Vercel COO 复盘智能体进销售与客服，线索团队 10 人缩到 1 人，93% 客服工单由 AI 处理

SDR Quotas Up 30%, a Support Agent Built In-House in 2 Months, and Why Vercel Will Drop Any Tool Without an API

"At @vercel we ran our AI SDR in shadow mode for 6 weeks.

Our best AI SDR reviewed every single output to improve the flywheel.

And then finally, she couldn't improve it anymore.

It performs like a 90th percentile rep now. 100% of the time." COO Jeanne DeWitt Grosser https://t.co/DjHvrWvfCJ pic.twitter.com/0uiyGG5TX5

— Jason ✨👾SaaStr.Ai✨ Lemkin (@jasonlk) November 5, 2026

Vercel used to have 10 people qualifying inbound leads. Six weeks after it turned on an agent, it had one. The agent doing the rest of the work costs under $5,000 a year in infrastructure and tokens.

Jeanne DeWitt Grosser, Vercel’s COO, set up a go-to-market engineering team six weeks after she joined. Its mandate was to put agents into every GTM function. Here’s where Vercel stood when she spoke at SaaStr AI:

The support agent handles 93% of total case volume, and most Vercel support cases are deeply technical.
The content agent did 96% of major marketing content updates last quarter.
Lead qualification went from 10 people to 1.
SDR quotas went up 30% in the quarter after Vercel automated 30 SDR workflows.

Vercel built all of this internally, mostly with fractions of an engineer’s time.

#1. The Lead Agent: 10 People Down to 1, Built by 20% of One Engineer

The lead qualification agent launched with about 20% of one engineer’s time behind it and a human reviewing its work. Over six weeks the 10-person team shrank to one person.

One person now runs it for the US. About 20% of another person covers Europe and APAC.

The economics:

Under $5,000 a year in infrastructure and tokens
About 20% of one engineer to maintain
32x ROI by Jeanne’s math, against the cost of the 10-person team
Runs 24/7, with faster speed-to-lead and quality metrics that match the human team

Jeanne said the people from that team moved into more valuable roles. For any GTM leader, the math is now on the table: a function that runs for $5K a year with an agent, or 10 salaries.

The build process:

A GTM engineer shadowed Vercel’s best SDR for a few days and logged every tab she opened: LinkedIn, BuiltWith, the company website, the CRM, Slack history.
Each of those data sources became a step in a tool-calling workflow.
The team documented the workflow and made it repeatable before involving any model.
The first version ran in production in shadow mode for 6 weeks. The top SDR reviewed every output and fed corrections back in.
When she had nothing left to correct, they took her out of the loop.

Jeanne says the agent now performs like a 90th-percentile rep on every lead. One engineer prototyped it over a weekend, and it was in production six weeks later.

#2. Every Agent Gets Built by the Same Three-Person “Tripod”

Vercel builds every internal agent with the same three people:

A GTM engineer
A data scientist
The single best performer in the function being automated

They sit side by side, write down how the best performer does the job, and turn that into the workflows that make up the agent. The agent runs with a human checking its work and can’t act on its own. When the expert stops finding things to fix, the human comes out.

Vercel ran this process on 30 SDR workflows, including event follow-up, product-qualified lead flows, and time-based campaigns. SDR quotas rose 30% that quarter.

The constraint here is your top performer’s calendar. Vercel’s agent took a few days of shadowing and six weeks of daily output review from its best SDR. Staff that with an average rep and you get an agent that performs like an average rep.

#3. Vercel Built Its Own Support Agent for ~$150K a Year with 3 Engineers

Vertex, Vercel’s customer service agent, powers vercel.com/help. Jeanne called it “our internal Sierra.”

$300 a month in infrastructure, plus about $12,000 a month in tokens
About $150K a year all-in
Thousands of technical cases a week
Three engineers on it
2 months from start to production

Vercel started with an off-the-shelf product, didn’t get the results it wanted, and built its own. Jeanne compared that with agentic support vendors that have 150 engineers running equivalent workflows at much higher cost.

She still revisits the decision every quarter. She asks her GTM engineering team whether anyone has built something better, faster, or cheaper, and if so, Vercel switches.

