Chase’s first customer-facing AI agent, live on voice and RCS
Chase wanted an AI agent that could talk to customers without creating legal risk. We started in home lending, where the rules are strictest and the loans are largest.
I designed Casey as one agent across voice and RCS. It qualifies customers, tells them where their application stands, and hands advice and quotes to a licensed associate.
Casey Voice has handled 3,000+ production calls at about 12% lead conversion.
Customers stalled halfway through mortgage applications and rarely came back.
Chase needed a way to reach those customers by phone and text. Any agent doing that speaks for the bank.
Three rules shaped everything: first, Casey cannot give advice or quote rates; second, every call and text needs consent and a disclosure; and third, voice and text follow the same rules.
In home lending, one wrong answer about a rate is a compliance issue.
If Casey quoted a rate or gave advice, the bank would be on the hook.
We started in home lending because it has the most regulation and the largest loans in consumer banking. An agent that works there can be reused across the bank.
Most of the design work defined when Casey hands a customer to a person.
One agent across voice and text
I designed Casey as one system with one set of rules: what it knows, what it may say, and when it hands off. Home lending was the first place we shipped it.
Instead of Separate voice and text bots owned by different teams.
Trade-off The first launch took longer, and I spent time aligning teams that would not use it for months.
Hand off early, with the conversation attached
Questions beyond Casey’s authority, signs of stress, and repeated confusion all go to a licensed associate. The transcript goes with the transfer, so the customer does not repeat anything.
Instead of Keep as many calls as possible away from associates.
Trade-off More conversations end with a person, which hurts a containment metric. In lending, that handoff is often where the lead converts.
Casey Voice qualifies callers, tells them where their application stands, and books time. Advice and quotes go to a licensed associate. Text customers who leave an application
When a customer saves an application and leaves, Casey sends an RCS message with a link back to the same step and a way to reach a person.
Instead of Email reminders or an unbranded SMS short code.
Trade-off Outbound messages need stricter consent and careful timing. A badly timed text feels like pressure.
A verified sender, one resume link, and a way to reach an associate. RCS shows the Chase brand, which plain SMS cannot.
Casey hands off when a question is out of scope or the customer is struggling.
Before each release, I ran edge-case QA against the guardrails with engineering. Every flow had to handle these states.
- GreetingSays it is Chase’s AI assistant and what it can help with.
- In scopeQualifies the customer, explains their status, and books time.
- Repeated confusionStops and offers a person.
- Out of scopeDeclines advice and quotes, and says who can help.
- Stress detectedTransfers to an associate right away.
- Human handoffTransfers with the transcript and intent attached.
Both channels shipped, and new Chase agents reuse the handoff rules.
- 3,000+Calls initiatedCasey Voice, production · measured
- ~12%Lead conversion from those callsCasey Voice, production · measured
- ~10kRCS messages sentCasey RCS, since launch · estimated
Chase trademarked the Casey voice. The handoff rules are now the reference for agentic work across the business, so new agents start with limits that already passed review in home lending.
I would bring legal into prototype reviews from the first week.
Their questions changed the flows more than any usability finding did. Next, the same handoff model goes to servicing, for customers who already have a loan.