Internal: quotations, concepts, internal minutes.
3. Personal: customer data, job applications, employee data.
4. Specially protected: health data, trade secrets, data under confidentiality duties.

For each class, define whether Claude is allowed, under which settings and with which account. For the most sensitive classes, running models locally can be the better choice; see sovereign AI.

Where else is Claude available?

Besides the direct API, cloud platforms offer Claude. According to the API documentation these are Amazon Bedrock, Google's platform for AI models, Microsoft Foundry and Claude Platform on AWS. Which features are available differs by platform.

The main reason for this route is existing cloud contracts. If framework agreements, identity management and billing already sit with a hyperscaler, you can use Claude there without onboarding another vendor. The documentation points out that feature availability varies by platform and that new features can arrive with a delay.

Check before deciding:

• whether the features you need (tool use, batch, caching) exist on the chosen platform,
• in which region data is processed,
• who is the processor under contract, Anthropic or the cloud provider,
• how usage limits and costs behave there.

Which use cases pay off first?

Claude pays off fastest where text or structured information is the core of the work and a person reviews the result. Typical candidates are quotation texts, service documentation, research, summaries and programming help.

Proven starting points:

• Sales and quotations: drafts from enquiries, specifications and product data. More in the page on quotation agents.
• Technical service: combining fault descriptions, manuals and ticket history and preparing reply drafts; see technical service.
• Knowledge management: answering questions on internal documents with source references; see knowledge management.
• Software development: explaining code, writing tests, proposing changes.
• Administration: summarising and comparing minutes, contracts and reports.

Less suitable are tasks where errors have serious consequences and no human checks the output, such as binding legal statements, approvals with financial effect or decisions about people. There Claude delivers drafts, not decisions. For exact arithmetic, a tool should calculate, not the language model.

How do you roll Claude out?

A rollout works in small steps: define data classes, start a pilot with a clear goal, measure, document rules and only then extend. Starting with licences and clarifying rules later produces shadow usage.

A workable sequence:

1. Choose a goal and a process. One concrete process, such as pre-sorting service tickets, beats the vague aim of trying out AI.
2. Settle data classes and accounts. Decide which data may go into which account.
3. Pick the plan. A pilot usually starts on Team; custom applications start on the API.
4. Name a pilot group. A small group that reports regularly on what works and what does not.
5. Measure. Record time spent before and after, error rates and satisfaction.
6. Write the rules. A short usage policy on who may use Claude for what.
7. Decide. Adjust the plan, expand or stop.

If you cannot staff responsibility for selection, rules and operations internally, someone has to take that role. One way to do so is described on the page about a fractional head of AI.

Where are the limits and risks?

Claude makes mistakes, occasionally invents details and knows your company only as far as you give it information. Add vendor dependency, changing prices and data-protection questions that only you can answer.

The main risks at a glance:

• Wrong statements: results must be checked for important decisions. Without source references, checking is laborious.
• Shadow usage: without clear rules staff use private accounts. Only a usable official offer prevents that.
• Cost surprises: with agents and long contexts, consumption rises quickly. Spend limits and monitoring belong in from day one.
• Vendor lock-in: prompts, tools and workflows depend on the model. An exchangeable layer in between keeps options open.
• Legal questions: legal basis, transfers to third countries, information duties and, in Germany, possibly works council co-determination remain your responsibility.
• Prices change: all prices quoted are as of October 2026 and may be adjusted at any time. Check the pricing page before every budget decision.

How do you get started?

Start with a single process, a small team and one clear data rule. Choose Team or the API first, measure over four to six weeks and then use the numbers to decide whether Enterprise or a wider rollout is justified.

Three questions prepare the ground. Which process costs the most time today and consists mostly of text? Which data does it involve, and who may see it? Who reviews the results? Once these three answers exist, choosing a plan is rarely a hard decision.

The following articles in this series cover setup, privacy settings, comparisons with other vendors and the use of Claude Code in detail, once they are published. Until then, the pages on AI agents and sovereign AI help you decide when a cloud service fits and when local operation does.