AI for Customer Care Routing at Wix - Making The First Step of Reinforcement Learning

Ofir Magdaci19 Nov 2024

Problem

Routing tickets to experts by waiting time alone ignores other drivers of satisfaction: tier matching, expert occupancy, abandonment and the chance the ticket gets resolved. Wix wanted a policy that balances five KPIs. A full reinforcement learning (RL) model would take long, carry high risk and offer an unknown gain over a simpler one. Operators and managers also had to trust it. Wix could not run an A/B test, only a full exposure test followed by release.

Example from the article: an idle top-tier expert is free and a low-tier customer is waiting. Should the ticket go to this expert or wait for a lower-tier one?

Solution in brief

Wix first built a one-step "Greedy" policy. For each ticket it estimates the expected reward of routing it to every candidate expert and picks the highest. A validated simulator evaluates the policy offline before a two-week live test. The simulator, reward and data were refined on this simple model before RL.

Questions this case answers
  • How to assign arriving jobs to heterogeneous workers so that several service KPIs are balanced rather than only waiting time minimized
  • How to validate a decision policy offline with a simulator when randomized online experiments are not allowed
  • How to route support tickets to experts when waiting time is not the only thing that matters?
  • How to ship a useful, explainable routing policy before a full reinforcement learning model is ready?

Tools and methods

  • Greedy one-step policy — scores each expert by expected reward for the current ticket and state
  • KPI estimators — exponential waiting-time model, linear regression for abandonment, supervised model for resolution
  • Queue simulator — replays tickets against a policy and aggregates KPIs and reward daily
  • Re-routing of waiting tickets — reassigns a queued ticket when an expert becomes idle
Continue reading — it's free

Google or a one-time email code. No password.