把 GTM 当产品来做：Vercel 如何用智能体重写销售流水线

Grosser defines go-to-market broadly: any function that touches a customer or makes a dollar. At Vercel that means marketing, sales, sales engineers, post-sales platform architects, customer success, support and partnerships. She has seen these teams run on a "Venn diagram strategy," each pursuing slightly different goals and segmentation frameworks that overlap but never quite align.

She expects the recent wave of hyper-specialization to reverse. Someone has counted 17 distinct roles inside go-to-market, and she hypothesizes many will collapse. Her advice is to map the jobs to be done from first awareness through a five-year, wall-to-wall customer, then orchestrate that lifecycle the way you would design a product.

Two shifts stand out to her. Consumption-based pricing made sales more consultative, because the first land is only a small slice of the eventual relationship. AI has intensified that, since everyone knows they must change but not what to change to. That is why forward-deployed engineering, a cousin of professional services, has taken off: embed with the customer, help them build, and feed what generalizes back into the product.

Grosser tells the story of Project Rosland at Stripe in 2017. Stripe ran lean ("efficiency is leverage"), so she was given four SDRs where most companies would have hired thirty. Her team built a database of every company on the planet with attributes that would allow a fill-in-the-blank email template, such as pitching Stripe Connect to marketplaces. The false-positive rate was too high and it never fully worked.

Today she is rebuilding the same idea at Vercel and it works. The difference is that she now has a go-to-market engineer rather than a systems admin. That person helps encode the human workflow in code, part deterministic and part agentic, using Vercel's own workflow tooling. Some of the people involved in the 2017 effort now work at OpenAI and Anthropic, and one has founded an AI-native GTM startup.

The remit of GTM engineering, Grosser says, is to break each go-to-market function into workflows and turn the ones where AI beats a human into agents. Her team started with inbound because it is the most "legible," meaning replicable and mostly deterministic. Outbound follows, starting at the lower end of the market where there is a single decision maker. In large enterprises an agent might only do the research.

The prize is time with customers. Annual benchmarks have shown sellers spend about 30 to 40 percent of their time in front of customers for as long as she has been in sales. She hopes agents can take on research and follow-up so that figure rises to roughly 70 percent.

To avoid the flood of obviously robotic outreach, Vercel always keeps a human in the loop. The GTM engineer first shadows the best performer in a role, such as an SDR with seven tabs open across LinkedIn, ChatGPT and internal databases. The agent then makes two calls on an inbound lead: is it likely to be qualified, and what should we say. It does deep research and drafts a response, and a person reviews it and presses send.

Vercel had ten SDRs on inbound. Now one person effectively does QA on the agent, and the other nine moved to outbound. The lead agent was built by one GTM engineer spending roughly 25 to 30 percent of his time over six weeks. Throughout, the team tracked normal SDR KPIs and held lead-to-opportunity conversion flat, while the number of touches fell because the agent replies faster than a queue of leads waiting overnight.

Asked whether AI will replace SDRs entirely, Grosser says it can replace a fair amount of prospecting, since the average SDR was not doing sophisticated research. The last holdout will be deep enterprise prospecting across many layers of an org chart. She notes that nobody went to college to become an SDR. It was simply where you were forced to start, and she thinks people can move straight into outbound or SMB closing.

For non-sales listeners, Grosser explains the roles. SDRs generate pipeline, either inbound (people who contact sales) or outbound (targeting parts of the market that haven't raised a hand). Account executives are closers who move a buyer from interest to conviction. Their deals get more complex as they go up from SMB, with a single decision maker, to mid-market with economic and technical buyers, to enterprise with procurement and committees.

She says GTM engineering will push companies to be more rigorous about sales early. Founders often assume sales is just talking to people, but it is a skill built on discovery, handling objections and showing value. You can't write agents without a point of view on best practice, so a team of around ten people with a documented playbook is probably the right time to bring in a GTM engineer. She also advises adding revenue operations earlier than you think.

On hiring a first salesperson, she agrees with the common rule of thumb of roughly a million dollars in ARR. The real test is repeatability: if your customers look nothing like each other, you still have an evangelist, founder-led sale. Once you can write down an ideal customer profile, hand over the reins, but stay close to customers for the R&D that tells you where the product resonates next.

Vercel's first three GTM engineers were sales engineers who had been front-end developers. Grosser says GTM experience matters. When she reviewed the lead agent's emails she realized she would not have sent some, because it had been modeled on a top rep with two years of experience rather than someone with her twenty. Vercel also acts as customer zero, building these agents on its own workflow SDK and AI gateway.

Grosser's favorite stack item is Gong, which she says became much more interesting once you can run agents against it. Vercel feeds Gong transcripts into a bot that first reviewed lost opportunities. For the biggest loss of the quarter, the account executive blamed price. The agent, reading Slack, email and calls, concluded the team never reached an economic buyer and failed to demonstrate ROI.

The team then pulled the idea forward into a live "deal-bot" that posts into per-customer Slack channels, for example noting that a deal is far along without having spoken to an economic buyer. Because Vercel ships so fast that enablement is hard, they also run the agent across calls to see where objection handling failed, and then treat those as bugs to fix in weekly sprints with new content, discovery guides and demo changes.

On build versus buy, Grosser says these agents are neither hard nor expensive. The lost-deal bot took about two days, or 40 hours. The lead agent runs on Vercel for around a thousand dollars a year, against well over a million in salaries for the ten SDRs it displaced. She argues that your own esoteric context and workflow are the key ingredient, so it is worth experimenting in-house before settling on platforms. She also warns buyers are ending up with twenty tools for twenty jobs, and she heard the term "experimental run rate" for revenue that may not renew.

