Vercel runs inbound sales on 14 rules, $1,000/year

Vercel's approach started with a single engineer spending roughly 20% of his time on the problem, and the first version of the agent was a prompt of about 125 lines written by the best SDR on the team. The company kept human SDRs in the loop from June through August, with the top SDR managing the agent like a new rep, reading 100% of its first hundred emails before shifting to sampling one out of every 100 outreaches. Over the following year, as Vercel added product surface area and moved upmarket into enterprise, the prompt grew to 1,000 lines.

The team then divided the prompt into two parts: rules and judgment. Engineers encoded the rules, leaving the model only the work that required judgment. This matches a pattern the author found in 14 production agent workflows, where 65% of nodes run as pure code and only 14% remain fully agentic. Vercel also runs a second agent that watches for breaches of the 14 rules and either fixes or blesses exceptions.

The stakes for Vercel appear to hinge on whether the rules-based system can keep pace as the business evolves, and on how well the monitoring agent catches breaches, since the model is no longer trusted to follow all qualification rules on its own.