About sixteen minutes into a podcast recorded in June, Guillermo Rauch describes a rule that most CEOs would never say out loud. “No one at Vercel has the right to block things,” he says. “The tokens must flow. The products have to be shipped. Customers need help. No one is here to say hold the production lines. Only security can do that.”
Notice which way the default runs. In most companies, work waits for a senior yes, and silence means no. At Vercel, work ships unless someone actively says no, and the effort of saying it sits with whoever wants to say it. An executive who wants to stop a launch has to read it, form a view and act before it goes out. Rauch applies the same rule to himself. He is copied on launch emails as an FYI, not as an approver. “You don’t get to relax,” he says. “Send me the email, and when I get time to review it, your product goes out.”
He is the founder of a company valued at $9.3 billion at its last round, with close to 800 people, whose infrastructure sits under OpenAI and Stripe. Of everything he could have chosen to explain about how the company runs, the rule he reaches for first is who is allowed to say no.
What caught me was not the rule itself. It was recognition. I have been building a version of this in my own work for the past year, agents that take the requests I used to make of people, and I had not put a name to what it was doing to the shape of the operation. Hearing it described at an 800 person company, as a settled operating model rather than an experiment, made it click. Rauch is not describing an AI strategy. He is describing what a company is for once the people inside it no longer need each other to get work done.
The company was a market for other people’s time
Here is the line. Around fourteen minutes in, asked whether even non technical hires go through the build exercise, Rauch says his marketing team is expected to build the agents that produce marketing artefacts, and then he describes the problem that expectation removes. “A common problem in marketing organisations is, can I beg, borrow and steal engineering resources or agency resources so that I can ship an idea into the world. That probably doesn’t exist ever, because you have the platform, you have the agents, and because you’re expected to create that factory. You’re not here to file Jira tickets.”
Read that back against your own week. How much of what you did was the work, and how much was the process of getting someone else to do the work? The brief you wrote so an agency could produce the thing. The ticket you filed so an engineer could change the thing. The message you sent to the analyst so you could get the number. The meeting you booked so three people could agree that a fourth person should do it. In most companies I have worked in, including large and well run ones, that second category is the majority of a knowledge worker’s day. The visible job is a role. The real job is procurement of other people’s hands.
That is what a company was, structurally. A market for borrowed time, with a currency of favours, seniority and budget, and a settlement system called the org chart. The people who got the most done were the ones who were best at extracting time from others. Being good at your job meant being good at asking.
Rauch’s second line closes the loop from the other side. He says he routes questions and research to the internal agent first, and then: “If I DM you, it probably meant that my agent couldn’t figure it out.” He frames it as respect. When he reaches a human, he is spending their attention on something a machine could not resolve. But notice what it does to the market. If the CEO no longer borrows your time for the things an agent can do, and the marketer no longer borrows the engineer’s, and nobody borrows the analyst’s, the trade that organised most of the company has stopped. What is left is the work, and the people who can do it.
That is the shift I want to map: from an organisation whose main internal activity was requesting, to one where every person is expected to deliver, because the means to deliver are on their desk. Speed is a side effect. The change is in what people do all day.
The org chart was always a map of who could stop you
An org chart officially shows who reports to whom. In practice, it shows who can stop what. Every box above you is a person whose sign off you may need. Every dotted line to legal, brand, finance or platform is a place where your work can stall for a reason that has nothing to do with whether it is good. The chart is a record of veto rights dressed up as a hierarchy, and it is also the price list for the market of borrowed time. Seniority told you whose request came first.
That was a rational design when the scarce thing was execution. If building took months and shipping was expensive, you wanted multiple people able to catch a mistake before it reached a customer, and you needed a queue, because there were more requests than hands. The veto was a quality mechanism and the hierarchy was a scheduling mechanism. Both cost you speed and bought you safety and order.
What Rauch is describing is a company where that trade no longer makes sense, because the platform, the hiring process and the agents already set the floor, and the queue has collapsed because most requests never reach a person. His phrase for the value is iterative greatness, and the mechanism beneath it is that the default state of any piece of work is out, not held. Only one function is allowed to break that default. Everyone else has to earn a block by doing the review themselves, quickly, with their name on it.
