# Private legal AI: what it should mean

By Josef Long. Published October 17, 2026. General information, not legal advice.

"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 differ, and the questions worth asking a vendor.

## Three tests for any "private" claim

1. Where does the AI actually run? On the firm's own machines or server, or on someone else's.
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, see https://vaultscribe.app/security.md.

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

| Model | Where client data sits | What to check |
|---|---|---|
| Public cloud AI service | On 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-premise | On hardware the firm controls. | Control is high, but patching, backups and keys become the firm's job. On-premise alone says nothing about encryption or logging. |
| Firm-controlled, on-device | On 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. 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.

## 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.

## 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.

**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?** The seven questions above.

## Where VaultScribe fits

VaultScribe is a private AI workspace for law firms built around these tests. Founding design partners sign a non-binding letter of intent, with terms agreed on a call: hello@vaultscribe.app.

Related: https://vaultscribe.app/legal-ai-agent.md and https://vaultscribe.app/local-vs-cloud-meeting-transcription.md