The Wager

In 1930, Keynes predicted his grandchildren would work fifteen hours a week. Economic growth would handle scarcity. The hard part, he thought, would be figuring out what to do with all the free time.

He got the economics right. Global GDP per capita is roughly six times what it was then. But he completely missed the psychology. We didn't work less. We invented new kinds of work, filled our calendars, and tied our identities to job titles so tightly that retirement became an existential crisis.

AI is reopening the question he thought he'd settled.

Dario Amodei wrote in "Machines of Loving Grace" that AI could free humans from drudgery. Same well Keynes drew from: when machines handle the labor, humans will find meaning elsewhere. Sam Altman put it more bluntly. A farmer from fifty years ago would look at what most of us do today and say: that's not real work. He's probably right. Most knowledge work exists because it was too expensive to automate. Not because it needed a human.

Altman predicted the first one-person billion-dollar company. Amodei said it could happen by 2026. Lovable, a Swedish AI startup, hit unicorn status in eight months with 45 employees. Cursor reached $500M in annual revenue with fewer than 50 people. The scale of what one person can operate is moving faster than anyone expected.

Thanks for reading! Subscribe for free to receive new posts and support my work.

I'm testing what that looks like when you actually do it.

My name is Peer. I live in Denmark. By day I head a team at AWS across Europe North, working with the largest companies in each country on AI transformation. I'm in the rooms where a single AI deployment decision takes six months of executive alignment, security reviews, and pilot programs. By evening I run the experiment. I built a real company, Mentilead, specifically as the test bed. Two Shopify B2B apps, real products, real merchants, real problems. Not because I wanted to run a Shopify business, but because the experiment only works if the stakes are real.

The question I'm testing: what happens when you give AI agents every operational role in a real company and measure what actually requires a human?

Not as a thought experiment. On real software, with real merchants and real European data protection law. I call this the Poker Principle. Poker without money is a card game. The stakes change the behavior. An AI agent making hypothetical decisions teaches you nothing. One making real decisions that cost real money teaches you everything.

The enterprise work and the solo work feed each other.

At AWS I sit in rooms where deploying one AI agent takes six months of alignment meetings. Then I go home and hand an entire business function to an agent in one evening. Enterprise measures AI in quarterly business reviews. I measure it in hours reclaimed this week. Enterprise asks "is this safe?" I ask "is this working?"

When enterprise best practice says "do X" and solo reality says "X doesn't apply here," that gap is where the interesting observations come from.

March 12, 2026. Late evening. I'd just closed the laptop on a full day of enterprise AI strategy and opened it again for something completely different. I told an AI agent to build my entire marketing operating system from scratch.

Three hours later it had produced 38 files. A complete ICP for my first app. A fully optimized App Store listing with 20 researched keywords. Two marketing experiments with hypotheses, success criteria, and tracking tables. A daily briefing system. It applied April Dunford's positioning methodology, PAS copy structure, ICE experiment scoring. Frameworks I know about but wouldn't have bothered applying manually at this stage for a zero-revenue side project.

I corrected the agent three times. Wrong customer segment. Activities that require customers I don't have yet. A product maturity assumption that was way off. And honestly, each correction annoyed me. If it's supposed to be an operating system, why am I catching basic context errors in the first hour?

But each correction took about thirty seconds. The agent absorbed the feedback, adjusted, kept going. That loop, "agent produces, human redirects, agent adjusts," ran in minutes instead of the two to three weeks this work would normally take a solo operator. The iteration cost dropped to nearly zero. I wouldn't have done this work manually. The agent wouldn't have known what to cut.

This isn't a twelve-month project. The technology isn't fully there yet. That's the point.

Running a real business on AI agents that aren't yet capable of full autonomy produces the most useful observations. You see exactly where the boundary sits. Over years, you watch it move.

The experiment runs in chapters, starting with foundation and baselines, then handing off functions one at a time, then testing what real autonomy looks like. Maybe in year four or five, something resembling a verdict.

I'll publish monthly field notes here. What the agents did. What they got wrong. What I had to override. Specific numbers. Failures in detail. No conclusions until the data supports them.

Keynes assumed freedom from labor would be welcomed. He didn't anticipate that we'd resist it. That we'd invent what David Graeber called bullshit jobs just to fill the void. That we'd rather be busy than free.

The AI transformation happening right now is asking that same question again. I don't have an answer. But I'm building a business that will push on it in a specific, measurable way over the next several years. From a desk in Denmark, late in the evening, after my actual job is done.

The first data point is in. The agent built the marketing OS. I corrected its course three times. The experiment is running.