Building Vercel's First GTM Agent

We went from 10 inbound SDRs to 1 and saved $2M+. 32x ROI on an agent we built in a weekend.

I recently broke this down with Alex Lieberman on the Human in the Loop podcast. This post is the cleaned-up version. Everything you need to build something similar, minus the ums and tangents.

The Problem: Inbound Leads Were a Mess

Before the lead agent, our contact sales form was a disaster. Spam everywhere (it's a public form). Response times hit 24-48 hours because leads got routed to whoever was "next," regardless of timezone. SDRs spent more time sorting crap than talking to customers. And we had no intelligent routing, no sense of who was the best SDR for a given lead or who was even online.

We were pinging people at all hours saying "Hey, look at this and decide if it's worth working." That's a horrible experience for everyone involved.

The Solution: Lead Agent

Here's what the system does now:

1. Form submission → enrichment (Clearbit, ZoomInfo, etc.)
2. Lead Agent kicks off → deep research on the company/person
3. AI makes a qualification decision → categorizes into buckets (qualified, follow-up, support, not sales-related)
4. Slack notification → pings the SDR with full context + reasoning
5. Draft email ready → for qualified leads, we've already written the response and pushed it to Outreach

The human's job went from "do research, qualify, write email, send" to "review AI's work, press send."

It's like we gave them the controller to a video game instead of making them play the boring parts.

Two Types of Agents You Should Build

This is how I think about where to apply AI in GTM:

Type 1: Efficiency Agents

Take work that's already being done and make humans faster at it.

Before: Rep spends 4 hours on tedious work, 2 hours on high-value work → 1 deal/day

After: Agent handles tedious work, rep spends 6 hours on high-value work → 3 deals/day

Same hours, 3x output. The SDR gets paid more (variable comp), has a more fulfilling day, and the business wins.

Type 2: "Should Be Doing" Agents

This is the bigger opportunity. These unlock work that wasn't happening at all.

Every company has a list: analyze why we're losing deals, track competitors, get field feedback to product, re-engage churned customers. These never happen because there's no time, no owner, always next quarter.

Anytime you hear "should" in GTM, think extra hard—that's probably a great spot for AI.

Examples we've built:

• Close/loss analysis — AI reads all the Gong calls and extracts why deals were actually lost (not the seller's excuse)
• Objection extraction — pulls objections from sales calls, determines if they were handled well, routes product gaps to the right engineering teams
• Field feedback pipeline — aggregates customer issues and surfaces them to product

These agents create new pipeline AND product insight that didn't exist before.

The Sweet Spot for AI in GTM

Current-gen AI is best at tasks that are:

• Boring — humans don't want to do them anyway
• Low-cognitive load — pattern matching, not novel strategy
• Repetitive — same workflow over and over

Lead qualification checks all three boxes. So does email drafting, research summarization, and data entry.

How to Build Your Own Lead Agent

Step 1: Shadow Your Best SDR

Before writing any code, I spent a week watching our top performer. Not just what she did—how she thought about leads.

• What signals made her excited about a lead?
• What made her immediately disqualify one?
• How did she research companies?
• What did she include in emails that got responses?

This is the "revealed preference" approach—watch what the best people actually do, not what the playbook says they should do.

Step 2: Prototype Fast

For early prototypes with no production data integration, just vibe code it. Move fast, validate the concept.

Once you're connecting to real systems (CRM, customer data), you need proper engineering.

Step 3: Focus on High-Impact, Low-Effort First

Pick the projects in the top-left quadrant. Lead qualification was perfect because:

• High impact (touches every inbound lead)
• Relatively low effort (structured workflow, clear inputs/outputs)
• Easy to measure (response time, conversion rates, SDR headcount)

The Architecture

Tech stack:

• Next.js — the app itself
• AI SDK — agent + structured outputs
• Workflow DevKit — durable execution for the multi-step process
• Vercel Slack Adapter — human-in-the-loop approval flow
• Exa.ai — web search for research

Key insight: Give your agent as much context as possible before it makes decisions. We implemented deep research (inspired by Anthropic's multi-agent research system) that does comprehensive company research before qualification.

The research feeds into the AI's decision AND becomes an artifact the human uses to verify the decision.

Get the Template

We open-sourced the entire architecture: Lead Agent Template

It includes:

• Form handling and enrichment
• Deep research agent with web search
• Lead qualification with structured output
• Slack integration for human-in-the-loop
• Email draft generation

Fork it, adapt it to your business, ship it.

The ROI Math

• Cost: ~$60K/year (engineering hours + AI API costs)
• Savings: $2M+/year (headcount reduction + efficiency gains)
• ROI: 32x

And that's not counting the incremental revenue from the outbound SDRs we moved those people into, or the faster response times improving conversion.

The Human Element

The remaining SDR? Their job is way better now. Output up, variable comp up, job satisfaction up. They review pre-qualified leads with full context and press send. The system does the tedious parts. They do the human parts.

If you have humans doing repetitive, boring work that follows patterns, you should be building agents. Shadow your best performers, prototype fast, focus on high-impact workflows. The lead agent was built in a weekend. The $2M+ in savings came from actually deploying it.

Watch the full episode: Human in the Loop - Lead Agent

Get the template: vercel.com/templates/ai/lead-processing-agent