The problem
The product was a marketing-analytics SaaS. New users had to connect data sources, define funnels and set up alert rules — a process that took 45–90 minutes and a phone call with customer success. Trial-to-paid conversion was painfully low.
What we built
An in-product setup agent: a chat sidebar powered by Claude Sonnet 4.6, with three private tools — connect_source, create_funnel, create_alert. The agent walked the user through onboarding in their own words.
Architecture in one paragraph
The browser sent user messages to a thin Node.js gateway. The gateway ran a Claude tool-use loop, executed the chosen tool against the SaaS\'s own API, streamed responses back. Total new infra: one serverless function and 600 lines of code.
The numbers, four weeks after launch
- Median onboarding time: 72 min ? 28 min.
- Trial-to-paid conversion: +41%.
- Customer-success tickets per new signup: down 63%.
What surprised us
Users asked the agent business questions it wasn\'t built for ("what should my first funnel be for an e-commerce store?"). The model answered well, and that conversational expertise turned out to be a bigger sell than the automation itself.
We build SaaS, CRM and dashboard products with practical AI baked in. Tell us about yours.