The numbers that should make you uncomfortable
Vercel took a 10-person inbound SDR team down to 1.25 people. The AI agent that replaced them costs about $5,000 a year in infrastructure and tokens. That's a 32x ROI on the headcount swap alone.
Jeanne DeWitt Grosser, Vercel's COO (previously ran GTM at Google and Stripe for roughly a decade each), stood up a dedicated GTM engineering team in June 2025. By August, they had a lead qualification agent live. The initial prototype took a weekend. Full deployment: six weeks.
After implementation, SDR quotas went up 30%. Not because the remaining human worked harder. Because the agent handled the volume and the human focused on complex deals.
Why this isn't just a Vercel story
The macro data tells you this was coming. SDR conversations have dropped 55% since 2014. Generating a dollar of new ARR costs 70% more than it did in 2021. Only 50% of teams hit pipeline goals. And by 2023, 79% of buyers were initiating vendor engagements themselves rather than responding to outbound.
The "tooling replaces headcount" trend accelerated hard in 2022–2023. Growth-stage B2B SaaS companies saw RevOps headcount jump from 2 to 8 while sales headcount barely moved (30 to 32). The infrastructure people are replacing the dialing people. That shift is structural, not cyclical.
But here's the trade-off nobody wants to say out loud: most companies can't replicate this. Some tech companies were still actively hiring SDRs through 2023 and into 2024. The prerequisite isn't buying an AI tool. The prerequisite is having your data, workflows, and documentation in a state where an agent can actually operate.
How Vercel actually built it
Three roles per agent build: a GTM engineer, a data scientist, and a subject-matter expert. That composition matters because the failure mode for most AI agent projects is treating them like software installs rather than operational redesigns.
Vercel's methodology follows a specific sequence. First, document the best practices a human SDR follows. Encode those into workflows. Then automate. The agent runs in shadow mode (outputs reviewed by a human) until it consistently outperforms the human baseline. Only then does it go live.
The architecture is headless and composable. Their Deal One agent plugs into Gong and Salesforce via APIs and webhooks. No UI dependency. Their data analyst agent, D0, answers internal queries in under a minute using a structured knowledge base. The whole thing is built on what Vercel calls Fluid compute, which they say cuts infrastructure costs by up to 85%.
They also automated 93% of customer support cases and 96% of major content updates last quarter. The lead agent wasn't an isolated experiment. It was part of a systematic approach to treating GTM like a product.
The five mistakes Vercel flagged (and the one that matters most)
DeWitt Grosser's team called out five common errors: assuming off-the-shelf solutions work out of the box, committing to costly architectures before testing at scale, confusing experimentation with production readiness, neglecting ongoing agent oversight, and building on weak data foundations.
The last one deserves the most attention. Stack consolidation is a prerequisite, not a nice-to-have. Without unified data and integrated workflows, adding AI agents increases fragmentation rather than reducing it. Experts tracking this space in 2023 emphasized the same point: if tools aren't intuitive and data isn't unified, adoption drops and the promised efficiency gains evaporate.
Over 30% of sales activities can be automated. That's the opportunity. But the gap between "can be" and "actually is" sits in your data layer.
What to measure before you try this
If you're a CMO or VP Marketing looking at this and wondering whether your org is ready, start with a diagnostic, not a vendor demo. Three questions worth answering first:
- Data readiness: Can you document the exact workflow your SDRs follow for lead qualification today, including decision criteria, handoff rules, and edge cases?
- Stack fragmentation: How many tools touch a lead between first signal and AE handoff? If it's more than four, consolidation comes before automation.
- Baseline cost: What does generating a qualified pipeline opportunity actually cost you in SDR comp, tooling, and management overhead? You need this number to calculate real ROI, not vendor-projected ROI.
The hypothesis (make it falsifiable): if you document and encode your top SDR's qualification workflow, then deploy an agent in shadow mode for 30 days, qualification accuracy should match or exceed human performance at less than 20% of the cost. If it doesn't, the bottleneck is data quality or workflow documentation, not the AI.
Vercel didn't start with the agent. They started with the documentation. DeWitt Grosser's team wrote down how their best reps worked before they automated anything. That's the part most teams skip, and it's the part that determines whether you end up with a $5,000 replacement or a $5,000 distraction.
