A Bridge Group survey of 351 B2B companies found that AI SDRs were the primary sales-development motion at 1% of respondents. One percent. So when Vercel's COO Jeanne DeWitt Grosser told SaaStr AI that her team went from 10 inbound lead-qualification people to roughly 1.2, the natural reaction is: outlier, not playbook.

Fair. But the operational details she shared are worth more than the headline numbers, because they describe a rollout method any marketing ops team can steal, whether or not you're ready to deploy an agent.

Shadow the Human Before You Touch the Model

Grosser's framework starts with a step most AI projects skip: a GTM engineer sat next to Vercel's best SDR for several days and logged every tab she opened. LinkedIn, BuiltWith, the company website, CRM, Slack history. Each data source became a step in a documented, tool-calling workflow. The team made that workflow repeatable before any model was involved.

This is process documentation dressed up as an AI project. It matters because the failure mode for most agent deployments isn't the model; it's that nobody wrote down how the human actually does the job. If your lead-qualification logic lives in one person's head and three browser tabs, you don't have a process to automate. You have institutional risk.

Six Weeks of Shadow Mode as a QA Layer

The first production version of Vercel's lead agent ran in "shadow mode" for six weeks, per the SaaStr talk transcript. The top SDR reviewed every output and fed corrections back in. When she couldn't improve the outputs anymore, they removed her from the loop.

For ops teams, this is the interesting part. Shadow mode is a control layer: the agent produces outputs, a human grades them, corrections tighten the workflow. Same logic you'd use to validate a new lead-scoring model or a routing rule change, except the feedback loop is daily and the iteration cycle is compressed. The constraint is real, though. Six weeks of daily review from your best performer is expensive calendar time. Staff that role with a median rep and you get a median agent. Grosser said this explicitly.

The Cost Numbers Are Compelling (and Incomplete)

The lead-qualification agent costs under $5,000 a year in infrastructure and tokens, plus about 20% of one engineer's maintenance time. Grosser cited a 32x ROI against the cost of the previous 10-person team. That $5K figure is eye-catching, but it excludes the fully loaded engineering cost. Twenty percent of an engineer's time isn't free; at market rates, that's $30K-$50K a year depending on your comp bands. The ROI is still strong, but the honest number is closer to $35K-$55K than $5K. Plan accordingly when you present this to finance.

The support agent tells a different cost story. Vercel's in-house support agent (handling 93% of case volume) runs on about $300/month in infrastructure plus roughly $12,000/month in model tokens, with three engineers building and maintaining it. That's around $150K a year before fully loaded engineering labor. If you're budgeting for agents, the gap between a lead agent and a support agent is the planning variable most teams miss.

The Metric Hygiene Problem Nobody Wants to Talk About

Grosser reported SDR quotas rose 30% in the quarter after Vercel automated roughly 30 SDR workflows. She reported the support agent handles 93% of case volume. Both are company-reported figures from a conference stage, not independently audited results.

The 30% quota increase is temporally associated with the rollout but not proven causal. The 93% support figure doesn't specify whether "handled" means fully resolved without human involvement, triaged and routed, or something else. None of this makes the numbers wrong. It makes them directional. And directional is fine, as long as you don't import someone else's undefined metric into your own reporting and call it a benchmark. If you're building your own agent, define "handled" before you deploy, instrument the measurement from day one, and separate triage from resolution in your reporting.

What to Steal (Even Without an Agent)

Grosser's framing was "process before AI." Observe a top performer, document the workflow, make it repeatable, then automate. That sequence works whether the automation layer is an LLM agent, a Zapier workflow, or a junior hire following a runbook.

The 1% adoption figure from the Bridge Group survey suggests most B2B teams aren't close to Vercel's position. The gap isn't model capability. It's process maturity, data quality, and the willingness to spend six weeks letting a human correct the machine before you trust it. The boring work is the load-bearing work.