Where AI actually belongs in a core banking program

Every banking RFP now asks about AI. The useful answer is rarely ‘we will put a model on the core.’ The ledger still needs to post, reconcile, and survive an examination. Intelligence belongs at the edges, in the work that currently burns people, not in the system of record.

Three places consistently pay for themselves. First, exception handling: payment repairs, KYC refresh, and ticket classification. Second, servicing: next-best action that is constrained by product rules, not a chatbot that invents offers. Third, risk operations: alert ranking so analysts are not drowning in false positives.

What does not belong in the first year is a generative layer that writes to customer balances or product terms. If you cannot explain a change to internal audit in one page, it should not touch the books.

Start with a single operational queue, a human in the loop, and a kill switch. Measure handle time and error rate for 90 days. Only then talk about expanding. Banks that skip this sequence end up with a demo and a risk committee that will not let it into production.