Rolling out Claude at a small company: a practical checklist

• First drafts of routine replies

  Owner: whoever leads support or the front office.

  Good looks like: Drafts built from your own templates and FAQ need only light edits, and nothing goes out unread. Track the time per reply and how often a draft has a wrong fact.

• Summaries of internal notes and documents

  Owner: the operations lead.

  Good looks like: The people who were in the meeting or wrote the document agree the summary is accurate and complete, and it takes less time than writing it by hand.

• Internal documents: procedures, checklists, job descriptions

  Owner: the owner of each document.

  Good looks like: Fewer rounds of edits to reach a version the owner signs off, with every fact and figure checked against the source.

• Spreadsheet formulas and data clean-up

  Owner: the finance or operations analyst.

  Good looks like: On data that isn’t sensitive, formulas are tested against rows whose answer you already know before anyone relies on them.

Leave for later: anything sent to customers without a person reading it, decisions about individual people, and work in regulated areas. Anthropic’s Usage Policy, in the version effective 12 November 2026, requires a qualified person to review the output, and the person affected to be told AI was involved, when it’s used for advice or decisions that can affect someone’s health, legal rights, finances, livelihood, or access to housing, education or essential services (checked 25 October 2026).

Keep work off personal accounts. Anthropic’s pricing page lists model training as “Opt-out” on its individual plans (Free, Pro and Max) and “None by default” on Team and Enterprise, and its commercial terms say consumer offerings are covered by its consumer terms instead (checked 25 October 2026). The two can run side by side: a team plan for people, and the API for one workflow.

1. Roles

   Team and Enterprise have four roles: User, Admin, Owner and Primary Owner. There is one Primary Owner per organization, and it can be a service account rather than a person. Admins manage members; only Owners and Primary Owners can open billing (roles and permissions, managing members). Enterprise adds custom roles.

   Our suggestion: make the Primary Owner a company-controlled account, and give a second person the Owner role so the workspace doesn’t depend on one login.

2. Sign-in and single sign-on

   Single sign-on is available on Team and Enterprise plans. Setting it up needs an Owner or Primary Owner, access to your domain’s DNS settings to verify the domain, and your identity provider (setting up SSO). Set it up before you invite people, so everyone signs in the way they already do.

3. Inviting people

   New members need an email address on one of the organization’s allowed domains, and a pending invitation takes up a seat straight away. On new Team organizations, members can invite teammates by default, with each invite going to an admin for approval unless the organization approves automatically; admins can turn member invites off (managing members). For a pilot, decide who invites, and invite only the pilot group.

4. The organization settings that matter

   For Team, Anthropic’s pricing page lists admin controls for remote and local connectors, organization-wide skills, organization instructions, usage analytics, and user and organization spend controls. Owners and Primary Owners can turn off the feedback buttons (“Rate chats”) under Organization settings, Data and Privacy (model training). On Enterprise only: SCIM, audit logs, role-based access and custom data retention.

Training

Anthropic’s Commercial Terms (effective 17 June 2025) say: “Anthropic may not train models on Customer Content from Services.” Its Privacy Center says that by default it won’t use inputs or outputs from its commercial products, such as Claude for Work and the API, to train its models. The condition: if someone explicitly reports feedback or a bug (the thumbs up or down buttons, for example) or otherwise chooses to allow it, that conversation may be used, and feedback is stored for up to five years.

Retention

On the API, Anthropic deletes inputs and outputs from its backend within 30 days, except for services with longer retention you control (such as the Files API), a zero data retention agreement, Usage Policy enforcement and legal requirements. On Claude for Work, chats stay until you delete them (how long data is stored).

Deletion

A deleted chat leaves the chat history immediately and Anthropic’s backend storage within 30 days. Chats flagged by Anthropic’s safety systems for Usage Policy violations are kept longer (same page).

Admin controls

Owners and Primary Owners can turn off the feedback buttons on Team and Enterprise. On Enterprise, they can set custom retention periods for conversations, projects and artifacts (custom retention).

