Skip to content
Reviews, articles and a sales glossary for SaaS sales leadersUpdated November 2026
Articles / Talk Review

AI Lead Qualification at Vercel: A SaaStr Talk Review

Key takeaway

Vercel's AI lead qualification agent is a clear, practical case study: shadow your best SDR, document what she does, encode it as a workflow, and run it with a human reviewer until the reviewer has nothing left to fix. The talk is worth 21 minutes for any sales leader, but its cost and headcount numbers are Vercel's own claims, and much of the infrastructure advice doubles as a pitch for Vercel's platform.

Jeanne DeWitt Grosser on SaaStr AI. Watch on YouTube

Talk: Going from 10 to 1 SDRs: How We Deployed AI Across our Sales Org Speaker: Jeanne DeWitt Grosser, COO of Vercel Where: SaaStr AI, on the SaaStr YouTube channel (about 21 minutes)

Most talks about AI lead qualification stay at the level of promise. This one gives numbers. Jeanne DeWitt Grosser, the COO of Vercel, the company behind the Next.js framework, walks through how her go-to-market team built an agent that qualifies inbound leads, and how that agent took a 10-person lead qualification team down to one person in the US plus a fraction of a person for Europe and Asia Pacific.

The talk is short, specific and easy to act on. It is also a talk by an infrastructure vendor, and part of it explains why you need that vendor’s infrastructure. This review separates the operating lessons a sales leader can copy from the parts that are Vercel’s product story.

What the talk covers

Grosser opens with the headline results. Six weeks into the job, she set up a go-to-market engineering team whose only goal was to bring AI agents into the company’s own sales, marketing and support work. Ten months later, she says, the customer support agent handled 93% of the total case load and a content agent made 96% of the major content updates in a quarter.

The sales story is the lead agent. It started with about 20% of one engineer’s time. With a human reviewing every output for six weeks, it replaced all but one person on a team of ten. Grosser stresses that the people moved into other roles rather than out of the company.

From there the talk has three parts:

  • The build method: how the team turns a human job into an agent.
  • Three lessons: headless products, a strong data foundation, and a new build versus buy calculation.
  • The agent tour: a meeting agent, a data analyst agent, a signal-based outreach platform and a support agent, each used to illustrate one of the lessons.

How Vercel built its AI lead qualification agent

The most useful five minutes of the talk is the build method. Every internal agent at Vercel starts with what Grosser calls a tripod: a GTM engineer, a data scientist and the single best person in the function. (A GTM engineer is a technical operator who builds automations and agents for the sales and marketing teams.)

For the lead agent, the steps were:

  1. Shadow the best rep. A GTM engineer sat beside the team’s top SDR for a few days and watched every tab she opened: LinkedIn, BuiltWith, the company website, the CRM and Slack history.
  2. Turn each tab into a step. Each data source became a tool call in a workflow. The team documented the process and made the workflow work well before adding any AI.
  3. Run in shadow mode. The agent ran in production for six weeks without acting on its own. The top SDR reviewed every output and fed corrections back.
  4. Remove the reviewer when the feedback stops. When she could no longer improve the results, the human came out of the loop.

Grosser’s claim is that the agent now performs like a 90th percentile rep, 100% of the time. The team then applied the same framework to 30 other SDR workflows, from event follow-up to product-qualified lead flows and timed campaigns. She adds that SDR quotas rose 30% in the quarter she described.

Vercel has published an open-source version of the pattern. The lead agent repository on GitHub runs a form submission through a research step and a qualification step, drafts an email, and posts it to Slack with approve and reject buttons. It is a reference build, not the exact system from the talk, but it shows the shape.

The numbers, and how to read them

Claim in the talk Vercel’s figure How to read it
Lead qualification team 10 people to 1 (plus about 20% of a person for Europe and Asia Pacific) Inbound qualification only, not the whole sales team
Lead agent running cost Under $5,000 a year in infrastructure and tokens Excludes the engineer’s time spent maintaining it
Upkeep About 20% of one engineer A real cost to budget for
Return 32x Vercel’s own calculation from salaries saved
Time to production Weekend prototype, live six weeks later With a strong engineer and a clean process already in place
Support agent 93% of support cases, 3 engineers, built in two months Vercel-reported; a technical support queue

Three cautions apply. First, these are self-reported figures from a conference stage, with no public methodology. Second, the 10-to-1 result covers one narrow job: qualifying inbound leads. Outbound prospecting and closing are not part of the claim. Third, the lead agent sits on top of a data layer that Vercel says its head of data science spent real effort building. That investment does not show up in the $5,000.

None of that makes the numbers wrong. It means a sales leader who wants the same result should budget for the process work and the data work, not only the model and the tokens.

Three lessons for sales leaders

1. Your best rep’s process is the specification

The tripod method is the part of the talk that transfers to any company. Before any AI, the team wrote down exactly what the best SDR did and in what order. That is a sales playbook in the most literal sense. If your team cannot describe how a good rep qualifies a lead, an agent cannot learn it either.