Her benchmark for the team: the average builder should be able to out-ship any vendor selling the same outcome. Teams that can’t yet should start by carving out a few hours a week to experiment.

We’ve seen the same at SaaStr, where 3 humans run the company with 20+ agents in production. For B2B + AI vendors, this sets a hard comparison. The buyer’s alternative is a weekend prototype maintained by 20% of a GTM engineer. Vendors that win against that have to unlock budget beyond the old software line item.

#4. If Your Product Has No API, Vercel Will Drop It

Agents call APIs and hit MCP servers. They don’t click through UIs. Jeanne’s first takeaway was that products need headless, composable architecture to stay in an agent-driven stack.

Vercel’s meeting intelligence agent ingests every sales call. It writes notes and action items, posts coaching suggestions to internal Slack channels, proposes CRM field updates, tracks competitive mentions and objections over time, and runs a post-mortem on every closed-lost deal.

When a rep tags it in Slack with a question, it searches Slack in real time, queries d0 (Vercel’s data agent) as a sub-agent, pulls Gong transcripts, searches the internal knowledge base, and streams the answer back into the channel. The rep stays in Slack. The agent has no UI at all.

That worked because Gong has an accessible API and Salesforce has webhooks Vercel could build automated workflows on. Jeanne said that if either tool had lacked a developer surface, Vercel would have replaced it.

The playbook platform follows the same design. Vercel’s best reps knew to reach out when a customer’s usage spiked or a high-value prospect landed on the pricing page. How often that happened depended on who owned the account and whether they checked their dashboard.

Now a usage spike or high-intent pricing page visit triggers the platform. It matches the signal to a play, drafts personalized outreach, and sends it to the rep for a one-click review inside their existing workflow. Every rep gets the top reps’ plays automatically.

For B2B vendors, this shows up as churn your health scores won’t predict. Customers’ internal agents decide which tools stay, and a tool they can’t call gets cut. Box, Notion, and Salesforce are all investing heavily in developer platforms right now. Jeanne’s to-do list: MCP servers, webhooks, APIs.

She also shared that one in three deployments from Vercel customers now comes from an agentic coding tool like Claude Code or Codex.

#5. d0 Answers in Under a Minute What Used to Take a Week of Data Tickets

Vercel’s most-used agent company-wide is d0, a data analyst agent anyone can query in Slack. It translates plain-English questions into SQL against Vercel’s analytics infrastructure. Questions that used to take a week through a data science ticket now come back in under a minute:

What was our token volume through AI Gateway last week?
Which UK companies do we have logo rights for?
Who are our highest-revenue startup customers using Sandboxes?

d0 runs on a semantic layer Vercel’s head of data science built on top of a model of Vercel’s revenue. The layer breaks the business into its smallest causal units and enriches them with first- and third-party signals.

Every agent at Vercel runs on that layer. The meeting intelligence agent uses it to make its analysis actionable, and the playbook platform uses it to pick which companies to contact and what to say. Jeanne said that without it, agents hallucinate or give generic answers.

Her summary: good data equals good agents. With accurate customer enrichment in the semantic layer, Vercel’s lead agent qualifies better than an MQL framework and writes more natural copy than an SDR.

For most founders, that means the data warehouse and semantic layer project that has slipped for two years now sits in front of every agent on the roadmap. Jeanne called this work unexciting and load-bearing. She also found that once the data was in front of people across the company, they started improving the data itself.

A Few More Learnings from Vercel

Costs jump at 100x to 1,000x pilot volume. Most companies haven’t run an agent at that scale yet. Cloud infrastructure was designed for request-in, response-out. Agents run long jobs, call LLMs, and call tools, and that needs different primitives. Vercel says Fluid compute cut early adopters’ compute costs by up to 85%, and its internal agents run on the same infrastructure its customers use.
Vercel wants its infrastructure to propose fixes. One of her three requirements for agent infrastructure is that it acts on problems itself, rather than sending alerts and dashboards to a human.
Enterprise buyers want advice on where to start. The executives Jeanne meets ask her where to begin with agents, what the architecture should be, and what breaks at scale. Few ask for a product pitch.
Vercel runs GTM like a product. Her closing point: document the GTM motion, design it, iterate on it, then automate it one piece at a time.