Security documents

Anthropic’s Claude Code docs point to its Trust Center for security documents such as its SOC 2 Type 2 report. Give them to whoever runs your security review.

1. Scope

   Who the policy covers and which tool: the company’s Claude workspace, signed in with the company account. Personal accounts aren’t used for work.

2. Data you may use

   Public information, your own drafts, and internal documents that aren’t confidential to a client or supplier. When unsure, leave it out and ask.

3. Data that never goes in

   Passwords, API keys and other secrets; payment card numbers; government ID numbers; health information; client data a contract keeps from third parties; anything on your list of restricted data.

4. Check before use

   You’re responsible for anything you use or send. Check facts, figures, names, quotes and code against the source before it leaves your hands.

5. Disclose AI use where it matters

   Tell people when they’re dealing with an AI tool rather than a person, when AI helped produce advice or a decision about them, and wherever a client contract or the law asks.

6. Decisions about people

   Don’t use it to decide about a person (hiring, pay, credit, health, housing) without a qualified person reviewing the output first.

7. Who to ask

   The named owner, by name and how to reach them.

8. Report a problem

   Pasted something that shouldn’t have gone in, or got an output that looks wrong or harmful? Tell the owner the same day. Nobody is blamed for reporting.

9. Review

   The owner reviews the policy every quarter, and when Anthropic changes its terms or your plan changes.

A starting point, not legal advice. It doesn’t cover your industry’s rules, your contracts or the law where you operate. Have your own lawyer review it before you adopt it, and add Anthropic’s Usage Policy, which applies to anyone who submits inputs to its products.

• Short sessions on real tasks

  An hour or less, built on the first jobs: each person brings a task they did last week and works through it. Cover the policy, the data list, and how to check an output before using it.

• A shared prompt library

  Keep the prompts that worked for each job in one shared place, with an example of a good result and what to check. Anthropic’s pricing page lists project sharing on Team and Enterprise.

• A named owner

  The rollout owner keeps the library, collects questions and problems, and runs a short refresher when the policy, the plan or the product changes.

Anthropic describes Claude Code as “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools”, in the terminal, IDEs, a desktop app and the browser (overview). Anthropic’s pricing page includes it on Team and Enterprise, and it can also be used with an Anthropic Console (API) account.

1. Repository access and approvals

   In Manual mode, Claude Code starts with read-only permissions and asks before it edits files, runs tests or executes commands, and before its file tools reach outside the folder it was started in. A Bash command you approve can still write anywhere your user account can. Auto mode, the starting mode for interactive terminal and VS Code sessions, has a separate classifier review actions instead, and an organization can turn it off (security). Decide which mode the pilot uses.

2. Shared permission settings

   Allow, ask and deny rules control what it may do; deny rules win, and the settings can be checked into version control for the whole team. Administrators can deploy managed settings that user and project settings can’t override, and can disable the mode that bypasses permissions (permissions).

3. Human code review stays

   Anthropic’s docs say the user is responsible for reviewing proposed code and commands before approving them. Keep every change going through your usual pull request review and tests, with the same branch protection as any other change.

4. Keep secrets out

   Keep credentials out of the repository and in a secrets manager. A deny rule can stop it reading a file such as .env. Claude Code keeps session transcripts on the developer’s machine in plain text for 30 days by default (data usage), so a secret pasted into a session is stored there too.

5. Pilot on one repository

   Start with one repository that isn’t your most sensitive, and a few developers. On Team and Enterprise, admins and owners can see a usage dashboard, with contribution metrics if GitHub is connected (analytics).

The Model Context Protocol (MCP) is an open standard that Anthropic introduced on 25 November 2024 for connecting AI tools to the systems where data lives: a system is exposed through an MCP server, and an AI application connects to it as a client (Anthropic’s announcement). Anthropic’s Claude Code docs say it doesn’t security-audit or manage MCP servers, and suggest writing your own or using servers from providers you trust (security).

