Private AI Email Sorting: Processing and Fallback

Private email sorting depends on the complete data flow, not just where the mailbox lives or the name of your model. The hosted application accesses message data to perform its work.

Which data is used?

The per-message classifier constructs its input from sender, subject and up to 1,000 characters of the body preview. Sender categorization uses sampled previews up to 500 characters each. These source-code limits do not establish the full footprint of every operation; fetched messages and forwarding have different processing needs.

History stores metadata and action information rather than a complete copy of the email body. Subjects can themselves include personal or confidential information. Evaluate logs, backups and provider retention for the deployment you use.

Does an own model keep everything local?

No. It does not relocate the hosted application. A custom endpoint can also have a configured default-provider fallback, so a failure can result in another model provider being tried. Production settings currently reject private and loopback endpoint addresses.

For strict local processing, an administrator must verify the application deployment, model choice, network policy and complete provider chain. Test normal operation and an outage with synthetic data; a local URL alone is not evidence of local-only inference.

Forwarding and recovery are separate decisions

Forwarding with a summary attaches the original message as well. Make sure the recipient is authorized to receive it. Undoing a folder move does not recall that copy. Permanent deletion cannot be undone through History.

Use exact native rules where they suffice, or preview a narrowly defined category without confidential messages. Keep matching, excluded and ambiguous examples in your review set. Do not assume automatic knowledge of your contacts or purchases.

See model settings, forwarding and the data-flow review. This page describes product behavior, not a legal compliance assessment.