Claude Code Enterprise: What to Decide Before Rollout

Claude Code Enterprise is not a product of its own. It is Claude Code, Anthropic’s tool for software work, on the Claude Enterprise plan. On new plans, the tool comes with the seat, and every token it uses is billed. Before your engineers use it, decide five things: permissions, repositories and secrets, review, provenance and spend.

Picture a Monday in your engineering team. One engineer installs Claude Code and gives it a task in the main repository. By Friday, half the team has done the same. Nobody has decided what the tool may read, who reviews its work or who sees the bill. This guide covers the five decisions to take first.

PromptPartner deploys AI in mid-sized and larger companies in Europe. This guide is our reading of Anthropic’s own documentation, opened on 7 October 2026. It is written for the CTO of a software company. The product changes often, so follow the links before you decide.

Key takeaways

On Team and on new Enterprise plans, Claude Code comes with the seat.^[1]
New Enterprise plans have no usage cap and no included usage. Every token is billed.^[2]
Current versions do not ask before each action. A second model reviews some actions, and most file edits in the project folder run without a check.^[3]
Five decisions come first: permissions, repositories and secrets, review, provenance and spend.

What “Claude Code Enterprise” means: a tool inside a plan

Claude Code is Anthropic’s tool for software work. Claude Enterprise is the plan that includes it. They are not the same thing: you buy the plan, and Claude Code comes with the seat.

Anthropic calls Claude Code “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools”.^[4] In plain words: an engineer gives it a task, it works through the steps, and the engineer reviews the result.

The plan adds what a company needs around the tool: chat, admin controls, and security and compliance features.^[2] Our page on Claude Enterprise describes the plan as a whole.

You can also run Claude Code without the plan, through the Claude Console or a cloud platform from Amazon, Google or Microsoft. Anthropic calls the Team or Enterprise plan “the default recommendation”, and the other routes lose some features.^[5]

Team or Enterprise? Both let you set the rules for Claude Code centrally.^[6] Audit logs are for Enterprise only.^[7] Anthropic’s guides for access by group and for limits by group are written for Enterprise.^[8]^[9] For Team, Anthropic’s own pages do not agree. The Team plan page lists role-based permissions and domain capture.^[10] The Claude Code documentation names both as additions of Enterprise.^[6] On Team, check your admin console before you plan around groups.

How access and billing work on Team and Enterprise

Claude Code is part of every Team seat and of the single Enterprise seat on new plans.^[1] It has no price of its own. The plans differ in how usage is charged, as the table shows. Seat prices are in our guide to Claude Enterprise pricing.

Source: Anthropic Help Center, read on 7 October 2026.
	Team	Enterprise, new plans	Enterprise, older contracts
Is Claude Code included?	Yes, in every seat	Yes, in the single Enterprise seat	Only on “Chat + Claude Code” seats and on Premium seats
How usage is charged	A usage limit for each member. Premium seats have more^[10]	No limit per seat. Every token is billed at API rates^[1]	By use on “Chat + Claude Code” seats. By seat on Premium seats

If your contract is older, check the seat type. On the old two-seat plans, a person with a chat-only seat cannot use Claude Code.^[8] Anthropic moves these contracts to the single seat at their next renewal.^[11]

Is Claude Code unlimited on the Enterprise plan?

In volume, yes. In cost, no. On new Enterprise plans, Claude Code has no usage cap, and all of its usage is billed: these plans “have no plan or seat-level usage limits”.^[2] An owner can set a limit to unlimited, “but consumption will still be billed”.^[11] On Team, each member has a usage limit. Work continues past it only if the company has switched on usage credits.^[1]

Anthropic’s cost page for Claude Code still describes an allowance for each seat on Team and Enterprise plans.^[12] The Help Center says that new Enterprise plans have none,^[2] and we follow the Help Center.

Why does Claude Code cost more than chat?

Claude Code costs more than chat because a coding session uses far more tokens, the unit that AI use is measured in. Each turn carries file contents, tool calls and several steps of reasoning. In Anthropic’s words, “one debugging session can consume more than a day of chat”.^[12]

Anthropic publishes its own average: “around $13 per developer per active day and $150-250 per developer per month”.^[12] Treat it as a first orientation, not as a budget. The page does not say which models or how many companies are behind the figure, and it says that costs “vary widely”.

Five decisions before your engineers use Claude Code

The five decisions are permissions, repositories and secrets, review, provenance and spend. Claude Code runs without any of them, which is why they are easy to skip. Each one has a setting or a record behind it, and each one needs an owner in your company.

Five decisions before engineers use Claude Code in a company: permissions, repositories and secrets, review, provenance and spend

Decision 1: permissions. Who may run what?

