Vercel's lead qualification agent has cut its team from ten people to one in six weeks and returns 32 times its cost.
Vercel's support agent handles 93% of its caseload and its content agent did 96% of major content updates.
Vercel applied its framework to 30 SDR workflows, lifting quotas 30%, and one in three customer deployments now comes from agentic coding tools.
Vercel's COO on Replacing Ten SDRs With AI Agents
Vercel turned its go-to-market organization into a test of whether in-house AI agents can do real work at scale. It built a GTM engineering team, paired engineers and data scientists with the best subject matter experts, and encoded documented workflows into agents with a human in the loop. The bet is that agents need composable APIs and a semantic data layer, not a UI, and that internal builders will out-ship vendors. Grosser says the infrastructure for real agents only arrived in 2026, so most companies have not yet hit the problems that come with scale.
Takeaways
Lead qualification agent replaced ten SDRs. Vercel launched it in August 2025 with about 20 percent of one engineer. After six weeks in shadow mode, with its best SDR reviewing every output, the company moved from ten people to one human in the loop. It costs under $5,000 a year and returns 32 times the investment.
Support agent handles 93 percent of caseload. Vertex, built in house in two months after an off-the-shelf tool fell short, handles thousands of technical cases a week. It costs $300 a month in infrastructure plus $12,000 in tokens, and three engineers work on it.
Build agents with a GTM tripod. A GTM engineer and a data scientist work with the best subject matter expert in the function. They document that function's best practice and encode it into tool-calling workflows. A human stays in the loop for QA until the expert runs out of feedback.
Data foundation and semantic layer matter. Vercel built a semantic layer so D0 can translate natural-language questions into SQL and answer in under a minute, where tickets to data science once took a week. Accurate enrichment lets the lead agent qualify better than MQL frameworks and write copy more naturally than an SDR.
Agents need headless composable developer surfaces. Agents do not live in UIs. They call APIs and MCP servers, so products without developer surfaces are invisible to agentic workflows. DealOne worked because Gong had an accessible API and Salesforce had webhooks. Grosser says Vercel would drop a tool that lacked one.
2026 infrastructure makes real agents possible. Grosser says 2025 was declared the year of agents but delivered assistance and prototyping. Cloud infrastructure built for request and response cannot support LLM calls and long-running tool use. Most companies have not run agents at scale, so they have not hit the rough edges.
SDR quotas rose 30 percent. Vercel documented the SDR workflow and made it legible before touching AI, then applied the framework to 30 workflows, from event follow-up to time-based campaigns. SDR quotas rose 30 percent that quarter.
Tap a name to go there · tap a line to see why they are linked