I wrote earlier this year about what replaces the org chart when agents arrive. This interview is the clearest worked example I have seen of the answer. The chart does not get flatter. It gets replaced by three design decisions: where the floor sits, where the front door is, and where the one gate lives.
Raising the floor before anyone is hired
Rauch says the requirements for a hire change weekly, so he stopped interviewing for requirements and started interviewing for output. Every candidate, technical or not, gets what Vercel calls a work prompt. Build something. Come back. Talk through it. He is watching for the same things you would watch for in a colleague: attention to detail, the side quests they took, whether they came back with an opinion about what was broken. He describes a candidate who delivered a strong piece of work and then told him where she had lost an hour to a rough edge in the product. That is the moment he was interviewing for. She arrived with a view on what she wanted to fix, and she had not needed anyone’s time to form it.
The old company hired for a slot in the request market. A marketer who could brief well. An analyst who could turn tickets around. A product manager who could get engineering to say yes. The work prompt hires for something different: can this person take a goal and return a finished thing, using whatever is on the platform, without a queue. That is a different person, and Rauch is explicit that it is a different deal. He tells candidates the project takes a decade and the hours are long.
The floor is also codified. Vercel runs skills.sh, a public directory of instruction files that tell a coding agent how to build a Next.js application the way Vercel’s own engineers would. People asked Rauch whether he had lost his mind giving away the company’s expertise. His answer is that the expertise was never the moat. The platform and the trust are the moat. Publishing the skills means a healthcare company or a bank can get Vercel quality output from an agent without hiring a Vercel engineer, which brings them onto the platform, which is the business.
Internally the same instinct runs the other way. Rauch says he has spent years learning how to communicate a product succinctly, and now his question is how to transfer that into the internal agent so the factory keeps producing at the same standard without him. The design engineering team, who own polish and the small surprising details in the customer journey, are building the review agent that carries their taste, because they know they cannot hire their way to coverage. In the old company, taste was a person you had to get time with. Here it is a file that anyone can load. Rauch mentions that Meta shipped something similar alongside its newest model, a skill whose whole job is to stop the model behaving like a model. I run one myself. Every piece I publish passes through a file that lists the words and shapes I do not want, built by ear over months of catching the same failures. That file is a skill. It is the floor for this publication, and it is the least glamorous and most valuable thing I have built this year.
One front door, and what it does to people
Most companies tried to raise the floor by buying everyone a subscription. Rauch is direct about what happened next. One person builds a prompt library and follows the right accounts and collects skills. The person at the next desk clocks in, types a question, clocks out. The gap between the top five per cent and everyone else did not close. It widened, because the tools rewarded the people who were already going to use them. And the request market carried on underneath, because the person who did not use the tools still needed to borrow the time of the person who did.
Vercel’s answer was one internal agent, called V, and one rule: you go to V first. Questions about the business, questions about a customer, a job you want automated, research you want run. V sits in Slack. Underneath it is a collection of agents, applications, skills and a team that maintains it. The chief operating officer built a dozen go to market agents for lead routing and the operational rhythm of sales, and Rauch says he does not personally know what they are, because he does not need to. He asks V.
The comparison he reaches for is the phone in your pocket. You do not think about which application handles which task. There is one object and it does everything. That is the design target for the internal agent, and the reason it matters for the argument here is that a single front door is what makes the request market close. If there is one place everyone goes, and it can answer, then the question never goes looking for a person. Human attention becomes the expensive path, reserved for what the machine cannot resolve. That sounds cold until you notice what it does to the humans on the other side. They stop being the routing layer for questions a system could answer, and they start being asked only the things worth asking a person.
The data organisation is his proudest example, and it is the request market in miniature. He once told his VP of data that he wanted every employee to be able to file a ticket and get an analyst’s answer within an agreed service level. That was the old model at its best: a queue, a promise about the queue, and a team whose job was to work it down. Today the service level is time to first token, measured in the observability platform, and the ticket queue has been replaced by V. The team that used to answer questions now builds the machinery that answers them, including the permissions model for who can see what. He describes this as bending the headcount curve, and he is careful with the verb. Bending, not cutting. The analysts did not leave. They stopped being borrowed.
