Years ago, I heard the late David Graeber speak about his Bullshit Jobs theory at a San Francisco bookstore. The audience, being the crowd that shows up at City Lights on a weekday evening, had not come to be persuaded so much as to nod along. They wanted to believe that vast categories of white-collar work, particularly the kinds they did not do themselves, were fake. If administrators, lawyers, and compliance teams, along with managers and middlemen of every kind disappeared tomorrow, the world would keep spinning, and spin better.

The mistake in both Graeber's critique and today's AI-flattened-org fantasy is the same: confusing work that produces no visible artifact with work that produces no value.

Anyone who has worked inside a company with more than a handful of people has seen the ritual forms of Graeber's work rant: meetings whose only output is other meetings, dashboards no one acts on (or even looks at), and strategy docs that work backward to memorialize decisions already made. Organizations do create pointless work. Status games turn headcount into prestige, so managers empire-build and find ways to keep their teams busy.

The interesting question is not whether organizations create fake work. Obviously they do. The question is why so much fake-looking work persists. My answer: The world is adversarial, and that makes some work irreducible.

Your company exists in a repeated game against competitors, regulators, customers, vendors, and even employees. It might be true that, in a better equilibrium, every company would need fewer lawyers. But if your competitor has lawyers and you do not, that does not mean you are living in that better equilibrium. It means you have volunteered to be the sucker at the table. The pressure that creates defensive waste also forces companies to become more creative, as your counterparts' optimization becomes your constraint.

Similarly, scale creates a second category of what Graeber sees as wasteful. Even inside a company full of smart, high-agency, good-faith people, coordination does not stay free for long. As you scale, Metcalfe's law kicks in, and the reporting chain defines what you can ship. When you want to change course, you need to bring people with you, and managers grease the wheels of the org so things move smoothly.

This is why the Silicon Valley recurring idea that “there are no managers, everyone is an IC” feels half-right and mostly wrong.

Sure, call everyone “Member of Technical Staff” if you want. Titles are cheap, but a shadow org will define your team dynamics. Someone decides on the roadmap, and approves or denies resourcing asks. Someone has to resolve conflicts between teams, coach the junior engineer to steer away from shiny objects, or give hard feedback to the brilliant staff engineer whose architectural judgment is matched only by their ability to corrode morale.

AI makes this confusion more tempting because it blurs roles once defined by scarce expertise. Engineers can now prototype product ideas without waiting on design, or dig into another team's implementation instead of stopping at the API boundary. PMs can query data and generate v0 specs. This is good. The old boundaries were often artificial. But collapsing boundaries does not eliminate the need for coordination. One person can now be PM, designer, and engineer, but coordination just moves outward, between these smaller, faster units. You can delay it until you have more scale, but unless you stop yourself from hiring, you cannot make coordination work fully disappear.

We have seen this pattern before: when production gets cheap, filtering gets expensive. The internet drove the cost of communication toward zero. It did not eliminate the need for editors, brands, taste, or trust. In fact, once everyone could publish, attention became the scarce resource. AI is producing a similar effect on building. As the cost of generating code, mockups, analysis, and product variants falls, the bottlenecks move elsewhere: judgement, sequencing, customer trust, and, again, attention.

A startup can cheaply create many versions of itself and grid-search over huge product diffs in parallel. It can generate thousands of landing pages, features, flows, emails, and experiments. But your customers are not infinite. They don't have enough patience. Every bad product swing has a reputation cost. Every confusing launch burns trust. Every “quick experiment” exposes real users to the internal chaos of your org. So you still put a team with judgement between your AI and your customers.

And, for the same reason, you put leadership between your employees.

People have limited tolerance for strategic thrash. Useful managers are filtering mechanisms. Managers are shock absorbers, taking on ambiguity before it gets sprayed across the team. They decide which signals matter. They protect attention. They coach people through the parts of work that are not reducible to tickets: conflict, motivation, fear, ambition, resentment, trust.

“But the best leaders are vibe-coding,” you say. Fine, you all want to be Tobi. But the reason executives and managers need to use these tools is not management turning into coding. It is managers being too far from the frontier, and that frontier moving faster than ever. Leaders need first-hand knowledge of what the tools can do, where they fail, and how jagged they really are.

Managers should be close enough to the work to have independent judgement, but detached enough from the critical path to make everyone else better. Maybe we will all spend less time on perf reviews once agents can summarize Slack conversations, GitHub commits, and sales calls. Great. But not all work becomes automatically legible to the machine. Counting lines of code and PR volume is strangely in vogue, but metrics get gamed and not every machine-readable signal is the right one. The best managers will use AI to get closer to the work while spending less time as human routers.

The most intelligent AI systems will not make people stop being people. They do not remove misaligned incentives. They do not make customers more patient, teams more honest, or strategy more obvious.

AI will destroy much managerial theater. It will not destroy the work. The job is still the same old difficult thing: turning human confusion into coordinated progress.

Thanks to Hannah Doherty, for her comments on early drafts of this essay.
Photo: Building and Rebuilding, by me. Previously posted in Budapest, 2025.