A lot of companies are pushing AI transformation right now. Step one is usually the same: buy a bunch of coding agent licenses, then roll out training and hope employee productivity goes up.

I think that math is fake.

If the org chart stays the same, trying to boost everyone’s output with training and AI tools is not realistic.

The AI Companies That Win Are Small and Sharp

Midjourney started with 11 people and hit $200 million ARR within a year. No funding. No paid acquisition. No sales team. That’s $18 million in revenue per employee.

Cursor had around 50 people early on. By early 2026, it had already reached $2 billion ARR with a $29.3 billion valuation. Even if you use its current headcount of 150, that’s still $6.7 million per employee. A normal SaaS company does maybe $150,000 to $250,000 per employee. Cursor is 30 times above the industry average.

But Anthropic itself is even crazier, especially the Claude Code team.

Claude Code writes 90% of its own code.

Boris Cherny, the founder of Claude Code, said on Lenny’s Podcast in February 2026 that he had not handwritten a single line of code since November 2025. Claude Code wrote 100% of it. He ships 10 to 30 PRs a day.

Across Anthropic, average engineer output is up 200% over the past year. Code written by Claude Code already accounts for 4% of all public commits on GitHub. Boris said it plainly on the podcast: coding is basically solved.

The numbers from other companies using Claude Code are even more extreme. At Stripe, one team migrated 10,000 lines of Scala to Java in four days. The original estimate was 10 engineer-weeks. At Wiz, migrating a 50,000-line Python library to Go took 20 hours. The original estimate was two to three months.

Then there’s Pieter Levels. One person. More than 70 projects. Even if 95% fail, the handful that work still make $3 million a year in profit. His exact line was: “Every extra employee slows the company down.”

Then there’s OpenClaw: 370,000 lines of TypeScript in a single repo, with a README that literally says “AI/vibe-coded PRs welcome.” Its founder, Peter Steinberger, runs more than a dozen Codex agents at the same time. One person, one day, small-team-week level output.

Andrej Karpathy said in March 2026 on the No Priors podcast that he hasn’t handwritten a single line of code since last December. He spends 16 hours a day talking to agents. His term for it was “AI psychosis.”

The ceiling on individual output was never determined by team size. It’s determined by leverage. AI is the new leverage. Karpathy has said that cloud plus AI makes the Pieter Levels model—one person running multiple companies—actually viable, and that a billion-dollar one-person company is no longer crazy.

Sam Altman has apparently started privately betting with CEOs on what year the first one-person billion-dollar company will show up.

Org Structure Decides Who Wins

Over the last two years, I’ve become convinced of one thing: company-wide AI upskilling will probably fail.

Conway’s Law says organizations design products that mirror their structure. In the AI era, that means traditional organizations will only ever build AI features, not AI-native products.

In a traditional company, the unit of collaboration is human to human.

Even if individual employees get faster, 80% of their time still goes into communication, process, meetings, decisions, and cross-functional coordination. Overall productivity barely moves.

You can’t expect a Formula 1 driver to floor it in an alley.

In an AI-native organization, the unit of collaboration is “human + agent.” One engineer can run five to ten agents at once. What used to take a small team a week can get done in a morning.

Go one layer deeper and this is really anti-collaboration: deliberately reducing human-to-human links and increasing human-to-AI links.

That’s what Cursor did. No product managers. Engineers spin up agents, write code, and decide product direction themselves. 150 people carrying $2 billion ARR isn’t because they’re smarter. It’s because the collaboration model is fundamentally different.

The Claude Code team at Anthropic works the same way. One of their engineers has a catchphrase: “My job now is figuring out how to get as many Claude Code instances working in parallel as possible.”

This is hard inside a traditional company. Not because of the tech. Because of the org chart.

Give an engineer in a traditional department 10 agents and their first reaction is not excitement. It’s this:

→ If my output goes up 10x, are they cutting headcount next year?

→ Can I still protect my team? If not, who gets cut first?

→ If an agent takes over my whole domain, what exactly do I write in my promotion packet?

→ If I do the work, the credit goes to AI. If prod goes down, the blame is mine. Why would I want this?

That’s why the people really using agents to do the work of 10 people are indie hackers, small-team founders, and people whose output maps directly to their own income.

What It Feels Like Firsthand

During the day I build frontend and backend systems inside a company. Nights and weekends I work on open source and side projects with a full AI workflow stacked on top.

The pace on those two sides doesn’t even feel like the same era.

At work, polishing a single API means going through the whole process. From kickoff to production, about a week. At night, I can build an MVP in one evening and put it in front of users the next day.

