AI Banking & Financial Services Agent

Description

A regional bank, so start from the constraint: nothing we build may give a customer a number we cannot trace back to a system of record, and nothing may leave our tenancy. Within that, the target is a customer-facing assistant covering balance and transaction questions, card blocking, and standing-instruction changes. Anything touching a limit, a loan or a dispute goes to a human, with the transcript and the customer already authenticated so the agent does not ask them to start again. You will work against a sandbox core-banking API. Expect it to be slower and less forgiving than the documentation suggests, and budget for that. We would rather see a narrow assistant that is right every time than a broad one we have to apologise for.

• Implement the balance, transaction and card-blocking intents against the sandbox core-banking API
• Enforce a strict allow-list: anything outside it escalates rather than improvises
• Carry authenticated session and transcript into the human handover with no re-authentication
• Log every response with the source record that produced each figure, retained for audit
• Handle core-banking latency and outage with explicit customer-facing messaging, never a stale figure

• Assistant deployed in our tenancy, integrated with the sandbox then production core-banking API
• Intent allow-list and escalation matrix, signed off by our compliance team
• Traceability log linking every stated figure to its source record
• Failure-mode documentation and rollback plan