Problem
A support organization routes each ticket to one expert, and the usual rule minimizes waiting time within hard topic, channel and language constraints. Fast and effective service often collide: a customer who waits little but is then escalated to a higher tier is worse off. Routing must also keep workload fair, or the best experts are overloaded while others sit idle. The team could not A/B test a new router, and logs only show what the current router chose.
Example from the article: Wix has thousands of experts, 10 languages, three channels and dozens of topics, and the old waiting-time-first routing in callbacks was replaced.
Solution in brief
Treat routing as a recommendation problem solved with reinforcement learning (RL). A model returns the top K experts per ticket and is trained and evaluated in a simulator, with a reward built from five KPIs and set per channel. The simulator replaces the A/B test and lets the model try actions the logs never contain. The article does not disclose the RL algorithm.
Questions this case answers
- How to assign incoming requests to a changing pool of heterogeneous workers when speed, quality and workload balance conflict and the new policy cannot be A/B tested
- How to evaluate and train a sequential decision policy when logged data reflects only the current policy and online exploration is unsafe
- How to route support tickets to experts when waiting time, resolution quality, expert tiers and workload pull in different directions?
- How to evaluate a routing policy when an A/B test is impossible and logged data comes from the current policy?
Tools and methods
- Reinforcement learning (RL) — the agent picks an expert for each ticket given the system state and learns from the reward.
- Simulator of the care system — trains the model, evaluates it before production, and serves as a playground for later changes.
- Five-KPI reward — waiting time, abandonment, resolution, tier matching and expert occupancy, scored per outcome.
Google or a one-time email code. No password.