PyData London 2023

Causal modelling of agent-customer pairing outcomes to optimise call centre performance
06-03, 10:15–10:55 (Europe/London), Warwick

Large scale call centres are the frontline of customer experience across many industries. Optimizing their operations is crucial for achieving better customer service. We model agent customer pairing as a “talent” allocation problem. In this talk, we discuss how we used uplift modelling to provide real time agent-customer pairings that drive a positive lift in overall interaction score (which can come from any arbitrary scoring function). We discuss the challenges of developing and deploying such models to make real-time interventions in call centres. Similar approaches can be used to drive uplift of any important business KPI with respect to an allocation decision.

This presentation is aimed for data science practitioners and data leaders. We start by introducing the theory of uplift models and provide context by giving examples of common use cases. We then proceed to take listeners through our story of optimising customer-agent pairings in large-scale call centres.

This story begins by introducing the business case. We then move on to the deeper technical detail of how we used uplift modelling to identify and address causal relationships between allocation decisions and the outcome of the decisions as measured by an interaction scoring function (which can be any arbitrary function of the measurable variables of the interaction). Attention is given to data requirements, modelling approaches and benchmark methodologies.

With the uplift model built, we then talk about the other half of value delivery: getting the model into production. Attention is given to overall architecture of data and model pipelines, system integrations and monitoring, as well as managing the constraints imposed by the call cent operational landscape.

Finally, we close our presentation with some general thoughts on how such approaches can be used drive value of any business KPIs with respect to an allocation decision.

High level talk outline:

  1. Causal inference
    - What data do we need to make meaningful inferences?
    - How can we use statistics and open-source packages to model uplift?

  2. Call centre operation and domain knowledge
    - What are the key operational metrics of a call centre that a real time ML based system must keep stable?
    - Why close collaboration with operations matter.

  3. Deploying ML models in legacy call centre systems and in cloud
    - Our technical solution in cloud, from BigQuery and DBT data pipelines, to Vertex AI, Kubernetes and CloudRun.
    - Technical integration with a CISCO call centre and some lessons learned.

Prior Knowledge Expected

No previous knowledge expected

Lead Data Scientist in Virgin Media O2. Petros is an ex astrophysicist who has been modelling solar eruptions as a postdoctoral researcher and Lectures at the University of St Andrews. Following his academic posts, he has applied machine learning in financial documents, genetic data and telecommunications. Petros is passionate about using data science and machine learning to model data, derive insights and productionise solutions.