Restructuring B2B Customer Service for 2 Million Bank Clients
Two million B2B clients reorganised into four defensible classes; ML-based segmentation board-approved.
Theme Customer Experience & Personalization · Also Strategy
In brief
Situation. A leading Latin American bank had two million B2B clients, each generating significant revenue, served through a customer-service structure that had grown organically over decades.
Complication. Service delivery was uneven, costs were rising, and the bank wanted to redesign the structure to match service categories to client characteristics, at scale, defensibly, without losing the high-revenue clients in the noise.
Resolution. I directed the data vertical of the strategic project. We analysed five years of client data, behaviour, attributes, service utilisation. Using factorial analysis and clustering, I identified distinct client segments. Linear optimisation and simulation were used to design a tailored service structure for each segment. The proposed structure was presented to the bank’s board for refinement and validation. Throughout, we worked with bank subject-matter experts to interpret a highly sensitive proprietary data environment.
Impact. Two million clients reorganised into four defined classes with clear assignment rules. Client experience improved; operational efficiency optimised. Demonstrated the application of complex data analysis to high-stakes financial-industry problems, with results that could be defended to a board.
The longer story
Banking has a particular cultural challenge with data science. The institution has had centuries to build internal expertise and political authority around how to manage clients. A data scientist walking in with a clustering algorithm is, structurally, a junior outsider telling the senior insiders that their intuitions are statistically wrong. This goes badly unless handled with care.
The single most important thing I did on this engagement was not the modelling. It was the time I spent with the bank’s subject-matter experts, relationship managers who had served corporate clients for twenty years, getting them to teach me what the data was actually showing.
Their domain knowledge was extraordinary. My job was to formalise their tacit knowledge into something a board could approve. Every cluster the algorithm found, I sense-checked with the relationship managers. If they nodded, the cluster was real. If they frowned, the cluster was a data artefact.
The board got four defensible client classes. The relationship managers got institutional validation of what they had been saying for years. Everyone won. The model was the visible artefact; the partnership between data and domain was the project.