How Vercel Handed Inbound to One Human and Sent Everyone Else to Outbound
By Creatives Takeover Editorial Team · October 22, 2026
Copy your best person before you automate them.
In 2025, a wave of AI startups promised founders they could fire their sales development teams and let software book the meetings instead. Most of those promises fell apart within a quarter. Vercel, the company behind Next.js, quietly did something different: it took a team of 10 SDRs working inbound leads down to one person in six weeks, held its conversion rate flat, and then sent the other nine people out to do outbound. The headline was the headcount. The real story is how they got there, and why it worked when so many others didn't.
A Contact Form Nobody Wanted to Own
Before any AI touched it, Vercel's inbound process looked like most startups' inbound process: a public "contact sales" form and a team of humans sorting through whatever came in. Drew Bredvick, the GTM engineer who later built the fix, described it bluntly as a disaster. The form was public, so spam poured in. Leads got routed to whoever was "next" in line regardless of time zone, which pushed response times out to 24 or 48 hours.
The people doing the work weren't the problem. The job itself was. Ten SDRs were spending more of their day sorting junk and researching companies than actually talking to customers, and they were getting pinged at all hours to decide whether a lead was even worth working. Every hour spent on that sorting was an hour not spent selling.
This is the part most "AI replaces SDRs" pitches skip over. The SDR model was already straining before AI arrived. Buyers were slower to respond, inboxes were fuller, and the most expensive people in the funnel were doing the most repetitive part of it.
They Copied Their Best SDR Before Writing a Line of Code
In June 2025, six weeks into her role, Vercel COO Jeanne DeWitt Grosser set up a go to market engineering team with a single mandate: bring agents to every part of how the company sells. Grosser had spent roughly a decade each running go to market at Google and Stripe. The first project was the inbound mess.
The team didn't start with a model. Bredvick spent about a week shadowing Vercel's top performing SDR, watching every tab she opened: LinkedIn, BuiltWith, the company's website, the CRM, old Slack threads. What made her excited about a lead? What made her drop one instantly? What did she put in emails that got replies? Each of those habits became a step in a workflow, and the workflow worked as a plain, predictable process before any AI was added to it.
Only then did the agent come in. Today, when someone fills out the contact form, the lead is enriched with company data, the agent researches the company and the person in depth, and it sorts the lead into a bucket: qualified, needs follow up, actually a support question, or not sales related at all. It then pings an SDR in Slack with its reasoning attached, and for qualified leads, a draft reply is already waiting. The SDR's job went from research, qualify, write, and send to review and press send.
The first version was prototyped over a single weekend. It launched in August 2025, then ran in "shadow mode" for six weeks, processing every real lead alongside the humans while the best SDR reviewed each decision and fed corrections back in. When she ran out of things to correct, the team went from 10 people to one in the US, plus about a fifth of one person's time covering Europe and Asia Pacific.
Why It Worked When the AI SDR Startups Didn't
At almost the same moment, the most hyped company in the space was learning the opposite lesson. 11x raised around $74 million from Benchmark and Andreessen Horowitz on the promise of an AI "digital worker" that could run outbound with no human involved. In March 2025, TechCrunch reported that the company had shown logos of businesses that said they weren't customers, including ZoomInfo, which said the product performed significantly worse than its human SDRs during a one month trial. 11x disputed several of the claims, but the category's reputation took the hit.
The broader numbers tell the same story. Industry analyses through 2026 estimate that 40 to 60% of AI SDR pilots were shut down within 90 days, and that AI booked meetings turn into qualified opportunities at roughly 15%, compared with about 25% for meetings booked by humans.
Vercel's agent beat that pattern for three reasons. First, it pointed AI at inbound, where the buyer had already raised their hand, instead of blasting cold outbound at strangers. Second, it never acted alone: every decision passed through a human before a single email went out. Third, it wasn't trained on a generic sales playbook. It was a copy of one specific person who was already the best at the job. The result held Vercel's lead to opportunity conversion rate flat, which Grosser summed up as the agent being as good as the humans were.
The cost is the part that made people stop scrolling. SaaStr reported the agent runs on about $5,000 a year in infrastructure and tokens, plus a fifth of one engineer's time to maintain. Bredvick's own fuller estimate, counting engineering hours, is about $60,000 a year against more than $2 million in savings, which is where the widely quoted 32x return comes from.
Nobody Got Replaced. The Job Moved.
When the 10 to 1 number went public, plenty of people read it as the end of the sales job. Vercel's own account is different. The remaining SDR now spends the day checking the agent's work, and the other nine were moved to outbound, the harder and more human part of the funnel where relationships are built from zero. That same quarter, Vercel applied the method to about 30 other SDR workflows, from event follow ups to campaigns triggered by product usage, and SDR quotas went up 30%.
The outbound side changed too. Vercel now runs what it calls a Playbook Platform: when a signal fires, such as a spike in product usage or a visit to the pricing page, the system matches it to a play, drafts personalized outreach, and hands it to a rep for a one click review. The agent does the timing and the drafting. The human decides whether to send.
The person who made all this possible holds a job title that barely existed two years ago: GTM engineer. Bredvick built the lead agent with about 25 to 30% of his time. Vercel now builds every internal agent with the same three person team: a GTM engineer, a data scientist, and the single best expert at the job being automated. The rule they follow is simple. Document the human, encode the human, check the agent until it beats the human, and only then step the human back.
What This Actually Means for a Founder Building Right Now
The tempting lesson is "buy an AI SDR and cut your sales costs." That's precisely the lesson the 2025 wave of failed pilots taught founders not to learn. What Vercel showed is that the value isn't in the tool. It's in knowing exactly how your best seller does the job before you ask software to do it.
If you're an early stage founder, you are probably your own best SDR. That's an advantage, not a limitation. Before you automate anything, write down what you actually do when a lead comes in: which pages you check, what makes you excited, what makes you walk away, and what you say in the first reply that gets a response. That document is the agent. Everything else is plumbing, and Vercel even open sourced its lead agent template so you don't have to build the plumbing from scratch.
Then point automation at the boring, repetitive part of the job, keep yourself in the loop until the agent stops making mistakes, and spend the time you win back on the part no agent can do yet: real conversations with people who haven't heard of you.
Five Things Worth Taking From This
The best training data for an agent is your best person, not a playbook. Vercel's agent performs well because it copies what one top SDR actually did, tab by tab, rather than what a sales manual says should happen.
Inbound is where AI earns its keep first. The leads had already asked to talk. Pointing automation at warm demand worked, while autonomous cold outbound became the category's biggest disappointment.
Keeping a human in the loop isn't a weakness, it's the method. Six weeks of shadow mode, with every decision reviewed, is what let Vercel trust the agent enough to step the humans back.
The win wasn't fewer people, it was people doing better work. Nine SDRs moved to outbound, quotas rose 30%, and the remaining SDR's job became reviewing instead of sorting.
Agents are cheap. Clarity is expensive. The compute costs about $5,000 a year. The hard, valuable part was documenting exactly how great work gets done before anyone wrote code.
While much of the industry spent 2025 trying to build a salesperson out of software, Vercel did something less glamorous and far more effective. It watched its best salesperson closely enough to turn her judgment into a process, let an agent run that process under supervision, and gave everyone else their time back to sell. The lesson isn't that sales teams are shrinking. It's that the people on them are finally being asked to do the part only people can do.