Set the starting mode of Claude Code and your company’s rules centrally, before the first install. This comes first because the default has changed. In current versions (documentation of October 2026), a session in the terminal or in VS Code starts in auto mode. The engineer is not asked before each action, and a second model checks the actions instead.^[3] It does not check everything: most reads and edits inside the project folder run without a check. On Enterprise plans, the checks count toward your token usage.^[3]

Anthropic states the limit itself: “Auto mode reduces permission prompts but does not guarantee safety.”^[3] If you have read that Claude Code asks before every change, that is Manual mode, and it is no longer the default.

Your company’s settings come first. Managed settings “take precedence over local developer configuration”.^[5] You set them in the admin console, or IT delivers them through device management. The two routes are not equally strong. Settings from the admin console are, in Anthropic’s words, “a client-side control, not a security boundary”. On a laptop that IT does not manage, a user can bypass them. Settings that come through device management are stronger, because the operating system can protect the file.^[13] Four things to set there:

The starting mode. An admin can switch auto mode off for the whole company.^[3] The mode that skips the permission prompts can be switched off too.^[14]
What is never allowed. Deny rules win: “If a tool is denied at any level, no other level can allow it.”^[14] Engineers can add to the company’s lists, but they cannot remove entries.^[5]
Who gets the tool. On Enterprise, you can limit Claude Code to one group, for example Engineering.^[9]
Which account. Two settings together require that every login belongs to your company, not to a private account.^[15]

One thing is not a control: a file of written instructions in the repository. “Permission rules are enforced by Claude Code, not by the model.”^[14]

Decision 2: repositories and secrets. What may it read?

Decide which repositories are in scope, keep passwords and keys out by rule and by sandbox, and know where the code goes. Claude Code reads the codebase it is started in, so name the repositories where engineers may start it.

A rule alone does not protect a secret. You can block a file with a deny rule for its path. But the rule does not cover “a command that reads files without naming them”.^[14] Anthropic’s answer is the sandbox: the operating system sets which files and network domains a shell command, and the processes it starts, can reach. It has a limit of its own: “The sandbox covers shell commands only.” Claude’s file tools, connected servers and hooks run outside it, so you need the rule and the sandbox together.^[16] Check two more things: “The sandbox is off by default.” And on native Windows, commands run without it.^[16]

Then the connected systems. MCP is the standard through which Claude connects to other tools. The default is open: “By default, anyone running Claude Code can connect any MCP server they choose.”^[17] An admin can replace that with a list of allowed servers.

Where does the code go? Claude Code runs on the laptop and sends prompts and model outputs over the network, encrypted in transit.^[18] A prompt can carry file contents, so your code is part of what travels.^[12] Four facts matter for a software company:

Training. Under commercial terms, which include Team and Enterprise, Anthropic does not train its models on your code or prompts, unless you choose to provide them.^[18]
Copies on the laptop. Sessions are stored there in plain text, for about a month by default.^[18]
Zero data retention. It means Anthropic does not keep prompts and answers after the response, with exceptions for law and misuse. It “is not included in the standard Claude for Enterprise plan”.^[19] Anthropic’s account team switches it on for accounts that qualify.^[18]
Europe. Anthropic states that “data is stored in the US”.^[20] For data residency in Europe it points to the cloud platforms,^[21] and that route loses some Claude Code features.^[5] What to ask Anthropic before you sign is in our pricing guide.

Decision 3: review. Who reads what the agent wrote?

A person reads and approves every change that Claude Code writes, before it is merged, as for any other pull request. Inside a session, Anthropic puts the duty on the user: “You’re responsible for reviewing proposed code and commands for safety before approval.”^[22] Apply the same standard at the merge.

An analogy helps here: Claude Code reacts like a junior employee. You ask a junior for the plan first, and you read the work before it ships.

Anthropic offers an automated reviewer, Code Review, as a research preview. It adds comments. Its findings “don’t approve or block your PR”, so a person still approves. It is billed separately, and Anthropic puts the average at $15 to $25 for each review.^[23]

Anthropic’s own guidance gives you a sentence for your engineering rules: “If you can’t verify it, don’t ship it.”^[24] Set the review as a rule in the code host, with branch protection, so that the code host enforces it.

Decision 4: provenance. Can you trace a change to the person who asked for it?

Only in part. By default, a commit made with Claude Code carries a line that names the model.^[25] A commit from a session in Anthropic’s cloud, or from a Remote Control session, also carries a link to that session.^[25] The model line shows that the tool took part, not what was asked. An engineer can switch both lines off, unless the company sets them in managed settings. Set them there.

The record that links an action to a person is separate. Anthropic calls it the audit data source. Each event carries the identity of the user who triggered it.^[26] The events travel over OpenTelemetry, an open standard, to your own monitoring system. Three limits to know:

The records are off until you switch them on, for everyone, through managed settings.^[26]
The text of the prompt is not recorded by default. Alerts and analysis are the job of your own monitoring.^[26]
The person is named only when engineers sign in with their Claude account. With an API key or a cloud platform, “there is no Claude account in the session”.^[26]

The audit logs of the Enterprise plan do not fill this gap. They record sign-ins, users, chats, projects and files, and their event list names no Claude Code action.^[7]

Decision 5: spend. Who sets the limits, and who reads the bill?

