Intelligent Ticket Routing for a 50-Person Support Org
Led the implementation of an ML-based ticket routing system that reduced misrouting from 35% to 8% and pushed SLA compliance from 72% to 94%. Turned a chaotic support queue into a predictable operation.
Challenge
Support tickets misrouted 35% of the time across a 50-person support org, causing SLA breaches to climb quarter over quarter.
Solution
ML-based classification model combined with a rules engine for auto-routing tickets to the right team and priority level.
Result
Misrouting dropped from 35% to 8%, SLA compliance improved from 72% to 94%, and average resolution time decreased by 28%.
The Problem
At a mid-size fintech company, our customer support organization had grown to 50 agents across five specialized teams — billing, technical, compliance, onboarding, and general inquiries. Tickets came in through email, chat, and an in-app form, and a small triage team manually routed them. The problem was scale: ticket volume had doubled in six months, but the triage team had not grown. Misrouting hit 35%, meaning over a third of tickets bounced between teams before reaching the right person. SLA compliance had dropped to 72% and was trending downward. Customers were frustrated, agents were frustrated, and leadership was asking hard questions.
What We Built
I scoped and led a cross-functional initiative to replace manual triage with an intelligent routing system. The approach had two components. First, an ML classification model trained on 18 months of historical ticket data — roughly 120,000 labeled tickets. The model analyzed ticket text, metadata, and customer attributes to predict the correct team and priority level. Second, a rules engine that handled edge cases: VIP customers, regulatory-sensitive topics, and escalation patterns that required human judgment.
I worked with our data science team on model development and with the support operations lead on defining routing rules and exception handling. We built a confidence threshold into the system — tickets where the model was less than 85% confident were flagged for human review rather than auto-routed. This was critical for building trust with the support team.
Rollout was deliberate. We ran the model in shadow mode for three weeks, comparing its routing decisions against human decisions without actually routing anything. When shadow-mode accuracy hit 91%, we moved to a phased rollout — one team at a time over four weeks.
The Outcome
Within the first full month of production, misrouting dropped from 35% to 8%. SLA compliance climbed from 72% to 94%. Average resolution time decreased by 28% because tickets reached the right team on the first try. The triage team was redeployed to handle complex escalations and quality assurance — work that was far more valuable than manual sorting. The support operations lead told me it was the single biggest improvement to their workflow in three years. Six months later, the model continued to improve as it ingested new data, and misrouting had further dropped to under 5%.