Her GTM engineers have computer science degrees and build directly on Vercel's AI gateway, sandbox, workflows and fluid compute. Vibe-coding and drag-and-drop workflow builders exist too, but she credits the speed to the platform itself. "Vercel builds Vercel with Vercel," she says.

Grosser's first job was on Gmail in 2004, when its technical lead over rivals lasted more than a year. A decade later cloud computing had narrowed such gaps, and she began to believe that the experience of being sold to would increasingly drive buying decisions. At Stripe, discovery calls were replaced by a whiteboarding session where customers drew their payment architecture. Stripe learned what it would have to displace, and customers left with an asset and a sense of a collaborative partner.

A second principle is adding value at every touchpoint regardless of whether the customer buys. After nine years at Stripe she watched many lost customers return half a decade later. At Vercel, outreach includes insight into a site's performance in absolute terms and against peers, since speed affects SEO, which feeds answer engine optimization. Even if the recipient never buys, they learn something real.

Lenny sums it up as thinking about every step of the prospect's journey the way a product manager thinks about flow, and Grosser agrees: it should feel like an experience rather than a transaction.

For tactics, Grosser says to invest in data that surfaces unique insights about a customer's suboptimal state. She also thinks companies stop at docs when larger customers want blueprints, like AWS's well-architected guides. Stripe had seen nearly every major marketplace, so a marketplace buyer didn't want generic documentation but the best way to set up marketplace payments.

She also stresses discovery. Founders get excited and jump to fixing problems, whereas excellent salespeople usually talk well under half the time, ask "five whys" and respond to a question with a question. At Vercel the insight content includes Core Web Vitals benchmarking, guidance on when an MCP server makes sense, and AEO. Even when it doesn't lead to a performance sale, it builds trust.

Grosser relays a round-number statistic that about 80% of customers buy to avoid pain or reduce risk, while only one in five buys for upside. Visionary pitches resonate with other founders, but enterprises worry about missing next quarter's target, being outdone by competitors or brand damage. Lenny adds that this matches what April Dunford has said about buying being a career bet. Grosser concedes it feels off-brand for startups but says it drives more buying behavior.

Grosser defines segmentation as carving up the world of companies by how they buy. Small, medium and large is the typical start, but she adds attributes specific to the offering. At Stripe the second axis was growth potential, because it was a consumption business and a company growing 200% a year is worth more than one growing 8%. It also cut by business model, as B2B, B2C, platforms and marketplaces each need different products.

At Vercel the same size and growth axes apply, plus traffic as seen in Google's Crux data and workload type. OpenAI, for example, is mid-market by headcount but a top-25 traffic site, which pushes it into enterprise-style selling. An e-commerce sale uses different language from a crypto company running on AWS. She built the framework in her first 30 days by asking the head of data science what predicts a hundred-thousand versus million-dollar customer, and where the company keeps winning.

The analysis showed deep e-commerce penetration but little in enterprise SaaS, where migrating millions of lines of code is a big lift, though SaaS companies are now adopting the AI cloud. Lenny's rule of thumb is three attributes, and Grosser agrees that beyond three it gets too detailed for a handful of sellers. She also insists segmentation is a company concern: she presents it to every new hire so product managers know who they're building for.

Grosser's litmus test for her account executives is that, put in front of ten engineers, it should take ten minutes to realize they aren't a product manager. That requires real product depth, which brings credibility and also lets sales act as R&D. A twenty-person sales team talks to many customers every week, so turning that feedback into signal for the roadmap extends product management, provided reps can tell an objection from a market opportunity.

She wants sales teams that think like a general manager: knowing when to say no, when to handle an objection, and when something heard on three calls is a real unlock. Company strategy, she says, is product strategy meeting go-to-market strategy, so a revenue leader needs to engage with product and pricing and not only with selling.

Grosser says PLG remains relevant for most companies unless they target large enterprises from day one, as Sierra does. It drove growth at both Stripe and Vercel. But it has a ceiling, because people will not spend a million dollars through self-serve, and companies stall when they wait too long to add sales. A replicable sales process takes a while to build, and outbound takes time to become predictable. She cannot think of a 100-billion-dollar company that is PLG-only.

On pricing she says to think of it like a product: understand where customers get value and where you incur cost, and align the two. Many companies underprice or default to freemium without it being a strategy. Stripe killed a free trial for Billing because integration effort already meant customers would stay.

At Vercel, a pricing change in August unbundled the Enterprise SKU after the team noticed about half of its users were startups wanting certain features. Moving those features to self-serve drove growth in the PLG funnel. In her view, getting those knobs right serves both customer experience and revenue.

Her hot take on compensation is that pay for performance is good but makes an organization less flexible, since you decide twelve months ahead what you value. When Vercel wrote this year's sales plans, the AI cloud did not exist yet and it launched mid-year. She wants the motivating upside of quota-based pay with the ability to change her mind, especially as annual planning gets harder.

On hiring, she favors a diversified portfolio: people with real sales experience paired with nontraditional backgrounds like consulting or banking. Those hires bring quantitative and consultative strengths and quickly learn sales is a skill, while AEs learn to think about P&Ls, talk to CFOs and present TCO analyses. The mix creates a richer learning environment.

In the lightning round, Grosser cites her mother's sayings, "when the going gets tough, the tough get going" and "where there's a will, there's a way," as useful in every quarter you're off pace. A former college diver who says she usually came third of three on her team, she notes that diving taught repetition, precision and getting straight back on the board. She also recalls a trainer's line: "yeses are great, nos are great, maybes will kill you."