How Litco holds your client’s documents.
This page describes what the software enforces, and where, rather than listing badges. After 2026’s privilege fights over vendor-held client data, “your documents never train anyone’s model and stay on infrastructure your firm controls” belongs in an engagement letter, and every claim behind it belongs on this page.
No client document ever trains a model. Cloud inference runs under zero-data-retention terms, so the provider keeps nothing after it processes a request. A firm can also keep a matter off the cloud entirely and run it on Litco’s own hardware.
One instance per firm, one encryption key per matter.
One deployment per firm
Litco deploys as a firm-controlled instance: your firm’s documents stay in your firm’s deployment, and the Matter Agent working your cases has no path into anyone else’s. Nothing your firm uploads is used to train any model, ours or a vendor’s.
Envelope encryption, per matter
Documents are encrypted in transit and at rest. Each matter’s documents are encrypted with that matter’s own data key, which is itself encrypted by a firm-level key. Compromising a stored object gets an attacker ciphertext for one matter, and hosted files stay encrypted in object storage.
Row security in the database
Postgres row-level security is forced on 192 tables, and the application connects as a role verified unable to bypass row-level security, so a mistake in application code hits a wall at the database. Privileged system paths exist for the jobs that need them, and those paths are reviewed as an audit surface.
Matter-based access control
Seats, roles, and matter membership gate every surface, including the Matter Agent’s. Per-user workspaces stay private to their user. Ethical-wall separations are enforced where documents are read, and again where they are shared.
Drafting runs in a sealed sandbox.
The Matter Agent drafts documents inside a sealed sandbox with no network connection—not a restricted connection, none. The drafting code runs as an unprivileged user, reads a fixed set of tools, and writes to a single folder that Litco erases when the run ends, whether the run succeeded or failed. If the sandbox cannot establish its isolation, the sandbox refuses to start.
Before producing documents, spending money, or changing anything that cannot be undone, the Matter Agent stops and asks a person.
Choose where inference happens, per matter.
Every matter chooses where its inference runs, whether a cloud provider under your firm’s own API keys, Litco’s platform keys, or Litco’s own hardware running an open-weight model with no third-party inference at all. Firms with strict obligations can keep a matter’s AI work off the public cloud entirely.
Litco measures every option before routing work there. LitigationBench publishes how each model performs on litigation tasks, including the self-hosted option, and the routing defaults inside the product follow those numbers.
Sharing with co-counsel and clients, governed.
External sharing runs through LitSpace links protected by passwords or one-time email codes, with watermarking on the way out and a record of what left. External recipients cost nothing, so there is no reason to share files outside the platform.
Litco keeps the full step record of every Matter Agent run, and the database records matter activity and import history append-only.
For security questionnaires and diligence reviews, write to hello@litco.ai.