Vercel’s COO Jeanne DeWitt Grosser ran go-to-market at Google and Stripe for roughly a decade each before joining Vercel. Six weeks into the job, in June 2025, she stood up a go-to-market engineering team with one mandate: bring agents to everything in GTM. That was before “GTM engineering” was a phrase anyone used.
Ten months later, the team has automated a real chunk of core company functions. Not a demo. Production, at scale, with the costs and the rough edges included.
The customer support agent now handles 93% of total case load. And Vercel’s support cases are not “reset my password.” They are deep, technical, infrastructure-level problems. The content agent did 96% of major content updates last quarter. And the lead qualification agent took a 10-person function down to roughly one and a quarter people.
The Lead Agent: 10 People to 1, for Under $5K a Year
Vercel launched a lead qualification agent in August 2025. It started as 20% of a single engineer’s time. With a human in the loop over six weeks, they moved the entire function from 10 people down to one person running it in the US, plus about 20% of a person covering all of Europe and all of APAC.
The agent runs about $5,000 a year between infrastructure and tokens. It takes 20% of one engineer to maintain. Jeanne’s math on that: a 32x ROI. You saved 10 salaries, replaced them with $5,000 of compute, and the thing runs 24/7 with faster speed-to-lead and human-equivalent quality.
When this came out publicly, plenty of people declared the end of the knowledge worker. Vercel’s read was different. They moved those 10 people into higher-value roles. The agent took the deterministic part of the job. The humans went up the stack.
That distinction matters for how you sell this internally. This is not “fire your team.” This is “stop having your best people do the part a workflow can do better 100% of the time.”
The Build Method Is a Tripod
Every internal agent at Vercel gets built the same way. Three people, shoulder to shoulder:
- A GTM engineer
- A data scientist
- The single best subject-matter expert for that exact function
They document the best practice for the function first, then encode it into workflows that become the agent. A human stays in the loop to QA every output. The agent does not autonomously execute. Over time, as the subject-matter expert runs out of feedback to give, you finally pull the human.
For the lead agent, this was literal. A GTM engineer shadowed Vercel’s best SDR for days, watching every single tab she opened. LinkedIn, BuiltWith, the company website, the CRM, Slack history. The engineer turned each of those into a step in a tool-calling workflow. They documented the whole thing and made it work as a deterministic workflow before any AI touched it.
Then the agent ran in shadow mode in production for six weeks. The best SDR reviewed every output and fed corrections back in. The flywheel ran until she could not improve it anymore. The architecture mirrors exactly what she did manually, except now it performs like a 90th-percentile rep 100% of the time.
A single engineer prototyped the first version over a weekend. It was in production six weeks later. They then ran the same framework across 30 different SDR workflows, from event follow-up to product-qualified-account flows to time-based campaigns. SDR quotas went up 30% that quarter.
The build was not hard. The discipline was. Document the human, then encode the human, then QA the agent until it beats the human, then remove the human.
The Three Things That Actually Matter
Jeanne pulled out three takeaways for anyone building agents, whether for your product or for your own internal teams.
1. Agents Need Headless, Composable Architecture
Agents do not live in UIs. They call APIs. They hit MCP servers. They consume webhooks. If your product is not composable and developer-accessible, you are invisible to agentic workflows. You are simply not in the stack.
This cuts both ways. Internally, Vercel’s “Deal One” meeting intelligence agent only works because Gong has an accessible API and Salesforce has webhooks you can actually compose into automated workflows. Jeanne’s line on the tools that did not have those surfaces: they would have ripped them out.
Deal One ingests every call, generates notes with action items, posts coaching suggestions into Slack, proposes CRM field updates, tracks competitive mentions and objections over time, and runs postmortems on every closed-lost deal. When a rep @-mentions it in a Slack channel, it searches Slack in real time, queries a second agent for data, pulls Gong transcripts, hits the internal knowledge base, and streams the answer back. The rep never leaves Slack. The agent never has a UI.
Same with the Playbook Platform. A signal fires, a usage spike or a high-intent visit to the pricing page, and the platform matches it to a play, generates personalized outreach, and surfaces it to the rep for a one-click review. The best reps’ instincts, made available to everyone, triggered automatically. No separate tool to log into. Signals flow in, drafted outreach flows out, reps review in the workflow they already use.
The to-do here is simple even if the work is not: go build the developer surface area now. MCP servers, webhooks, APIs. Box, Notion, and Salesforce are all making major bets on their developer platforms for exactly this reason. The companies whose products can be called by an agent will be in the stack. The ones that can’t, won’t.
2. Invest in Your Data Foundation
Good data equals good agents. None of the clean-warehouse, semantic-layer, knowledge-base work is fun to build, but all of it is load-bearing.
Vercel’s most popular internal agent is D0, a data analyst agent the entire company reaches through Slack. Questions that used to take a week to ticket through the data science team now get answered in under a minute. Things like: what was our token volume through AI Gateway last week? Which companies do we have logo rights for in the UK? Who are our highest-revenue startup customers using sandboxes?
D0 translates plain-English questions into SQL against the analytics infrastructure, so anyone can get an answer without writing code or waiting on the data team. To power it, Vercel built a structured, queryable knowledge base, a semantic layer sitting on top of a model of their revenue, broken into the smallest causal units and enriched with first-party and third-party signals.
