From 3 Days to 3 Hours: AI Routing in Practice

Starting Point: Ticket Routing Chaos

The company processed approximately 800 incoming requests per week—from billing questions to technical issues and contract renewal requests. All of them were sent to a single support@ inbox, from where the on-duty manager manually assigned them to teams. The average first response time was three business days.

The problem wasn't a lack of personnel—the 14-person support team was physically overwhelmed. The problem was the system: each request required a human classifier, who spent 5 to 20 minutes just determining who to route it to. This consumed over 60 hours of team time weekly.

AI Solution Architecture

We implemented a three-tier classification system based on a language model. The first tier identified the request type from 12 categories with 96% accuracy. The second tier assessed urgency and priority based on sentiment and keywords. The third tier automatically assigned a resource from among the currently available specialists.

The key decision was to abandon rigid rules in favor of a trained model. Rules are fragile: one non-standard query breaks the logic. The model coped with formulations written by a person in a hurry, with spelling errors, and with a mixture of languages.

Results of the First 90 Days

Three months after launch, the average time to first response dropped from 72 hours to 4.3 hours—a 94% reduction. Routing accuracy was 96.2%, better than that of a human dispatcher (historically, around 89%).

The results were particularly significant for VIP clients: the system learned to recognize them by domain and account size, automatically assigning the highest priority. Response time for this segment dropped from 48 hours to 1.5 hours.

Lessons and Pitfalls

The main lesson: don't try to automate everything at once. We started with classification and routing, leaving the final response to the client to human agents. This maintained the quality of communication and gave the team time to trust the system.

A pitfall we didn't foresee: during the first two weeks, employees sabotaged the system—not maliciously, but simply by opening the shared inbox out of habit. The solution turned out to be organizational, not technical: we closed direct access to incoming calls and left only personal queues.