Skip to main content

Sender Smart Categories: Reuse and Subject Checks

Sender Smart Categories can reuse a saved sender-to-category assignment instead of asking the model again. This reuse has eligibility conditions; it does not mean all future messages from an address have the same purpose.

Which categories can be cached?

The current processor permits categories whose primary action is Move or Delete. Keep categories are not eligible for cached routing. A saved assignment that the AI learned still needs the subject guard to pass before reuse.

The guard normalizes subject words, ignores numbers and common reply prefixes, and compares learned patterns. An unrelated subject can cause individual classification. An empty or digits-only subject, or missing stored patterns, instead permits a match fallback. This is a heuristic, not a guarantee that changed message types are detected.

A worked example: one sender, two kinds of mail

Take a fictional hosting provider that writes from billing@lumenhost.example. The mailbox has a category “Receipts” with the primary action Move to folder.

  1. “Your invoice 10432 for March” arrives. The sender is new, so the model is asked which category fits this sender. It sees the address, the display name, the subject and a short preview of the text, and answers “Receipts”. Because Receipts moves mail, the assignment is saved together with the words of that subject.
  2. A month later “Your invoice 10518 for April” arrives. Numbers are ignored, and at least half of the remaining words match the stored pattern. The saved assignment is reused without a model call, and the message is moved.
  3. Then “Spring sale: 30% off hosting” arrives from the same address. No word matches a stored pattern, so this one message is classified individually against all enabled categories. If the model picks “Marketing”, the message follows the actions of Marketing. The saved Receipts assignment stays unchanged.

When an individual check confirms the saved category, the new subject is added to the sender’s patterns. Up to ten patterns are kept per sender; the oldest is dropped first.

Where to find it in the app

  1. Open the mailbox and go to Rules. In the Auto-Processing card, the checkbox Check every email individually controls the feature. It is unchecked by default, which means senders are classified once and the result is reused. If you check it, the model reads every message and assignments learned by the AI are skipped.
  2. Open Senders. Each row shows the address, the number of emails, a source label and a category selector. The source is AI when the model decided, Pattern when the address matched a built-in pattern, and You when you set it.
  3. To correct a sender, choose another category in the selector. The source changes to You.

What a rule you set yourself does differently

A choice you make in Senders is treated as an instruction, not as a learned guess. It applies to every later message from that exact address without the subject check, it also works for Keep categories, and it still applies when Check every email individually is on. Choosing Uncategorized for a sender keeps its mail in the inbox and stops the model from deciding again for that address.

One exception: a disabled category never acts. Mail from a sender assigned to a disabled category is classified individually among the enabled categories until you enable it again.

Review mixed-purpose senders

Test an invoice and a promotional message from the same fictional address. Inspect actual categories before applying moves. Review existing saved assignments as well as the setting for automatic categorization of new senders; do not treat a setting change alone as proof that every stored decision is bypassed.

The direct subject-guard checks show the observed code behavior with synthetic inputs. They are not a model accuracy benchmark or a live mailbox test.

Common questions

Does a saved assignment cover a whole domain?

No. The match is on the exact sender address, separately for each connected mailbox. billing@ and news@ at the same domain are two senders with two assignments.

Why does the model still read every message from some senders?

Assignments to Keep categories are not saved by the AI, so senders whose mail stays in the inbox are classified message by message. If you change a category’s primary action to Keep in inbox, a learned assignment to it is dropped the next time that sender writes.

What does the source “Pattern” mean?

If the mailbox has an enabled category named exactly “Newsletter”, a new sender whose address begins with a term such as noreply, no-reply, newsletter, marketing, promo, news or updates before the @ is assigned to it without a model call. Without a category of that name the pattern is not used.

Is the saved assignment used when I classify manually?

No. Classify All in the Inbox tab reads each selected message individually. When you apply the results, senders of messages in Move or Delete categories are recorded in Senders. A rule you set yourself is not overwritten.

Keep actions reviewable

Start with a dedicated folder and inspect new subjects. A cached assignment cannot correct an unclear category prompt. See multi-action controls and History and Undo for the recovery limits.

Ready to put your inbox on autopilot?

Get Started Free