The operator received more than 25,000 customer service emails a month. Engineers spent time triaging tickets that were not real incidents, classification errors sent work to the wrong teams, and troubleshooting took too long.
What we built- A no-incident detection module filters out the roughly 40% of tickets that need no engineering work, using time-series analysis, a rules engine and decision trees.
- A root-cause localisation module learns from historical tickets to suggest likely causes.
- Classification models read incoming emails and tickets and route them to the right engineer.
- Known issues are resolved automatically.
- Misallocated engineering effort fell sharply.
- Root-cause analysis that engineers once did by hand is now automated.
- Net promoter scores rose as customers got faster answers.