BankUnited

BankUnited, a regional bank headquartered in Miami Lakes, Florida serving SMB customers across Florida, New York, and Texas, faced a persistent operational challenge: frontline employees could not quickly retrieve accurate answers from a policy and procedures library spanning approximately 400 documents. In community and regional banking, where relationship quality and response speed directly influence SMB retention, this friction had measurable consequences — extended call handle times, inconsistent guidance to customers, and heavy dependency on a dedicated support team to field routine procedural questions. The status quo created bottleneck support costs and undermined the adviser role that differentiates regional banks from larger institutions.

BankUnited developed SAVI (a generative AI assistant) on Amazon Bedrock using Anthropic's Claude 2, a large language model, paired with Amazon Kendra for intelligent enterprise search across the 400-document policy corpus. Employees interact with SAVI through natural-language queries — no Boolean search syntax or document navigation required — and receive grounded, policy-accurate answers in under 10 seconds. The retrieval-augmented generation (RAG) architecture grounds Claude 2's responses in BankUnited's own documentation, controlling hallucination risk in a regulated banking environment. The AWS-native stack allowed BankUnited's BI and analytics team to focus on application functionality rather than infrastructure integration, accelerating time to deployment and enabling a shift to a 24/7 self-service support model at a fraction of the prior staffing cost.

SAVI delivered measurable gains across speed, accuracy, and support economics from deployment:

• 95% accuracy rate on policy-related queries, verified against BankUnited's documented procedures
• Sub-10-second response times, enabling employees to get answers while actively on a call with a client
• 24/7 support availability established without proportional headcount increase
• Reduced inbound call and email volume for procedural questions as frontline staff self-served through SAVI
• Shortened employee training lifecycle, with new hires able to look up unfamiliar procedures on demand

Qualitatively, the shift freed human advisers to focus on client relationships rather than documentation lookup, reinforcing BankUnited's service-oriented positioning in the SMB segment.

• A RAG architecture anchored to internal documentation is a low-risk entry point for generative AI in regulated industries — the model answers from your policies, not general training data.
• Winning a focused internal use case (policy lookup) builds organizational trust and creates a documented proof point for expanding AI to customer-facing applications.
• Reducing training burden matters as much as reducing call time — AI knowledge assistants can compress onboarding for new branch staff.
• Moving to 24/7 self-service for procedural queries is only viable when accuracy is demonstrably high; establish a verification benchmark before decommissioning human support coverage.
• Build an explicit roadmap (BankUnited used "OpportunitiAI") to channel momentum from a successful pilot into broader enterprise adoption.