How Rocket Money built its personal finance agent with Claude

Overview

Rocket Money, a personal finance app used by millions, built Rowan, a financial agent reachable by text message, on Claude via Amazon Bedrock. Rowan routes each request through a Haiku 4.5 classifier to single-responsibility subagents, with Sonnet 5 handling most tasks and Opus 5 reserved for financial tasks needing deeper reasoning, and automates routine tasks like subscription cancellation as deterministic code that an agent repairs when a website changes. During Rowan's beta the team saw near-zero hallucinations, with apparent errors tracing to source data rather than the model, while monthly code commits grew 11x from 11 in December to 128 in July as the team built the agent.

The challenge

Rocket Money's users benefited from tracking spending, canceling subscriptions and negotiating bills, but that value still depended on people doing the work themselves inside an app, and many people who want to manage money well don't have the time, aptitude, or interest to sit in a dashboard. As Aaron Dignan, VP of Product & AI, put it, the kind of ongoing financial attention the wealthy get from a personal advisor had never been available to everyone else.

The solution

Rocket Money built Rowan, a financial agent reached by text message, running in production on Claude through Amazon Bedrock. Each incoming request is read by a Haiku 4.5 classifier and routed to a single-responsibility subagent: most run on Sonnet 5, with Opus 5 reserved for financial tasks needing deeper reasoning, and a final subagent formats every response for a consistent voice. Routine tasks like subscription cancellation are codified as deterministic scripts (Agent Skills) that an agent repairs when a website changes, and every action is auditable: structured outputs let the team run evals against deterministic results, classifier fall-throughs catch off-topic requests, and Claude classifies whether an action actually completed so failures trace back to the exact classifier or call that went wrong.

Reported business value

Across Rowan's beta the team saw near-zero hallucinations, with apparent errors tracing to source data rather than the model. Monthly code commits climbed from 11 in December 2025 to 128 in July 2026, an 11x increase in seven months, partly by widening who writes code: Rowan's product manager ranked ninth of 76 PR authors across the whole repository within three months of her first commit. Rowan is now beginning a phased private rollout starting with its first paying users.