Set spend limits before the first Claude Code session, by group and by person, and give finance one report by department. Do not count on a default. Anthropic’s consumption guide is plain: “No limit anywhere = no limit.”^[9]

On Enterprise, admins can set limits at three levels: the whole company, a group and a single person.^[9] Use the company limit with care. When it is reached, everyone stops at the same time. At a limit, the running request finishes and the next one is blocked. A task with many steps can stop halfway.

Three facts help engineering and finance read the bill the same way:

There is a report. It shows token use and estimated spend for each person and each model, updated daily.^[27] On Enterprise, the cost data can also be grouped by department.^[9]
The engineer’s screen and the invoice differ. Claude Code shows costs at list price. If you pay contracted rates, the figures on screen will not match your bill.^[12]
The model changes the bill. On Enterprise, a model can be switched off for everyone, or allowed for some roles only.^[9]

The general steps to keep an Enterprise bill under control are in our pricing guide.

Who decides what: engineering, security and finance

No single person should own all five decisions about Claude Code. The split below is our suggestion.

Who decides what for Claude Code in a company: what the engineering lead, security and finance each decide and need to see

The engineering lead decides which teams start, which commands and tools they need, and who reviews.
Security decides the starting mode, the rules set centrally for the company, which systems may connect, and where the records go.
Finance decides the limits by group and by person, and reads the monthly report by department.

The Enterprise plan has roles for this split. Anthropic gives two examples. A Finance role has access to billing and no access to Claude Code. An Engineering Lead role has Claude Code and a view of the analytics.^[8] One detail decides who can read the bill: “Enterprise Admins can view all analytics except Spend.”^[27]

What to roll out first

Our advice: roll out Claude Code in the same order as the rest of a Claude rollout. The rules first, then one team, then weekly practice. Anthropic describes its own practice as an admin-paced rollout: a small group first, then watch the records, then widen access.^[28]

The rules. Set the managed settings and the limits, and switch on the records, before the first engineer installs the tool.
One team. Pick a team with a lead who wants it. Give it the tool for a few weeks and measure what it uses. That is the baseline for your budget.^[12]
Weekly practice. Give the team a fixed hour each week for real tasks from their own work. Then widen access, one team at a time.

Engineering does not have to be the first department. In a software company, the return from AI often shows fastest in go-to-market: sales, marketing and customer success. Finance and engineering follow, on the same rules. See our guide to an AI governance framework and our page for software and SaaS companies.

Skills come with the account: once an engineer is signed in, the skills and plugins that the company provides “load in Claude Code automatically”.^[1] Claude for business covers skills and the five steps of a rollout across teams. Our guide to the Claude desktop agent covers the agent that works in files, folders and connected apps.

PromptPartner’s own reference case is a go-to-market rollout, not an engineering one. At a European software company above €100 million in annual recurring revenue, we shipped five working tools into the systems the team already used. One of them turns a week of work on sales material into minutes. That rollout took longer than four weeks. Read the case study.

If you do not yet know where AI pays off in your company, measure first. Our method starts with a four-week assessment. You can stop after it. After the assessment, the standard programme rolls out Claude Enterprise or ChatGPT Enterprise on a governed foundation, with use cases for one department and training for your team.

This guide summarises Anthropic’s public pages. It is not legal advice, and your own contract decides what applies to you.

Frequently asked questions

Can I use Claude Code with the Enterprise plan?

Yes. On new Enterprise plans, Claude Code is part of the single Enterprise seat. On older contracts it depends on the seat type, so check which seats your contract has.

Is Claude Code the same as Claude Enterprise?

No. Claude Code is Anthropic’s tool for software work. Claude Enterprise is the plan that includes it, together with chat and the admin controls. You buy the plan, and Claude Code comes with the seat.

Why is Claude Code so expensive on the Enterprise plan?

On new Enterprise plans the seat includes no usage, so every token is billed. A coding session with Claude Code uses far more tokens than a chat, because each turn carries file contents, tool calls and several steps of reasoning.

Is Claude Code unlimited on the Enterprise plan?

In volume, yes. In cost, no. New Enterprise plans have no usage cap and no included usage. Everything is billed, and a limit set to unlimited is still billed.

Does Claude Code ask for permission before it changes files?

Not by default. In current versions, as of October 2026, a Claude Code session in the terminal or in VS Code starts in auto mode. The engineer is not asked each time. Most reads and edits inside the project folder run without a check. A second model reviews other actions, for example shell commands and network access. An admin can switch auto mode off for the whole company.

Can a European company keep its code in Europe with Claude Code?

Not on the Enterprise plan alone, as far as Anthropic’s public pages show. Anthropic states that data is stored in the US. For data residency in Europe it points to the cloud platforms, and that route loses some features.