That layer is what makes Deal One’s analysis actionable instead of just interesting, and it is what lets the Playbook Platform contact the right company with the right message. Every agent runs on top of it. Without it, agents hallucinate or give generic answers. With it, they are grounded in your actual business.
There is a second-order effect here too. Put that data in front of smart people and the data itself gets better, not just the agent using it.
3. The Build-vs-Buy Calculus Has Flipped
With the rise of B2B software, enterprise software was a procurement exercise. You bought because building was too slow and too expensive. That assumption is gone. Robust applications and agents can be built and run at scale in a matter of weeks.
The cost numbers make the point. The lead agent: under $5,000 a year, 32x ROI. Vertex, the in-house customer service agent that powers the help site, costs $300 a month in infrastructure plus about $12,000 in tokens. Call it $150,000 a year, with three engineers, handling thousands of technical cases a week. Vercel started with an off-the-shelf tool, did not see enough results, and built in-house in two months.
Compare that to some of the agentic-support companies running 150 engineers on equivalent workflows at dramatically higher cost. The difference is not magic. It is that Vercel used infrastructure built for agents from the start, instead of discovering the cost problem after committing to an architecture.
But building is not automatically the answer, and most people skip this part. Every quarter Jeanne asks her GTM engineering team the same question: has someone out there built something better, faster, or cheaper than what we already have? If the answer is yes, the in-house agent goes. The new build-vs-buy is not “always build.” It is “your average builder should be able to outship any vendor selling you the same outcome, and you should keep checking whether that’s still true.”
The Part Nobody Wants to Talk About: Scale Breaks Things
Most teams have not run anything at real scale yet. The MCP server gets traffic, the internal agent is in an experimentation phase, the AI workloads look impressive. That can give you a false sense of readiness.
Production scale is what reveals whether your architecture is durable. When an agent that was running at 1x suddenly runs at 100x or 1,000x, you hit infrastructure problems and cost surprises you never anticipated. The cloud infrastructure we all grew up on for 20 years was built for a simpler request-and-response world. Agents are different. They think, they call LLMs with long-duration execution, they call tools. Those need different infrastructure underneath them.
This is why Vercel built what they call agentic infrastructure and runs all of these agents, lead agent, Playbook Platform, Deal One, D0, on Fluid compute. Fluid only triggers compute when needed and reuses existing resources before spinning up new ones. Early adopters cut compute costs by up to 85%. The point for your team: infrastructure choices matter far more than most teams realize at the start, and the bill arrives at scale, not in the demo.
One signal of where this is all heading: one out of three deployments from Vercel’s own customers now comes from an agentic engineering tool like Claude Code or Codex. The code is increasingly being written by agents. The infrastructure has to meet it there.
Go-to-Market Is Turning Into Consulting
The strategic shift underneath all of it is bigger than headcount. The executives Jeanne talks to are not asking for a pitch on Vercel. They are asking her to help them figure out where to start with agentic AI, what best practices look like, what the architecture should be, and what breaks at scale. GTM is moving closer to consulting than to selling.
That is the real story of running a company on agents. It is not a cooler tool. It is a different shape of company. Headcount an order of magnitude smaller for a given function. Humans reviewing agent output instead of producing it. Engineers becoming shepherds. Marketing and sales teams shipping software.
Treat Your Go-To-Market Like a Product, Then Automate It Piece by Piece
The companies that win this are the ones that treat go-to-market the way they treat product. They document it, design it, iterate on it, and then automate it one workflow at a time. That is the whole method. Shadow your best person, encode what they do, QA it until it beats them, pull the human, then ask every quarter whether you should keep it or replace it.
The gap is opening now. A single engineer prototyped Vercel’s lead agent over a weekend. The teams that start carving out a few hours a week to build will be on one side of the shift to agents. The teams waiting for it to feel safe will be on the other. The delta between those two groups is going to be the widest in B2B, and it compounds every quarter you wait.
The 5 Mistakes Vercel Says Will Bite You
Jeanne was direct about the traps, including the ones Vercel walked into before getting this right.
- Buying off-the-shelf first and assuming it’ll work. Vercel started Vertex, their support agent, on a third-party tool. It did not deliver. They ripped it out and built in-house in two months. The lesson is not “always build,” it’s “don’t assume the vendor solved your workflow until you’ve seen it perform on your actual cases.”
- Committing to an architecture before you understand the cost at scale. The cost problem shows up at 100x and 1,000x, not in the pilot. Teams pick an architecture that works in the demo, then get hit with infrastructure bills they never modeled once usage climbs. Pick infrastructure built for agents from the start, or you pay to re-platform later.
- Mistaking experimentation for readiness. An MCP server getting traffic and an internal agent in an experimentation phase can feel like proof. It isn’t. Production scale is the only thing that reveals whether the architecture is durable, and most teams haven’t run anything at real scale yet.
- Letting agents drift with nobody watching. Things break quietly. An agent slowly stops doing its job and nobody notices until a number moves. Someone has to own an hour or two a day reviewing output, even after you’ve pulled the human from the loop.
- Building on a weak data foundation. Skip the knowledge base, the clean warehouse, and the semantic layer and your agents hallucinate or give generic answers. None of that plumbing is exciting to build, but the agents are only ever as good as the data underneath them.
