# An operations copilot that triages 2,400 tickets a day - Industry: Logistics - Engagement: 4 month engagement - Summary: We rebuilt their ticket pipeline first, then layered routing and reply drafting on top. Mean handle time dropped 41% without changing headcount. - Industry: Logistics - Duration: 4 months - Outcome: 41% drop in MHT --- ## Where they were A regional logistics operator was processing roughly 2,400 inbound support tickets per day across email, web form, and an aging Zendesk instance. Routing was manual: a rotating triage team read every ticket and assigned it to one of 14 specialist queues. Average mean handle time was 11 minutes; backlogs grew on weekends. Their first instinct was to plug an LLM into the reply box. We pushed back. The routing layer was already wrong about ~18% of tickets; layering generation on top of that would have produced confident, well-written, wrong answers — at scale. ## What we changed Before any model touched a ticket, we rebuilt the pipeline so that routing was deterministic and testable. We replaced a 600-line if/else chain with a small set of rules, a labelled evaluation set of 3,000 historical tickets, and a CI job that flagged regressions. Routing accuracy went from 82% to 96% on the eval set — and that was the foundation everything else stood on. Only then did we introduce the model layer: classification refinements for the long tail, structured extraction of customer intent, and reply drafting for the high-volume categories. The drafts were never auto-sent; an agent always approved before send, which gave us a continuous stream of supervision signal. ## What it costs to maintain The eval set is the artifact. New ticket categories show up monthly, and the team adds labelled examples to keep the regression net tight. The model side is intentionally boring: one provider, one prompt template per category, swappable. ## Outcome - Mean handle time dropped from 11 minutes to 6:30, a 41% reduction - Routing accuracy moved from 82% to 96% - Headcount unchanged; the team absorbed a 30% volume increase the following quarter without expanding The team that previously triaged now reviews drafts. The work is more interesting and the SLAs are tighter. --- ## Navigation - [Back to case studies](/case-studies) - [Home](/) - Markdown for agents: