How does private AI handle sensitive data?
Private AI handles sensitive data safely only when the system documents and controls every material route: collection, processing, model calls, retrieval, tools, logs, support, backups, exports, and deletion. Local hardware can reduce exposure, but privacy comes from the complete operating boundary—not the machine’s address.
Reviewed 12 October 2026 · General operating guidance, not legal or compliance advice.
Prove the route before making the privacy claim.
The workload needs an approved route record with a named owner and evidence for each control.
Name the data
Define approved data classes, prohibited material, purpose, authority, and minimum necessary use.
Trace every path
Map prompts, files, embeddings, outputs, tools, telemetry, logs, backups, support, and external calls.
Limit access
Use identity, least privilege, encryption, environment separation, and explicit tool permissions.
Preserve authority
Keep consequential decisions with named people and fail closed when authority or evidence is missing.
Control persistence
Set retention, deletion, cache, export, backup, and recovery rules for each data class.
Test the boundary
Inspect access, routes, exceptions, recovery, and prohibited flows before and during operation.
Four claims that require evidence.
A location label or vendor promise is not enough.
Document the route before selecting the model.
Start with data classes, permitted uses, users, systems, authority, and prohibited flows. Do not upload confidential source material to the assessment.