Vercel built a lead agent that researches and qualifies every inbound sales lead and drafts the reply for a human to send, according to a blog post by Drew Bredvick, who built the agent, published on January 23, 2026.
TL;DR: The lead agent starts when a prospect fills in the contact sales form. It enriches the lead, runs deep research on the company and the person, sorts the lead into one of four buckets (The Solution), and sends the reasoning to a sales rep in Slack. For a qualified lead, a draft email is ready. Drew Bredvick reports that the team went from 10 inbound sales development reps (SDRs) to 1 (per the post), saved more than $2M a year, and got a 32x return on about $60K a year of cost (The ROI Math). He also reports that the first version took a weekend to build (The Human Element). All numbers are the author's own, with no stated method.
| Fact | Detail |
|---|---|
| Source | View source |
| Source type | Blog post by the builder, 5 min read |
| Source published | 2026-01-23 |
| Speakers | Drew Bredvick, author and builder of the agent |
| Industry | Developer platform and cloud hosting |
| Function | Sales |
| Techniques | Enrichment, deep research, structured output, human review in Slack, durable execution |
| Tools named | Next.js, AI SDK, Workflow DevKit, Vercel Slack Adapter, Exa.ai, Clearbit, ZoomInfo, Outreach |
Independent summary of public material. Vercel is not affiliated with Taskade.
The system Vercel built
Before the agent, the contact sales form at Vercel was a public form with a lot of spam. The author reports response times of 24 to 48 hours (The Problem), because the next rep in line got the lead whatever the time zone. Reps spent more time on sorting leads than on talking to buyers, and no logic picked the best rep for a lead.
The lead agent now does the sorting and the first draft. The human job changed from research, qualify, write and send to a short review. In the author's words, the rep's task is now to "review AI's work, press send."
The post names a second class of agent: work that the team agreed was useful but never did. The examples are a loss analysis that reads sales call recordings, an objection extractor that routes product gaps to engineering, and a pipeline that collects field feedback for the product team.
Architecture of the system
The lead agent runs as a chain of stages:
- A form submission arrives, and enrichment services add company and contact data.
- The agent runs deep research on the company and the person. The author says that this design took ideas from Anthropic's published multi-agent research system.
- The model makes a qualification decision with structured output. Each lead goes into one bucket: qualified, follow-up, support, or not sales-related (The Solution).
- A Slack message gives the rep the full context and the reasoning.
- For a qualified lead, the agent writes a reply and puts the draft in Outreach.
The research has two jobs (The Architecture). It feeds the decision, and it is the record that the rep reads to check the decision. The author gives one main lesson: give the agent as much context as possible before it decides.
Before any code, the author spent a week (Step 1) with the top SDR to learn which signals made her want or reject a lead, how she researched a company, and what got replies. The first prototype had no production data. The author reports that real connections to the CRM and customer data needed proper engineering. The team published the architecture as an open-source template.
Results the source reports
- Drew Bredvick reports that inbound SDR headcount went from 10 to 1 (per the post). The other reps moved to outbound sales roles (The ROI Math).
- He reports savings of more than $2M a year from fewer heads and faster work, against a cost of about $60K a year for engineering time and AI usage (The ROI Math). He calls the result a 32x return.
- He reports that the team built the agent in a weekend, and that the savings came from the deployment, not the build (The Human Element).
- He reports that the remaining SDR has higher output, higher variable pay and a better job.
- The post gives an example for an efficiency agent: a rep moves from 1 deal a day to 3 deals a day in the same hours (Type 1). The post presents this as an illustration, not as a measured Vercel result.
Critical assessment
All numbers come from the builder and appear in a short blog post. The post does not give the period of the savings, the cost of the moved reps, or a before-and-after conversion rate. The $2M figure (The ROI Math) mixes headcount and efficiency gains, so the share that the agent alone produced is not clear. The stack in the post is mostly Vercel's own products, so the post also works as a product example. The post is a cleaned-up version of a podcast episode. No outside party checked the figures. The architecture is simple and specific, and it is the most reusable part of the source.
Build this in Taskade
- Build an AI agent with web search and persistent memory that researches each new lead and writes a short brief. Start at AI agents.
- Start from a template in SDR agents.
- Run the stages around the agent, such as routing and follow-up tasks, with sales automations.
- Send each brief to your sales channel with the Slack integration, one of 100+ integrations.
- Start an automation from your own form or another system with a webhook trigger, on Pro or higher. See the webhook trigger.