Start with the work Claude is allowed to do
A pilot can pass a login test and still fail its first document review. The missing decision is often which files, actions, and users belong inside the approved workflow.
Claude Enterprise governance starts with that boundary, before invitations go out. A firm of 25 to 200 lawyers needs a launch plan its IT team can operate and its risk owner can approve.
For Atlanta law firms and regulated mid-market teams across the Southeast, the practical question is what a technical rollout will permit. Legal guidance and implementation planning serve different jobs.
This guide explains technical deployment and governance decisions for a bounded pilot.
Separate Enterprise access from Cowork authority
Claude Enterprise access and permission to use Cowork deserve separate decisions. Anthropic documents both local and cloud Cowork sessions, with distinct controls for cloud execution and browser access.
That distinction changes the architecture review. A chat workflow using an approved upload differs from an agent working with a folder or an external tool.
Firms that also run Microsoft 365 Copilot still need a separate Claude pilot boundary. See our Claude Enterprise vs Microsoft 365 Copilot guide for how the governance choices differ.
Draw the intended data flow before choosing settings. Show the user’s device, approved files, Claude service, identity provider, and monitoring destination. Mark every place information can leave the approved workflow.
Use Anthropic’s Team and Enterprise Cowork guidance to check the controls available in your tenant. Feature availability and defaults can change.
Launch Now (MVSL): keep the first boundary small
Cocha’s deployment motion starts with Launch Now, the Minimum Viable Secure Launch. Fast-Follow and Harden & Scale address later expansion under separately agreed scope.
The launch recommendation is deliberately narrow: Enterprise as the floor, autonomy disabled, connectors and Model Context Protocol (MCP) disabled, and approved local files only. Browser and computer use stay outside the initial scope.
These are Cocha’s recommended launch conditions, not a statement about Anthropic’s factory defaults. Verify how each restriction is enforced before giving users access.
If the available controls cannot support the approved boundary, defer that capability. Cowork can wait while the team validates a more limited workflow.
Local files still need a data decision
A local folder is an access boundary. It does not establish that model processing stays on the device or that information never reaches a service.
Start testing with synthetic or public documents. For any later use of internal material, record the approved data categories and the person authorized to approve them.
Exclude privileged or client-restricted matter material wherever the approved policy requires it. Review client restrictions with the firm’s responsible lawyer before admitting matter documents.
Avoid mounting a user’s home directory or an entire synchronized document library. Use a dedicated folder with only the approved test material, then check what the workflow can read and change.
A useful first exercise is deliberately ordinary: summarize a public policy document, compare the answer against the source, and inspect the resulting files. Record the steps rather than assuming the task stayed inside scope.
Test identity, egress, and evidence separately
The architecture needs more than a green sign-in screen. Each boundary needs its own test and named owner.
Identity and endpoint readiness
Confirm the Enterprise organization, intended identity provider, and pilot membership. Test SSO and MFA through the approved identity path, including a user who should be denied access.
Document who can administer Claude and who can change the pilot’s scope. Test removal of a pilot user and record any remaining sessions or access questions for investigation.
For desktop use, confirm supported devices and managed configuration deployment. Keep an inventory of the app versions used in the pilot because control behavior can differ across versions.
Our Claude Enterprise deployment runbook covers the administrator setup sequence. This launch decision also needs evidence about the workflow that setup permits.
Egress needs a route-by-route check
Anthropic documents that Cowork network egress permissions do not apply to web fetch, web search, or MCPs, including Claude in Chrome. A network allowlist therefore needs a clearly defined scope.
List each enabled route and its controlling policy. Verify that prohibited browser access, connectors, and MCP tools remain unavailable in the pilot.
Repeat the test in a new session after changing network settings. Anthropic says network changes take effect when a new Cowork session is created.
Monitoring has its own data boundary
Do not assume a general audit export proves what happened inside an agent session. Test which events reach the system your security team will use during an investigation.
Current Compliance API documentation includes Cowork sessions. Its coverage differs from the narrower audit log export.
