AI Triage Rescues the Operations Inbox

Every operations team has one: the shared inbox that everyone dreads opening. Two hundred emails a day. Supplier queries mixed with client requests mixed with automated notifications mixed with newsletters nobody subscribed to. Someone - usually the most junior person on the team - spends the first 30 minutes of every morning sorting the pile.

The architecture

We built a 4-workflow n8n system: two main workflows - one for hourly categorisation, one for a daily digest - each calling a shared sub-workflow per mailbox. Adding a new mailbox takes minutes, not a rebuild.

Three categories, deliberately simple

• NO_REPLY_NEEDED - newsletters, confirmations, automated notifications. Marked as read automatically.
• NEEDS_REVIEW - forwarded messages, CC'd threads, anything below confidence threshold. Starred and labelled for human review.
• REPLY_WAITING - direct questions, client requests, urgent issues. Starred, labelled, surfaced at the top.

How the AI reads email

The critical design decision was making the AI read the full email thread, not just the latest message. An email that says 'Yes, approved' means nothing without the context of what was requested three messages earlier.

We set the confidence threshold at 85%. Below that, the email defaults to NEEDS_REVIEW. A triage system that miscategorises urgent emails as noise is worse than no system at all.

Deduplication and self-healing

Every processed email is logged to a relational database by message ID. The hourly workflow checks this table first and skips anything already categorised. The system also monitors its own labels - if someone accidentally deletes a Gmail label, the next run recreates it with the correct colour.

The daily digest

Every morning at 9am, the digest workflow summarises everything that arrived overnight: email counts per category, top senders, and draft response suggestions for REPLY_WAITING emails. One email gives the team lead a complete picture before the first coffee is finished.

The result

Thirty minutes of manual sorting became ten minutes of reviewing what the system flagged. A 66% reduction in triage time - from four n8n workflows, an AI model, and a database table. The person who used to sort emails now reviews exceptions and handles the replies that actually matter.