The incident partners fear
An associate asks ChatGPT Enterprise to “draft a status update to the client and opposing counsel.” The model pulls language from internal strategy notes. The draft cc's the wrong party. It ships before a partner reviews tone, privilege, or settlement posture.
Training slides said “always review AI output.” There was no enforced stop between draft and send — and no log that shows who approved what left the firm.
Why permission cleanup isn't enough
Ethics opinions focus on competence and confidentiality: know what you feed the model, verify output, avoid waiving privilege. That is lawyer responsibility — but firms also need firm-wide controls when dozens of staff use AI daily.
- Matter walls in your DMS don't always match SharePoint sharing
- Copilot doesn't know which paragraph is work product vs. client record
- External sends are where malpractice carriers see claims — not summarization
Controls a 120-person firm can run without an AI practice group
Client matter access boundaries
AI may read only matters tied to the user's active cases — not firm-wide search across all clients.
Mandatory approval for external email
Any message to domains outside the firm or client approved list routes to responsible attorney before send.
Block bulk document export
Prevent AI-assisted zip downloads or mass copy of discovery folders to personal drives.
Audit trail for privilege review
Log draft → approval → send so risk management can reconstruct decisions.
Questions your managing partner should get clear answers to
- Can AI email anyone outside the firm without a lawyer approving?
- Can AI access matters the user isn't staffed on?
- If a client asks what AI touched their files, can we produce a log?
- Does governance apply to Copilot and ChatGPT the same way?