Banks don't have an AI problem — they have a data platform problem

Summary

The post argues that banks’ difficulty scaling AI stems less from model capability than from fragmented data, weak governance, and limited real-time access. Examples from CBA Live 2026 include model drift in credit scoring, fragmented customer signals, and a collections model that predicted, with 85% accuracy, how many days newly delinquent accounts would take to cure using governed data. It also presents Erica’s 3.2 billion interactions since 2018 as evidence that production AI requires continuous tuning, monitoring, and edge-case management, while generic models decay in frontline settings. The proposed Databricks architecture combines Lakehouse, Unity Catalog, Lakeflow, Lakebase, MLflow, Model Monitoring, Online Feature Store, Genie, and Agent Bricks to support governed analytics, low-latency decisions, auditable models, and constrained agent actions. Its conclusion is that a shared data platform should precede additional AI use cases because it improves deployment speed, trust, explainability, and regulatory defensibility.

Context

The post describes banking institutions struggling to move AI initiatives beyond pilots because customer, risk, and product data are fragmented, governance is limited, and systems lack the real-time consistency required for personalized services, model monitoring, and agentic workflows. It also identifies model drift, explainability, compliance, and vendor capability as practical concerns.

Approach / What changed

The post maps these challenges to a Databricks platform architecture. Lakehouse, Unity Catalog, Lakeflow, and Lakebase provide governed batch, streaming, transactional, and operational data foundations; MLflow and Model Monitoring support model lifecycle management; Online Feature Store and Genie support real-time decisions and governed natural-language access; and Agent Bricks and Databricks Apps support constrained, auditable agent workflows.

Takeaways

• A collections session described predicting how many days a newly delinquent account would take to cure with 85% accuracy, using account, digital engagement, credit bureau, deposit, and historical resolution data in a governed way.
• Erica has handled over 3.2 billion customer interactions since launching in 2018, and its production operation requires continuous data tuning, monitoring, and review of edge cases rather than a set-it-and-forget-it model.
• A speaker reported that only around 5% of several thousand vendors in a large-scale audit had genuine AI in their products, with others relabeling robotic process automation or standard automation logic.

Related reading