The one gate they kept
Which brings us back to the veto. If the floor is high and the front door is single, the remaining design question is where you place the one gate you cannot remove. Vercel put it on security. Rauch and his CTO spend real time studying that machinery: where the choke points are, what the throughput is, how frequently product improvements are being communicated to customers. He asks his agents to measure the change log rate. That is a founder treating the organisation as a system with a measurable flow, and choosing the single place where friction is allowed.
There is a phrase he uses that explains why. Everything is code. A slide deck is code. A document is code. A spreadsheet was always a sequence of tool calls and a short script, disguised by the fact that you dragged columns around with a mouse. Once you accept that, the organisation itself becomes something you can engineer, and the question of where to put the gate is an engineering question about latency and risk, not a political one about who is senior enough to hold things up.
Where this breaks
Vercel is a developer platform run by an engineer, selling to the most agent native customer base on earth. Rauch is upfront that the company had less of the adoption gap than the anecdotes he hears from customers, including one that invites AI CEOs to fireside chats to persuade staff to use the tools at all. A regulated business cannot make security the only veto, because legal and compliance are genuine gates rather than political ones, and pretending otherwise is how you get fined. The work prompt interview selects for a particular kind of person, and the deal Rauch offers is at least a decade of long hours. That is a culture, not a template, and it will not transplant into a company that promised its people something different.
The numbers are also vendor stated. Vercel reported at its June event that agent driven deployments had gone from under three per cent to more than half in six months, and that token volume through its gateway had risen roughly tenfold. Those are the company’s own figures about its own platform, and in the same interview Rauch points to a GitHub outage that day, which he attributes to agentic demand the infrastructure could not carry. The load is real. The infrastructure to hold it is uneven, including at the companies selling it.
There is a human cost in the shift too, and Rauch does not dwell on it, so I will. The request market was inefficient, but it was also where a great deal of belonging lived. Being asked was how you knew you mattered. A company where nobody borrows your time because a machine did not need to is a company that has to find another way to tell people they are needed, and most have not thought about that yet. Rauch’s own answer is that the DM now carries more weight, because it only arrives when a human was required. That may be true for a CEO. It is less obviously true for the analyst whose queue disappeared.
And there is an older tension I have written about before, the codification trap. When you turn a person’s taste into a skill file, you make the company less dependent on that person. Rauch frames this as a gift, and inside a founder led company with a decade long deal it may be one. In a company where the person who wrote the skill can be made redundant the quarter after, it reads differently. The mechanism is the same. The meaning depends on who holds the file.
The question to take into your own building
Here is the test I would run on any company, including your own. Take last week’s calendar and inbox and sort everything you sent into two piles: the work, and the requests for someone else to do the work. Most people who try this are unsettled by the ratio. Then look at the second pile and ask, request by request, whether anything on your desk today could have done it. Not perfectly. Well enough that you would not have needed to ask.
That ratio is your company’s borrowed time. It was installed when execution was scarce, when the only way to get a thing made was to find the person who could make it and get in their queue. Rauch has built a company where that queue is close to empty, and the three decisions that got him there, a codified floor, one front door, one gate, are available to anyone willing to make them.
The tooling is the easy part. The harder part is accepting that the skill you were promoted for, the ability to get things out of other people, is worth less every quarter, and that the skill replacing it, the ability to finish something yourself, was always the job underneath.
The full interview is Guillermo Rauch on Knuckle Up with Nakul Mandan. I have quoted the parts that serve this argument, and there is a great deal more in it about hiring, taste and what a platform is for. Well worth the hour and a half. Watch it here.
If this landed, subscribe. The next piece takes my own OpenRouter bill, line by line, and sets it against what the same work would have cost in salary.
Craig Hepburn is an AI strategist and Perplexity Fellow. Twenty years building at the frontier of digital, from Microsoft and Nokia to Art Basel and UEFA. Now building at the frontier of agentic intelligence.