What should an AI acceptable-use policy say, and how do you enforce it in Claude?

An AI acceptable-use policy should state which AI tools and accounts are approved, what data may go into them, which uses are banned or need human review, who approves new connectors and agents, how output is checked and disclosed, and what is logged. In Claude, powered by Anthropic, many of those rules can be enforced by settings on Team and especially Enterprise plans. The rest need a process with a named owner.

Why does a written policy need product settings behind it?

A policy that says 'do not paste client data into unapproved tools' is only as good as people's memory. Settings turn some of those sentences into rules the system enforces: a personal account cannot sign in from the corporate network, a write action in a connector asks for approval, a chat is deleted after the retention period. The goal is to enforce what you can, monitor what you cannot enforce, and train people on the remainder. Each rule in the policy should name which of those three it relies on.

What sections should an AI acceptable-use policy have?

• Purpose and scope: which people (employees, contractors, temps), which tools (chat apps, coding assistants, add-ins, agents) and which company systems the policy covers.
• Approved tools and accounts: the named tools, the rule that work happens only in company-managed accounts, and how to request a new tool.
• Data rules: which data classes may go into which tool, tied to your existing data classification. Spell out regulated data such as PHI, cardholder data and client-confidential material.
• Prohibited uses: activities banned outright, including anything barred by the vendor's own usage policy.
• High-risk uses requiring human review: decisions or advice about individuals in areas such as employment, credit, insurance, legal and medical, with who must review and how that review is recorded.
• Output responsibility: the user owns the output, must check facts, figures and citations, and must not present unreviewed AI output as final work.
• Disclosure: when customers, candidates or other affected people must be told AI was used.
• Connectors, Skills, plugins and agents: who may add them, the approval steps, required owners and review cadence.
• Software development: rules for coding assistants, including which repositories, which MCP servers and what review applies before merge.
• Records and retention: how long AI conversations are kept, legal holds and how AI content fits the records schedule.
• Monitoring and privacy notice: what the company logs and reviews, so employees are told before monitoring starts.
• Incidents: how to report a wrong output that reached a customer, a data leak or a suspected prompt injection.
• Training and attestation: required training before access and annual acknowledgment.
• Exceptions and review: who can grant exceptions and how often the policy is reviewed. Quarterly is reasonable while products change this fast.

Which Claude settings enforce which policy rules?

The table maps common policy rules to the control that enforces them, with the plan each control requires as of October 2026. Team covers 2 to 150 users and includes SSO, domain verification, connector admin controls, organization-wide Skills and usage analytics. Enterprise adds role-based access, SCIM, audit logs, the Compliance API, custom data retention, IP allowlisting, network-level access control, inference hooks and spend limits.

What can a Team plan enforce, and where do you need Enterprise?

A Team plan can carry a sound policy for a smaller company: SSO, domain verification, owner-controlled connectors, organization-wide Skills with the option to turn off user-created Skills, organization instructions, and usage analytics. What it cannot do is vary rules by role, push identity changes automatically through SCIM, export audit logs, feed a compliance archive through the Compliance API, shorten retention, or inspect content with inference hooks. If your policy has rules like 'only finance may use the ERP connector' or 'chats are deleted after 90 days,' those are Enterprise features today. Our comparison of Claude Team and Enterprise covers the full difference.

Which rules need a process instead of a setting?

Some of the most important rules have no switch. Name an owner and an evidence trail for each.

• High-risk recommendations about people. Anthropic's Usage Policy, in the version effective November 12, 2026, requires that a qualified person meaningfully review a high-risk AI recommendation, with authority to change it, before it is delivered or acted on, and that affected individuals be told AI was used. Its high-risk areas include legal, medical, finance, credit, insurance, housing and employment. Internal drafting and analysis that is not the final recommendation is excluded. Mirror this in your policy, since the vendor's terms bind you anyway.
• Accuracy of figures, citations and legal statements. Require the user to verify against the source before anything leaves the company.
• Client and contractual confidentiality. Some client contracts restrict AI processing of their data even when your own policy allows it. Legal or account owners keep a list.
• Prompt injection reporting. Tell people what odd behavior looks like, such as Claude trying to send data somewhere unexpected after reading an email or file, and where to report it.
• Intellectual property. State who reviews AI-generated content used in products, marketing or code before release.

What does good policy language look like?

Short, testable sentences work better than principles. These sample clauses are starting points for counsel to adapt.

• Approved tools: Employees may use AI tools for company work only through accounts provisioned by IT. Approved tools are listed on the AI tools intranet page. Using personal AI accounts for company work is prohibited.
• Data: Restricted data, as defined in the Data Classification Standard, may be used only in AI tools marked approved for restricted data, and only for the purposes listed for that tool.
• Connectors: No connector, plugin or agent may be connected to a company system without approval from IT Security and a named business owner. Write access requires a documented reason and per-action approval unless IT Security grants an exception.
• Review: The employee who uses AI output is responsible for it. Figures, citations and statements of fact must be checked against the source before external use.
• High-risk decisions: AI output may inform, but not make, decisions about hiring, performance, credit, pricing for individuals or benefits. A qualified reviewer must approve and record the decision.

How do you roll the policy out?

• Draft with IT, security, legal, HR and one or two business leaders who will actually use Claude. Policies written only by risk teams tend to ban what people need and get ignored.
• Configure the settings in the control map before access opens, then test each one: try a blocked connector, a blocked model and a personal account from the office network.
• Publish the policy with a one-page summary and require acknowledgment as part of training.
• Tell employees what is logged, which you will need anyway if you enable the Compliance API or inference hooks.
• Review quarterly. Each review should check new Claude features against the policy, since a new capability may need a new rule or a new toggle.