Case Study: Rocket Money builds a personal finance agent with Claude

Rocket Money runs Rowan with near-zero hallucinations using Claude

Rocket Money, a personal finance app, wanted to provide its users with an expert-level financial assistant. Their challenge was to build a highly reliable and accurate AI agent that could handle sensitive financial tasks via text message without the hallucinations common in other AI systems. They chose to build this new product, named Rowan, using Anthropic's Claude models and the Claude Platform (API).

The solution implemented using Claude was an innovative "agent-workflow hybrid" architecture. This system uses a classifier to route user requests to specialized subagents, which are powered by different Claude models like Haiku, Sonnet, and Opus, each selected for specific tasks. This approach, combined with structured output and a self-healing automation philosophy, resulted in near-zero hallucinations and 100% accuracy for financial actions. The vendor, Anthropic, provided the reliable and consistent model performance Rocket Money needed to build a trusted consumer AI product in the high-stakes finance domain.

Rocket Money