Uploaded July 2026 | Updated September 2026, 2 weeks ago
Vercel's support agent resolves 91% of tickets fully autonomously, built in three weeks by two engineers after their vendor gave them 30 days to leave. The industry's dedicated support-agent companies typically hit 60–75%.
Recorded live at Vercel Ship 2026 in New York City, Kathryn Middleton (Engineering Manager, Vercel) walks through exactly how it was built. Seven months in production, the agent has handled 130,000+ cases, over 7,000 in a single week, saving support engineers 5,000 hours a month, the capacity of 31 full-time engineers, while costing less to run than one.
CHAPTERS
0:00 2 a.m., 11 tabs, cold coffee
2:51 Offboarded in 30 days: why we built our own
4:54 From ticket to resolution: enrich, research, act
6:02 The agent loop with AI SDK
8:19 AI Gateway: failover and one-line model swaps
9:37 Vercel MCP: +15% resolution from real data
11:57 A knowledge base with Upstash Search
14:49 Demo: billing, invoices, and approval UIs
15:47 Finding your first agent
LINKS
Upstash Search on the Vercel Marketplace → vercel.com/marketplace/upstash/upstash-search
Ship 2026 recap → vercel.com/blog/vercel-ship-2026-recap
Start building → vercel.com
FOLLOW VERCEL
X → https://x.com/vercel
LinkedIn → linkedin.com/company/vercel
GitHub → github.com/vercel
#Vercel #VercelShip #AIAgents #CustomerSupport #AI #DevTools
Vercel's support agent resolves 91% of tickets fully autonomously, built in three weeks by two engineers after their vendor gave them 30 days to leave. The industry's dedicated support-agent companies typically hit 60–75%.
Recorded live at Vercel Ship 2026 in New York City, Kathryn Middleton (Engineering Manager, Vercel) walks through exactly how it was built. Seven months in production, the agent has handled 130,000+ cases, over 7,000 in a single week, saving support engineers 5,000 hours a month, the capacity of 31 full-time engineers, while costing less to run than one.
CHAPTERS
0:00 2 a.m., 11 tabs, cold coffee
2:51 Offboarded in 30 days: why we built our own
4:54 From ticket to resolution: enrich, research, act
6:02 The agent loop with AI SDK
8:19 AI Gateway: failover and one-line model swaps
9:37 Vercel MCP: +15% resolution from real data
11:57 A knowledge base with Upstash Search
14:49 Demo: billing, invoices, and approval UIs
15:47 Finding your first agent
LINKS
Upstash Search on the Vercel Marketplace → vercel.com/marketplace/upstash/upstash-search
Ship 2026 recap → vercel.com/blog/vercel-ship-2026-recap
Start building → vercel.com
FOLLOW VERCEL
X → https://x.com/vercel
LinkedIn → linkedin.com/company/vercel
GitHub → github.com/vercel
#Vercel #VercelShip #AIAgents #CustomerSupport #AI #DevTools