• Start read-only

  Connect a system for reading first, such as a knowledge base or a ticket queue. Add write access later, for one action at a time, once you’ve seen how it’s used.

• Least privilege

  Give each connection its own account with only the permissions that job needs, not an administrator’s. On Team and Enterprise, the pricing page lists admin controls for connectors.

• Keep an audit trail

  Log what each connection reads and changes in the connected system, and review it. Anthropic’s pricing page lists audit logs on Enterprise.

• Time on the chosen jobs

  Time a sample of each job for a week before the pilot, then the same job during it. Compare like with like.

• Quality checks

  Have the job’s owner spot-check outputs each week: errors caught, rework, complaints. Faster but wrong isn’t a gain.

• Adoption

  Who uses it each week, for which jobs. Anthropic’s pricing page lists usage analytics on Team, and Claude Code has its own dashboard.

• When to expand or stop

  Agree the rule before you start: expand where a job got faster at the same quality; change the approach where it didn’t; stop where it doesn’t earn its cost.

1. Weeks 1–2

   Decide and set up

   • Name an owner, pick two or three first jobs and write down what good looks like for each.
   • Choose the plan, write the one-page usage policy and decide which data never goes in.
   • Set up the workspace: owners and admins, sign-in, invitations and the data settings.

2. Weeks 3–6

   Pilot with a small group

   • Invite the people who do the first jobs, run a short session on their real tasks and start the prompt library.
   • For developers: one repository, shared permission settings, review as usual.
   • Check results every week against the baseline, and fix the policy where people got stuck.

3. Then

   Expand, or stop

   • Expand team by team where the jobs got faster without losing quality; stop or change course where they didn’t.
   • Connect other systems one at a time, read-only first.
   • Review the policy, the settings and Anthropic’s terms each quarter, and when Anthropic changes them.

• Sensitive data pasted in

  Without a list of what never goes in, people paste customer records, contracts and secrets. Write the list first, and make reporting a slip easy.

• No owner

  Nobody keeps the policy, the settings or the prompt library, so each drifts. Name one person before the first invitation.

• No policy

  Everyone guesses what’s allowed, and the careful people stop using it. One page is enough to start.

• Measuring nothing

  Without a baseline, nobody can say whether it helped, and the decision to expand or stop becomes a matter of opinion.

• Everyone on day one

  A company-wide launch spreads mistakes before you’ve seen them. Pilot with the people who do the first jobs, then expand.

• Plans and pricingAnthropic (claude.com), checked 25 Oct 2026
• API pricingAnthropic’s platform docs, checked 25 Oct 2026
• Commercial Terms of ServiceAnthropic, checked 25 Oct 2026
• Is my data used for model training?Anthropic Privacy Center, checked 25 Oct 2026
• How long do you store my organization’s data?Anthropic Privacy Center, checked 25 Oct 2026
• Configure custom data retention controls for Enterprise plansAnthropic Privacy Center, checked 25 Oct 2026
• Roles and permissionsAnthropic Help Center, checked 25 Oct 2026
• Manage members on Team and Enterprise plansAnthropic Help Center, checked 25 Oct 2026
• Set up single sign-on (SSO)Anthropic Help Center, checked 25 Oct 2026
• Claude in Amazon BedrockAnthropic’s platform docs, checked 25 Oct 2026
• Claude Models on Google CloudAnthropic’s platform docs, checked 25 Oct 2026
• Claude in Microsoft FoundryAnthropic’s platform docs, checked 25 Oct 2026
• Usage PolicyAnthropic, checked 25 Oct 2026
• Claude Code overviewClaude Code docs, checked 25 Oct 2026
• Configure permissionsClaude Code docs, checked 25 Oct 2026
• Claude Code securityClaude Code docs, checked 25 Oct 2026
• Claude Code data usageClaude Code docs, checked 25 Oct 2026
• Claude Code analyticsClaude Code docs, checked 25 Oct 2026
• Introducing the Model Context ProtocolAnthropic, checked 25 Oct 2026
• Anthropic Trust CenterAnthropic, checked 25 Oct 2026