Knowledge base
AI labels
Say what a label is for in plain words, and AI adds it to new mail that fits.
Details
- The label editor has three controls for it. “What is this label for?” takes a description in plain words, such as “invoices and receipts” or “anything from investors”. “Apply to new mail automatically” switches it on, and “Try it on recent mail” shows what it would catch.
- One AI call per arriving message covers every AI label of the workspace and adds the ones that fit. It skips spam, the Bin, encrypted mail and muted conversations. It only adds labels: it never archives or moves mail.
- “Try it on recent mail” reads the newest 20 inbox conversations and lists the ones the label would catch, without labelling any of them. A try spends one AI action of the person who runs it.
- AI labels are part of the paid plans, and a workspace may have up to 10 of them. Each workspace also has a daily cap on the messages they read: 200 on Starter, 1,000 on Business and 5,000 on Enterprise. Labelling arriving mail does not spend your daily AI actions.
- The model reads the sender, the subject and the start of the text, as summaries already do, through routing that denies data collection.
- On the REST API a label carries aiInstructions and aiEnabled, set with POST /labels and PATCH /labels/{id}, and POST /labels/{id}/test tries one. The SDK has the same fields on labels.create and labels.update, plus labels.test. The MCP server has them on createLabel and updateLabel, plus testLabel.