Automating 90% of support tickets: How we built Vercel's support agent
summary
Vercel’s support agent resolves 91% of tickets fully autonomously after being built in three weeks by two engineers when a vendor demanded offboarding in 30 days. It runs seven months in production, handled 130,000+ cases and 7,000 in a single week, saves 5,000 hours/month (capacity of 31 engineers) while costing less than one, and improves resolution by +5% after swapping models (Opus 4.6 → Gemini 3.5 Flash); stacks used: AI SDK, AI Gateway, Vercel MCP, Upstash Search, Next.js.
AI-written summary · about
chapters (9)
- 0:002 a.m., 11 tabs, cold coffee
- 2:51Offboarded in 30 days: why we built our own
- 4:54From ticket to resolution: enrich, research, act
- 6:02The agent loop with AI SDK
- 8:19AI Gateway: failover and one-line model swaps
- 9:37Vercel MCP: +15% resolution from real data
- 11:57A knowledge base with Upstash Search
- 14:49Demo: billing, invoices, and approval UIs
- 15:47Finding your first agent
related talks
19mAutomating 90% of support tickets: How we built Vercel's support agent
19m1,200 analytics requests a day: How we built Vercel's data agent
25mBuilding high performance, AI-native engineering teams
50mOpening Keynote
46mBecoming an AI engineering company
54mKeynote
16mSoftware that ships itself: self-driving apps
16mFrom storefronts to shopping agents: Ecommerce