Automated Email Triage
An AI capability that reads inbound customer emails, classifies them, identifies the customer, extracts the details, and decides what can be handled fully automatically — introduced measurement-first.
The challenge
A high-volume customer inbox where a large, repetitive share are "holiday stop" requests — each needing manual reading, a customer lookup and careful date interpretation.
What we built
We built a general, LLM-powered triage framework that classifies the email, matches the customer in the database, computes precise start and return dates from loose wording (e.g. "from next Monday for two weeks"), and decides whether the request is clear and complete enough to automate end-to-end. The AI's judgement is cross-checked by strict rules; anything ambiguous — no customer match, unclear or contradictory dates — is routed to a human. It never guesses on a customer's behalf.
- LLM email classification (is this a holiday stop?)
- Customer identification against the database
- Date extraction from natural language
- Automation-readiness decision, cross-checked by strict rules
- Human-in-the-loop routing for anything uncertain
- Measurement-first "watch & report" rollout — no live impact until switched on
- A reusable framework that extends to cancellations, address changes and delivery issues