The tools aren’t different. My company also pays for Claude Code.

The difference is whether the organization can actually let the tool unlock productivity. The answer is no.

So in most companies, “AI transformation” is just a PowerPoint project. Buy tools, run training, send announcements, set OKRs—and keep the same org while every workflow is still blocked by the same process overhead.

An AI-native individual can now do in a few days what used to take a traditional team a few months. Claude, Cursor, and Rakuten have already proven that at the company level. The problem is that most organizations simply do not allow individuals to operate in an AI-native way. The pace gets choked by process almost immediately.

The team is no longer a flywheel. It’s a liability.

Worse, traditional teams are not just slow. They permanently lock you out of becoming AI-native.

The org you built, the contracts you signed, the layers you set up, the middle managers you hired—every single one of them is on the side of not changing. You’re not going to truly reorganize. At most, you’ll wedge in an “AI innovation team” and then watch it die in KPI warfare.

That is the real script behind what 90% of companies call “AI transformation.”

The Breakout Move: Cut First, Then Hire

On Lenny’s Podcast, Boris Cherny mentioned a management principle that’s actually very interesting: “underfund things a little bit.”

His exact line was: “There’s this interesting thing when you underfund everything a little bit, because then people are kind of forced to Claude-ify.”

If teams are deliberately understaffed, under-budgeted, and under time pressure, they’re forced to use AI to close the gap.

The real problem in mature companies is not “too much work, too few people.” It’s “too many people, not enough real work.”

When there are too many people and not enough work, people are not working. They’re coordinating who should work. Meetings, alignment, process, status updates—and the week is gone.

All of that was designed for the productivity limits and friction costs of the previous era. If you just jam AI into it, AI can only help around the edges. It never reaches the main flow.

Cut half the team, and the logic flips.

When there are fewer people and more work, nobody has time for meetings whose only purpose is to justify meetings. Everyone is forced to face problems directly, forced to orchestrate agents, forced to own work end to end. People aren’t getting cut because they’re bad. They’re getting cut because that team size was designed for the last era.

In the new era, that team size itself is the biggest obstacle to lower costs and higher output.

It’s brutal, but it works. In 2026, more and more companies will go this route—not because they want to, but because they don’t have a cheaper option.

The AI Upside Is Not Fair

AI only makes coders code faster, product people validate faster, and people with taste produce better work. It amplifies what you already have. It does not give you what you don’t.

And the people at the top are operating at a level most people can’t even picture. Top indie developers are writing 10,000 to 100,000 lines of code a day.

Karpathy spends 16 hours a day coding through agents and gets anxious if he doesn’t burn through enough tokens by the end of the month.

Boris Cherny ships 10 to 30 PRs a day. Peter Steinberger runs more than a dozen Codex agents at once.

Y Combinator’s Garry Tan said he’s been sleeping only four hours a night because working with Claude Code is so stimulating he doesn’t need anything to keep himself awake. He also said a third of the CEOs he knows are in the same state.

If the top people are already twisted into knots like this, what are ordinary employees supposed to do?

First: drop the employee mindset.

“Wait for instructions, then execute” is already something AI does better than people.

If you can’t break down problems, design solutions, and judge what’s worth doing, then when layoffs come, there’s a good chance your name is on the list. Pure executors are cheap in the AI era.

Second: think full-stack.

Even if you “just write code,” you need to be able to carry work from requirement to deployment on your own—product design, technical approach, launch, operations, all of it. No passing the buck.

In orgs like Cursor or Claude Code that don’t even have product managers, that’s already how engineers work. Being able to run the entire product loop alone is one of the rarest skills in the new era—and one of the hardest to replace.

Third: build business sense.

Real leverage is using AI to orchestrate resources—turning ideas into products, then getting those products sold.

Don’t be satisfied with using AI at work to build a few product features. Learn how your company thinks about product, sales, and execution.

That way, if layoffs do hit, and you can both build and sell, you still have a shot at carving out your own path through entrepreneurship.

Final

Replacing the basic unit of collaboration across the organization—from human-to-human to human-to-agent—is the unavoidable path of enterprise AI transformation.

If a company can’t do that, buying more Claude Code licenses won’t matter.

Organizational inertia, contracts, KPIs, politics, relationship debt—every one of them pushes against the switch, and most companies simply won’t have the nerve to do it.

So more and more of them will choose the blunt option: cut headcount first, fully equip the people who remain with AI, then hire back only when needed. It may not be the optimal path, but it’s general-purpose and it works.

by Ren

Thanks for reading.

If you’re interested in careers, startups, investing, or AI, follow along.