Attorney-client privilege is the bright line that governs every AI decision a law firm makes. Get the architecture right and AI fits cleanly into intake, drafting, and communication. Get it wrong and you introduce privilege risk that a single careless disclosure could surface in litigation. This post is about getting the architecture right.
The privilege question in one paragraph
Attorney-client privilege protects communication between an attorney and a client that was made in confidence for the purpose of seeking or providing legal advice. The privilege can be waived by disclosure to third parties — including, in the right conditions, AI vendors. The architecture around how data flows determines whether privilege is preserved or broken.
The six architectural commitments
- Private or access-controlled deployment. The model runs inside a boundary the firm controls — cloud tenancy or on-premise. No shared public inference.
- No training on firm-specific client data. Contractually enforced. The vendor does not improve its general model on your content.
- Role-based access controls. Attorney-level, paralegal- level, intake-level. Nobody sees more than their role requires.
- Minimum-necessary data handling. Each automation layer sees only the fields it needs. The intake workflow doesn't need discovery materials.
- Segmented workflows for sensitive matter types. The most sensitive matters run in tighter-access environments than the routine ones.
- Review checkpoints before attorney-facing output is relied upon.No AI-generated content is released without attorney review.
What automation can and can't do
Automation supports intake, organization, first-draft generation, and client communication. It does not replace attorney judgment on substantive legal work. That line is architectural: the workflow is designed so the attorney is always the last gate before client-facing output.
BAAs and the vendor relationship
AI vendors that touch firm data should be contractually bound as service providers under the firm's confidentiality obligations. The BAA or equivalent agreement should cover: no training on firm data, data-use limitations, breach notification, subcontractor flow-down, audit rights. See HIPAA BAAs for AI vendors for the healthcare-specific parallel.
Privilege logs and AI-generated content
Firms should treat AI-generated drafts and intake records the same as any other firm work product for privilege-log purposes. The architecture should make this straightforward — clean, auditable records of what was generated, when, by whom, and with what review.
The specific deployments we run
Our voice agents for law firms run inside access-controlled environments with the posture above. Intake automation and document automation follow the same architectural rules. For the broader bar-rules context see Bar rules and AI for Virginia, DC, and Maryland attorneys.
If you want to talk through your specific firm's posture, scope an engagement. This is not legal advice and nothing here substitutes for your firm's own ethics analysis.