The admission decision comes before the settings review
A vendor demo can look clean and still leave one question unresolved: which information is allowed into the workflow at all.
For law firms and regulated mid-market teams in Miami-Fort Lauderdale and across Florida, that question shows up when chat uploads give way to agent tools that can reach folders, connectors, or local files. The technical team needs an approved-input decision, not only a list of product toggles.
This guide focuses on the intake boundary for confidential and restricted material. It is educational deployment guidance, not legal advice.
Separate three information classes before you enable tools
Start with classes the firm can defend in a review, then map tools to those classes. Keep the labels operational so administrators and risk owners use the same language.
- Public or already published material. Policies, statutes, marketing pages, and other documents the firm would treat as public. These are the safest first pilots.
- Internal administrative material. Firm procedures, non-client operational docs, and synthetic test files created for the pilot. Still control retention and access.
- Restricted client or matter material. Anything counsel treats as confidential, privileged, or otherwise restricted. Admission here is a recorded decision, not a default.
Do not treat “the user can see the file” as approval to send it into a model or an agent session. Visibility on a workstation and authorization for an AI workflow are different permissions.
Turn vendor review into a short decision record
Florida lawyers reviewing generative AI tools often begin with Florida Bar Ethics Opinion 24-1 (January 19, 2024). The opinion page states that advisory ethics opinions are not binding. Read it with the firm’s responsible lawyer. Do not treat this article as an interpretation of confidentiality, consent, billing, or advertising duties.
What the technical team can prepare is the evidence counsel and risk owners usually ask for when a third-party tool may receive firm or client information:
| Vendor question | What to retain |
|---|---|
| Data retention after a prompt or upload | Contract or admin documentation stating retention and deletion |
| Whether inputs are used for model training or self-learning | Current vendor policy for the Enterprise plan in use |
| Who can access stored prompts, files, or embeddings | Role model and support-access notes |
| Breach or process notice commitments | Security addendum or equivalent notice language |
| What happens when the service ends | Export and deletion path, including backups if documented |
Record the plan, the date checked, and the person who checked it. Revisit the record when the vendor changes terms or when the firm expands from chat to agent tools.
Opinion 24-1 discusses researching a program’s policies on data retention, data sharing, and self-learning as part of protecting confidentiality. Keep legal conclusions with counsel. Keep the technical checklist above as operational evidence for that review.
Build an approved-input decision record
Use one record per workflow, not one vague firm-wide statement. A chat summarization pilot and a folder-based agent pilot should not share a blank approval.
| Field | Example content |
|---|---|
| Workflow name | Public policy summarization pilot |
| Tool and plan | Named Enterprise product and environment |
| Allowed information classes | Public only; synthetic test files |
| Explicitly excluded classes | Client matter documents; privileged correspondence |
| Human review owner | Named role before external use of outputs |
| Approval path for confidential information | Named lawyer or risk owner; no self-approval by end users |
| Stop instruction | What users do if a restricted file appears in scope |
| Review date | Next check after configuration or vendor-policy change |
Keep the approval path short enough that users follow it. If every exception routes through an unnamed committee, people invent workarounds.
Test the boundary with ordinary tasks
Before admitting restricted material, run tasks that make the boundary visible:
- Summarize a public document from an approved folder. Confirm the agent cannot open neighboring folders that hold internal or client files.
- Attempt a prompt that asks for content outside the approved mount. Record whether the tool refuses, fails, or quietly reaches further than expected.
- Save the output to an approved location and inspect what else was created. Unexpected copies are part of the intake decision, not only a monitoring problem.
These are technical acceptance tests. They do not prove legal compliance. They show whether the configuration matches the written classes.
Connect intake to the shared launch posture
Intake decisions sit next to the launch boundary in Cocha’s Claude Enterprise governance work for Southeast buyers. Launch Now keeps Enterprise as the floor, autonomy off, connectors and Model Context Protocol (MCP) off, and approved local files only. That posture is explained in Claude Enterprise governance for law firms and in the live Claude Enterprise deployment runbook.
Do not expand connectors or MCP to “make intake easier.” Expand only after the allowed classes, human review owner, and stop path are written down. After a task runs, use the companion AI agent audit evidence checklist for what administrators should capture and when to stop.
Florida firms evaluating Claude or similar agents should ask a technical provider to show how approved inputs are enforced in the tenant they will use. Defaults and labels change. Verify the controls in the target environment.
Quick answers
What is an approved-input decision?
A recorded choice about which information classes may enter a named AI workflow. Visibility on a workstation and authorization for an AI workflow are different permissions.
Which information classes should a firm separate first?
Public or already published material; internal administrative material; and restricted client or matter material. Admission of restricted material is a recorded decision, not a default.
What vendor evidence should the technical team retain?
Retention and deletion terms, training or self-learning policy for the Enterprise plan in use, who can access stored prompts or files, breach or process notice commitments, and export or deletion paths when the service ends.
When should a firm pause information admission?
Pause if restricted material is reachable without an approval record, if vendor retention answers are missing, or if nobody owns human review of outputs.
Close with a recorded go / no-go on information admission
Pause the workflow if restricted material is reachable without an approval record, if vendor retention answers are missing, or if nobody owns human review of outputs. Record the blocking condition and the evidence needed to reopen the decision.
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 which workflows are ready for a bounded pilot and where intake controls still need work.
