Reference implementation In production since September 2025

Priority inbox triage

An agent that ranks email threads by urgency, verifies each against project tools, and presents only the items that require action.

~90 → ~8 min daily review time
568 → 7 messages to priority items on a typical day
7-9 → 2-3 false urgent flags a day after cross-checking
~$1.20 monthly model cost

Context

Three mailboxes produced several hundred messages a day. Manual review took about 90 minutes and still missed time-sensitive threads.

Approach

Every thread defaults to "no action". The classifier must justify promotion to P0, P1 or P2, using only the latest message, and is cross-checked against ticket status and sent-mail history before it reaches the user.

Architecture

  • Mail ingestion on a scheduled poller
  • Classification service with an evaluation set
  • Postgres record of threads, decisions and actions
  • Native desktop review surface
  • Private endpoint behind a tunnel with token authentication

Controls

  • Read-only mailbox scope
  • No automatic sending
  • Every classification stored with its reason
  • Subset testing before any prompt change

Result

Daily review fell from about 90 to about 8 minutes after four weeks of iteration, with false urgent flags reduced by roughly two thirds.

What this means for you

The same pattern applies to shared inboxes, support queues and approval workflows. See Custom AI Tools.

Start here

Have a problem AI might solve?

Describe the outcome you need. Within two working days you receive a written view on whether AI is the right tool, the risks to manage, and a proposed first step.