An inbox mixes sales requests, client questions, supplier messages and automated notices. Norml builds an AI email triage workflow around the categories your team uses. We connect the selected mailbox, test the classification rules against representative messages and route the result to a named owner. Unclear messages stay visible for review.
The starting scope uses Gmail, n8n and Claude. We agree which messages the model can read and which label or internal queue each category should produce.
Define categories such as new enquiry, existing client, supplier and needs review. Include representative edge cases and decide whether one message can belong to several categories. Explicit sender or account rules can take precedence over a model suggestion.
Map accepted classifications to Gmail labels and an agreed owner queue. Keep the original message and thread reference available. Classifying a message does not grant permission to forward its contents to a different mailbox or external service.
Configure a visible fallback for unmatched, conflicting or invalid results. Retain the message ID and processing result so another trigger does not repeat the same handoff. Later replies receive their own assessment when the conversation changes.
Identify the mailbox, included folders or labels, account owner and permitted data. Exclude personal mail and attachments unless they are explicitly included. We verify the supported connection before relying on a shared-inbox setup.
Assign someone to ambiguous messages and routing failures. Define urgency using your operating policy. A model-generated confidence number cannot establish reliability without evaluation against labelled examples.
This workflow categorizes and assigns work. Reply drafting, sending, deletion, spam removal and automatic archiving require separate rules. The first release can leave all source messages in place while the team reviews classifications.
Your team categorizes examples, including messages with several requests and categories that are easy to confuse. We use these examples to define expected behavior.
Connect the mailbox and model, map allowed labels, and record the chosen category with its source. Unmatched output must reach review instead of being discarded.
Test repeated events, a new reply, invalid model output and an unavailable destination. Confirm that a reviewer can correct the category and resume the missing handoff.