Personalised upsell
Personalized upsell for each bank customer improved conversion rate in-branch and over call center to >4% from previous 2.3%.*
*Estimated by monthly tracking of customer interactions
What the business got
Challenge
Bank was sitting on a large internal data set from marketing collection that was used manually by analysts to design next offers in a general sense.
These offers conversion rates hovered under or at industry benchmark 1-2%, yet costed a lot of work to create them.
Strategy
With most banking upsell happening in-person through branch or customer center calls, but the bank increasingly digitalized channels too, increase of the conversion rate just by 1-2% offered a great opportunity.
Bank chose to build a new ML model to personalize upsell and needed effective feedback loops to be integrated throughout their core banking user interfaces along the mobile apps.
Our delivery
We built the new event channels extremely fast reusing the choreographed micro-services architecture we delivered half a year earlier.
Then created a new user interface in React specifically to integrate into the existing branch and call center workflows.
With a temporary help of a core banking engineer, we added a new web browser stage into FlexCube, presenting our model's personalized upsell for the currently served customer, right during the customer exchange happening, giving the best chance for the employee to engage and close a deal.
Country-wide roll-out of the new functionality under 1 month, with one no-incident roll-back localized to a single employee through feature flags avoided any outage typically part of a core modernization.
Outcome
As the model was fine-tuned over the quarter after our delivery, the conversion rate shot up above 4%.
The advanced analytics lab delivered company wide dashboards for executive decision making and continued improving the feedback loops for 5 years after this success.
Client testimonial
"Sandor's team cut through the unnecessary and delivered exactly what moved the needle."
— lab sponsor.
The team setup
Scope
Build · high effort
Integrate · moderate effort
Reuse · least effort
Arch. & CI/CD · -3 months
Squad
Product owner and tech lead
5 engineers
1 flexCube engineer
4 data scientists
Whose people
ours
the client's
The container architecture
Advanced Analytics built on containerized micro-services to abstract away the external world's legacy interfaces, to enable fast development encapsulated and allowed us to fully automate the developer experience.
Adapter services pattern separate the effects and the two different technical worlds, including the FlexCube core in this case.
Choreography based pub-sub: entity property changes trigger any reaction from independent worker services, reacting only on properties each one interested in. A wave of reactions triggered as a cost of total decoupling of workers into independently maintained and released components.
App State Manager: the only high-speed single instance service (per key), that merges the states and writes to the loan database by append only (each record is a history of change too for regulatory reports).
Adapter services: one side they keep the architecture decisions of this system, on their other side everything allowed to make it work with the designated external integration. They dampen external system changes and enables us to end-to-end auto-test even without 3rd party support over core systems.
The delivery calendar
FlexCube deployment went without incidents, we had to turn off a feature flag temporarily for an integration fix, never risked the day to day operations.