What we learned building agents at Vercel
By Steven Van ·
Vercel shares its internal method for picking AI agent projects, plus results from a lead-qualification agent and an abuse-report agent.
Vercel has published a methodology for choosing which internal tasks to hand to AI agents, based on its own experience automating work at the company. The team's finding: current-generation agents do best on human work that has low cognitive load but high repetition, such as data entry, research, qualification, and triage, since those tasks are too variable for traditional automation but predictable enough for AI to handle reliably.
The post walks through two examples built internally. A lead processing agent now handles initial research, qualification, and follow-up email drafting for inbound sales leads, with a human approving each email over Slack before it sends; the change let one person do work that previously took a team of 10. An abuse platform agent analyses reported URLs for phishing or copyright issues and recommends an action for a security engineer to confirm, which cut time to close a ticket by 59% in its first iteration.
Vercel has also open-sourced templates for building similar agents, covering lead processing, a data analyst agent that turns natural language questions into SQL, a conversational flight booking assistant, and a Slack bot for collaborative storytelling. Read the full writeup on the Vercel blog.
