What we learned building agents at Vercel (vercel.com)

Vercel shared a practical playbook for building internal AI agents that prioritize ROI and reliability over hype. After months of experiments, they recommend focusing on the “agentic sweet spot”: tasks with low cognitive load but high repetition (data entry, triage, research, qualification) where current frontier models are predictable enough to save time and maintain quality. Their methodology is simple—ask teams what boring, repetitive work they hate, prototype an agent that automates the grunt work, and keep a human-in-the-loop for validation—recognizing production constraints and model limitations today. They illustrate the approach with two production agents and measurable outcomes. A lead-processing agent automates deep research, uses a generateObject-based classifier for qualification, drafts personalized emails, and posts results to Slack for human approval—one person now handles work that used to need ten. An anti-abuse agent ingests URLs, runs visual and textual analysis, recommends actions, and routes cases for engineer review, cutting time-to-close by 59% in iteration one. Vercel open-sourced agent templates (lead processing, natural-language-to-SQL data analyst, fault-tolerant flight booking, Slack storybot) and offers a hands-on program to help teams discover and build high-ROI agents, emphasizing reproducible workflows, multi-phase reasoning, and pragmatic human supervision.