Every sentence an AI says to a customer is a sentence the bank said. Voice and text are also the two channels federal law watches most closely, so the real problem was not building an agent that talks well. It was building one that knows what it is not allowed to say.

We started with mortgage because it has the most regulation, the biggest dollar amounts, and the longest customer journey in consumer banking. Anything that survives there works anywhere. So I designed Casey as a reusable conversational system rather than a mortgage feature: one agent, two surfaces, and a shared model of what it knows, what it may say, and when it steps aside.

Casey can qualify, explain, and schedule. It cannot advise or quote, because that requires a licensed human, so the work went into making that wall feel like a door. Any question beyond its authority, any sign of customer stress, and any repeated confusion routes to a person with the full conversation attached, and that routing logic is the piece every future Casey inherits.

Casey Voice shipped July 2025. Casey RCS shipped May 2026 on iOS and Android. Chase trademarked the voice, and the conversational patterns are now the reference model for agentic work across the business.

Chase needed a customer-facing AI agent that could not create legal exposure every time it opened its mouth. I designed Casey as a reusable system across voice and text, proved it in home lending because that is the hardest environment in the bank, and spent most of the project on the boundary where an AI hands a person to a licensed human. Both surfaces shipped, Chase trademarked the voice, and the pattern is what the next agents are built on.