Confidentiality is the product.

All data is encrypted at rest with AES-256 and in transit with TLS 1.3. Keys are managed in a hardware-backed KMS, rotated on a defined schedule, and segmented per tenant — the key used to encrypt one firm’s data cannot be used to read another’s.

Backups inherit the same controls as primary storage. Snapshots are encrypted and access-logged.

Every firm is a distinct tenant with isolated storage, isolated document indices, and isolated retrieval contexts. No query issued by one firm can reach another firm’s data. There is no shared vector store. There is no cross-tenant retrieval.

We do not train models on firm data. Not base models. Not fine-tuned ones. Not "anonymized" ones. The contents of your matters stay inside your tenant and inside the inference context of your own requests.

Case Watch recommends. It never acts. Every calendared date is presented before it’s set. Every draft is held in a queue, unsent, until the attorney reviews it. Nothing leaves the firm’s environment through Case Watch — no auto-filed motion, no auto-sent email, no auto-signed anything.

This isn't just a policy. It's enforced in the architecture. There is no code path that sends a document to a court without a signed attorney confirmation.

Single sign-on via Microsoft Entra ID is the default. Roles follow the firm: partner, associate, paralegal, legal assistant, external co-counsel. Access is scoped by matter, not just by user. A paralegal who doesn't work on a matter does not see its documents.

Every read, every edit, every draft generation, every model call is written to an immutable audit log available to the firm's admin in real time and exportable on demand.

We use large language models to read documents, extract structured fields, and draft routine language. We chose vendors whose enterprise contracts prohibit training on input and guarantee no human review of customer data without explicit consent. We inspect those contracts carefully. We can share the list on request.

Models make mistakes. That is why the attorney reviews everything, that is why low-confidence extractions are flagged rather than filed, and that is why drafts sit in a review queue. The system is designed around the assumption that the model will be wrong sometimes — not around the wish that it won't.