The rules we designed for
Anne is built around the duties that make generative AI hard for law firms: ABA Model Rule 1.6 (confidentiality) and 1.1 cmt. 8 (technological competence), Rule 5.3 (supervision of non-lawyer assistants — where AI sits today), Formal Opinion 512 (you must know where the data goes before using generative AI), and the risk that sending matter content to a third party waives attorney-client privilege or work-product protection.
Fail-closed, not best-effort
In confidential mode, matter work is answered only by a model running on the firm's own hardware. If that local model is offline, Anne refuses to answer rather than falling back to a cloud provider. An error can never become a disclosure. This is enforced in the routing code, not in a policy PDF.
Proof, not promises
Every brain interaction on a matter is recorded in a tamper-evident audit trail: which model answered, whether confidential mode was on, and what retrieval was used. Any matter can produce a privilege report — including a PDF you can hand to a client or carrier — showing per answer whether content stayed on-network. Most vendors promise privacy in a contract; Anne shows you the log.
What leaves the network
| Data | Where it goes |
|---|---|
| Matter content (confidential mode) | Never leaves the firm's network — local model only, fail-closed |
| Matter content (cloud mode) | Only the minimized, relevant excerpt goes to the cloud model — never the full record; every send is audited |
| Citation checking | Citations are extracted locally; only the citation strings (public by nature) reach the research API — never your draft |
| Documents & database | Single-tenant storage on your deployment; offsite backups leave only encrypted client-side, with keys the tenant holds |
| Credentials in output | A redaction layer strips detected credentials before any reply leaves on any channel |
Data handling you control
Each firm sets its own retention window (closed matters and dead leads purge automatically; the audit trail never does), can freeze everything with one legal hold switch, and records its data region. Purges themselves are audited — deletion leaves a paper trail.
Draft work product, always
Every deliverable is stamped draft work product — attorney review required in code, is grounded in the matter record with footnoted sources, is checked by an independent second model before you see it, and has its case citations verified at approval. Anne never signs, files, or advises a client — that is the supervising attorney's role, and the system refuses to cross it.
Questions your ethics counsel wants answered in more depth?