Private legal AI: what it should mean

"Private" is easy to put on a web page and hard to build. Here are three tests for any privacy claim, how the common deployment models really differ, and the questions worth asking before a firm trusts a vendor with client material.

Josef LongOctober 17, 20266 min readGeneral information, not legal advice

Three tests for any "private" claim

Marketing pages use the word freely. These three questions cut through it.

  1. Where does the AI actually run? On the firm's own machines or server, or on someone else's. This decides who can reach the data at all.
  2. Who holds the keys? Data that is encrypted but decrypted by the vendor is only as private as the vendor's staff and systems.
  3. What can leave, and can you see it? A product can run locally and still send data out. Ask for a log of every outside request.

For how VaultScribe answers these tests, in full, see the security and architecture page.

Cloud, private cloud, on-premise and firm-controlled

People use these terms loosely, so here is how they usually differ for a law firm. These are general patterns. Any one product may differ, which is why the questions below matter.

ModelWhere client data sitsWhat to check
Public cloud AI serviceOn the vendor's shared servers.Who at the vendor can read it, whether it trains models, how long it is kept.
Vendor "private cloud"On the vendor's servers, in an environment set aside for the firm.Isolation from other customers is a real benefit. The vendor still hosts the data, so ask who holds the keys and who can reach the environment.
On-premiseOn hardware the firm controls.Control is high, but so is the work: patching, backups and keys become the firm's job. On-premise alone says nothing about encryption or logging.
Firm-controlled, on-deviceOn each lawyer's own computer and, optionally, a firm server.The firm holds the keys and sees the log. This is the model VaultScribe is built around, and it needs capable hardware and an owner for IT.

Where each model runs

A private tool should let the firm see which AI model handles which task. VaultScribe lists the models on each machine by tier, the firm's own AI server, and any cloud provider. Cloud providers are off by default and each one appears in the log of outside requests. Routing rules decide where a privileged matter may go.

Design preview: models and providers, showing models on each machine by tier, the organization's AI server, approved cloud providers that are off by default, and the order of the automatic routing rules
Models and providers: what runs on each machine, on the firm's server, and with any approved cloud provider, and the routing rules in order. Design preview. Sample data.

Seven questions to ask any private legal AI vendor

  1. Where does inference run, and is any of it on your servers?
  2. Who holds the encryption keys, and can you recover without the vendor?
  3. What leaves the firm, and is every outside request in a log you can read?
  4. How does deletion work, including backups and search indexes?
  5. Is our content used to train any model, yours or anyone's?
  6. Can we check the claims ourselves, for example with our own firewall?
  7. What happens to our data if we stop paying? You want to keep reading and exporting what you already have. VaultScribe is designed that way: if a license lapses, existing recordings stay readable and exportable, and new capture and AI features stop after a grace period. The license is checked offline, so nothing phones home to confirm it.

Questions

What is private legal AI?

AI for legal work that runs inside the firm's own environment, under the firm's own keys and access rules, with no vendor holding client data by default.

Is private AI the same as on-premise AI?

Not exactly. On-premise means the software runs on hardware the firm controls. Private is the larger claim: who holds the keys, what can leave, whether there is an audit trail, and how data is deleted. On-premise is usually part of private, but it is not enough alone.

Is a vendor's private cloud actually private?

It can be isolated from other customers, which is a real benefit. But the vendor still hosts the data, so ask who holds the keys, who at the vendor can reach the environment, and what is logged.

What should a law firm ask a private AI vendor?

Where the AI runs, who holds the keys, what can leave and whether it is logged, how deletion works, whether content trains models, whether you can check the claims yourself, and what happens to your data if you stop paying.

Where VaultScribe fits

VaultScribe is a private AI workspace for law firms built around these tests. I am looking for founding design partners, with a non-binding letter of intent and terms agreed on a call.

Read the security and architecture page Become a design partner

Related: Legal AI agent for law firms and Local vs cloud meeting transcription. General information, not legal advice.