2. Data quality decides agent quality

Grosser’s line is “good data equals good agents.” Vercel built a knowledge base and a semantic layer, which is a structured model of the business that every agent queries. Her claim is that with accurate customer enrichment underneath, the lead agent qualifies better than any MQL framework, and its emails read more naturally than a rep’s. Without that layer, she says, agents hallucinate or give generic answers. For most sales teams, this is the real project: clean CRM fields, consistent definitions and reliable enrichment.

3. Build versus buy has shifted

The talk argues that one strong engineer can now build a workflow-specific agent in weeks, so buying is no longer the default. Vercel started its support work on an off-the-shelf product, did not get the results it wanted, and built its own. Grosser balances this with a quarterly review: if a vendor has built something better, faster or cheaper, the internal agent goes. That is a sensible rule for anyone weighing AI SDR tools against an in-house build.

The other agents in the talk

Deal One (meeting intelligence). Ingests every sales call, writes notes and action items, posts coaching to Slack, proposes CRM updates, tracks competitor mentions and runs a post-mortem on every closed-lost deal. Reps ask it questions in Slack, and it pulls Gong transcripts and searches internal knowledge.

D0 (data analyst). Answers questions in Slack by turning plain English into SQL. Grosser says questions that used to take a week through the data team now take under a minute.

The Playbook platform. Watches for signals such as a usage spike or a high-intent pricing page visit, matches the signal to a play, drafts personalized outreach and puts it in front of the rep for a one-click review. This is the best example in the talk of AI making every rep act like the top rep, rather than replacing reps.

Vertex (support). The customer support agent behind the Vercel help site, which the company says handles thousands of technical cases a week.

The common design choice is that none of these agents has its own interface. Work flows in from Slack, the CRM and call recordings, and drafts flow back into the tools reps already use. Grosser notes that Deal One only worked because Gong and Salesforce offered APIs and webhooks. Tools without that kind of developer access, she says, get dropped.

Where the talk falls short

It is partly a product pitch. Several minutes of the talk explain why classic cloud infrastructure struggles with long-running agents and why Vercel’s own compute product fixes it. Grosser cites early adopters cutting compute costs by up to 85%. That may be true for some workloads, but it is a vendor claim, and a sales team building a lead agent will rarely hit infrastructure limits first. The Fluid compute documentation is the place to check the details if that part matters to you.

The hard parts get one sentence each. The talk moves quickly past the data foundation, the change management for the nine people who moved roles, and how the team measured “90th percentile.” A sales leader will want more on each before promising the same result to a board.

It fits a technical company. Vercel sells to developers, has strong engineers in every team and already runs on Slack. A company without an engineer to spare will find the build path harder than the talk suggests.

Who should watch it

  • Sales and revenue leaders planning where AI fits in their team: the build method and the three lessons are the takeaway.
  • SDR managers worried about headcount: the story is about moving people to higher-value work, and the Playbook platform shows AI supporting reps.
  • RevOps and GTM engineers: the tripod and shadow-mode approach is a repeatable project plan.
  • Skip it if you want a vendor comparison. This is one company’s build story.

The AI SDR entry in the Library covers the product category this talk sits in. For what an AI agent might send once a lead qualifies, real follow-ups from Vercel, Gong and hundreds of other companies are in the Sales Templates library.

Verdict

This is one of the more useful public talks on AI lead qualification because it shows a method, not a demo. The headline numbers come from Vercel itself and cover one narrow job, so treat them as a ceiling rather than a forecast. The lesson that holds up anywhere is the simplest one: document your best rep’s process, fix the data underneath it, and keep a human reviewer in the loop until the agent earns trust.

FAQ

What is AI lead qualification?

AI lead qualification uses an AI agent to research each inbound lead, score or categorize it, and decide the next step, such as routing it to sales, sending a follow-up or passing it to support. Vercel’s version researches the company, qualifies the lead and drafts the outreach for a person to approve.

How did Vercel go from 10 SDRs to 1?

According to Grosser, a GTM engineer shadowed the best SDR, turned each of her research steps into a workflow, and ran the agent in shadow mode for six weeks while she reviewed every output. When her feedback stopped improving the results, the agent took over the inbound qualification work, and the other team members moved into other roles.

How much does the Vercel lead agent cost to run?

Grosser says it costs under $5,000 a year in infrastructure and tokens, plus about 20% of one engineer’s time to maintain. She describes that as a 32x return against the salaries of the original team.

Can I use the Vercel lead agent?

Vercel publishes an open-source reference version as a template and a GitHub repository. It covers research, qualification, email drafting and Slack approval, and is meant to be adapted to each company’s own process.

Is the talk worth watching?

Yes, for any sales leader planning AI work. It is about 21 minutes, the build method is practical, and the numbers are specific. Watch the infrastructure section with the knowledge that the speaker sells that infrastructure.

Sources

Checked November 16, 2026.

  1. youtube.com/watch?v=wXHLbr4W-8s
  2. saastrannual.com/session/going-from-10-to-1-sdrs-how-we-deployed-ai-across-our-sales-org-with-vercel-s-coo
  3. saastr.com/vercel-took-a-10-person-sdr-team-down-to-1-the-whole-thing-costs-5000-a-year-with-vercels-coo-jeanne-dewitt-grosser
  4. github.com/vercel-labs/lead-agent
  5. vercel.com/templates/next.js/lead-processing-agent
  6. vercel.com/docs/fluid-compute