Anthropic also documents OpenTelemetry monitoring for Cowork. Treat telemetry configuration as a data-handling decision: exported tool details and file paths can contain sensitive information.
Choose the event content, access permissions, and retention deliberately. Verify a sample task appears at the destination before relying on that evidence during launch review.
Read local guidance without turning it into a settings checklist
Georgia firms can read the State Bar’s Generative AI Toolkit with their responsible lawyer. Its July 9, 2026 version describes a living guidance document, not a formal advisory opinion.
Use that discussion to identify questions for the deployment team. Keep legal interpretation with counsel and record which approved workflows the technical configuration must support.
Write the acceptable use policy around the pilot
A pilot acceptable use policy should tell a user what to do when a document falls outside the approved scope. Broad encouragement to use AI responsibly leaves that decision unresolved.
Name the approved workflows and data categories. State who reviews outputs, where generated files belong, and how users report unexpected access or behavior.
Include a stop instruction for suspicious content that asks Claude to ignore the task or expose information. Keep human review in the workflow before any output reaches a client or an external system.
The firm’s responsible lawyer owns legal interpretation and client-specific decisions. Technical controls provide evidence for that review; they cannot settle those questions on their own.
Make go/no-go a recorded decision
Residual risk acceptance should identify what remains unverified and who can accept it. A launch deadline is no substitute for that record.
Use a short readiness record with these fields:
| Decision | Evidence to retain |
|---|---|
| Pilot scope and prohibited data | Approved workflow and folder boundary |
| Identity and device readiness | Sign-in, denial, and removal test results |
| Disabled capabilities | Configuration record plus user-level tests |
| Monitoring readiness | Sample events and unresolved visibility gaps |
| Operational ownership | Support contact, incident route, and stop procedure |
| Residual risk | Named owner, accepted limitations, and review date |
Pause if restricted material is reachable, required approval controls cannot be enforced, or nobody owns the remaining risk. Record the blocking condition and the evidence needed to reconsider.
Quick answers
What should a firm decide before inviting users to Claude Enterprise?
Which files, actions, and users belong inside the approved workflow. Claude Enterprise governance starts with that boundary before invitations go out.
How do Enterprise access and Cowork authority differ?
Enterprise access and permission to use Cowork deserve separate decisions. A chat workflow using an approved upload differs from an agent working with a folder or an external tool.
What belongs in a Minimum Viable Secure Launch boundary?
Enterprise as the floor, autonomy disabled, connectors and MCP disabled, and approved local files only. Browser and computer use stay outside the initial scope. These are Cocha’s recommended launch conditions, not Anthropic factory defaults.
What evidence does monitoring need before go/no-go?
Test which events reach the system your security team will use during an investigation. Verify a sample task appears at the destination before relying on that evidence during launch review.
When should a pilot pause?
Pause if restricted material is reachable, required approval controls cannot be enforced, or nobody owns the remaining risk. Record the blocking condition and the evidence needed to reconsider.
Expand one capability at a time
Fast-Follow can address provisioning, monitoring integration, or data controls that need further work. Priorities should follow the pilot’s findings rather than a promise to enable everything next.
Harden & Scale can evaluate connector and MCP access after the underlying permissions are understood. Start with read-only access where appropriate, verify the tool behavior, and require explicit review before enabling writes.
Give every expansion its own acceptance test. A successful document summary does not establish that a connector can safely reach a matter repository.
Related guides cover the next decisions in more detail: what information may enter an AI agent workflow, the evidence an administrator should capture after each agent task, how to approve a new AI workflow before it scales, approval gates for adding AI connectors after a pilot, and how to enforce an AI acceptable use policy after launch.
Atlanta and Southeast buyers should ask a technical deployment provider to explain that progression, including who makes changes and who approves the result. Agree whether the engagement is advisory or hands-on before work begins.
For a focused starting point, request the free AI Readiness Snapshot. It makes no changes to your environment and requires no software installation.
Use the findings to decide what your team can handle and where deeper work is needed. The aim is AI use your firm can defend.
