The Bottleneck Was Never the Machine.

AI is going to take these jobs — not assist them, take them. That isn’t the apocalypse everyone braces for; it’s the oldest pattern in technology. Every big upgrade replaces a large class of ordinary workers with a much smaller class of a new kind, and the arrow only ever points one way:

a crowd of ordinary workers → a handful of a new kind.

More output, fewer people, a higher bar — every single time. I’m not calling this from the outside or reading it off someone’s slide deck. I run an AI company, and I’m watching it happen at my own desk this quarter.

The execution seats

Start with what I can see up close. At Thunderbit, a single AI runs our entire influencer outreach, end to end, and no human touches the chain:

1. It scores each creator — worth contacting or not.
2. It writes the first email, answers the reply, and negotiates the deal.
3. It follows up on its own when someone goes quiet.

Support works the same way, and so does most of what used to be a wall of tickets. We didn’t build one big general-purpose AI to do all of it — we build a separate, narrow agent for each kind of situation. What used to fill a room now takes the machine plus exactly one person: it does the work, and the person makes the one call it can’t — is this creator worth it, is this refund fair, yes or no.

The communication got automated. The judgment didn’t.

The polite version of this is that AI “elevates people to higher-value work.” I don’t buy it. The execution seats didn’t get elevated — they got absorbed, and not gone-someday: gone. The last wave did the same to travel agents, down about 60% from their peak, and to newsroom jobs, cut in half in a decade. Desks like that don’t get upgraded; they empty.

Retail already ran this play

Retail is the cleanest version of the pattern, and everyone has lived through it. For a century it ran on people — clerks, cashiers, stockers, buyers, the middlemen between a factory and your closet. There were millions of them, and the job asked for a high-school education, not a degree.

Then e-commerce arrived, which is really just internet plus retail. A company like Amazon served a bigger market than any chain in history, and it didn’t do it with more clerks. It did it with software, and software is written by a far smaller number of college-educated engineers.

Here’s the move everyone skips: the clerks didn’t become the engineers. One class of worker was swapped for a smaller, higher-credentialed one — the cashier wasn’t retrained into a backend developer. Output went up while the headcount went down, and the education bar climbed the whole way. That’s not a side effect of the story. That is the story.

Every upgrade needs fewer people

Retail isn’t special. The same shape turns up everywhere you look, across a century of technology:

• Farming. A hundred years ago about 40% of Americans worked the land; today it’s under 2%, and they grow more food than ever.
• Industry. GM employed more than 600,000 people at its peak; Nvidia became the first four-trillion-dollar company with a fraction of that headcount.
• Schooling. As computers spread, the college wage premium roughly doubled — from 39% to 79% — then held there.

AI is simply the next tier up, and it pushes the frontier up once more. The engineer’s golden decade is closing, and the new elite is the researcher: the handful of people who can move the models themselves, paid like no engineer ever was. Stack the tiers and the whole century compresses into a single line:

retailer → engineer → researcher — each tier fewer people, each tier a higher bar.

Every step trades a crowd for a smaller, sharper few. The pyramid doesn’t just rise. It narrows.

Who makes the cut

So who makes the cut? The ones who kept the judgment and learned to direct the machine. At Thunderbit I’ve tried to make that shift literal and total:

• I bought every employee a microphone and put the whole company on our own voice tools.
• The core skill now is telling an agent, precisely and out loud, exactly what you want — and then having it build the thing.
• A marketer on my team doesn’t file a ticket for a tool anymore; she describes it, and the agent builds it.

The effect is that everyone becomes a little bit technical — not a coder, a builder. Software creation stops being the engineering department’s private job and leaks into every seat in the building. That reshapes who I hire: I’ll take a business person only if they’re hungry to build, and an engineer only if they genuinely care about the business.

The pure specialist doesn’t come along. Even the people who make these tools will tell you the interface is plain language now, not syntax — so the moat isn’t coding. It’s judgment, taste, and knowing what to ask for.

The winning teams are tiny and technical

There’s a last turn of the screw, and as a founder it’s the one I feel most: the companies running this play are shrinking too. The teams building the frontier tools are astonishingly small and almost entirely technical:

• Cursor reached roughly $100M a year with about sixty people, most of them engineers.
• Midjourney did a couple hundred million with a dozen, and took no outside money at all.

hundreds of employees → a dozen — each one building leverage, not running a process.

I feel the same pull in my own hiring. For every seat I’m about to add, I now ask the same question — could an agent do the execution while a person keeps the judgment? — and more and more, the answer is yes. The company that wins my market won’t be the one with the biggest team; it’ll be the one where a few people with real technical judgment aim a lot of machines in the right direction.

That’s why I don’t believe the next great AI companies get built by non-technical founders handing specs to a contractor. You have to be close enough to the machine to know what to ask of it, and what to trust. The founder needn’t be the best engineer in the room, but the founding team has to carry that judgment in its bones. Technical judgment stopped being one department; it’s the whole company now.

The only variable is time

None of this happens overnight. It runs on human time, not compute time — five to ten years, if I’m honest with myself. The machine is ready. We aren’t.

You can actually measure the gap. In one controlled trial, experienced developers working with AI felt about 20% faster — and were clocked 19% slower. The capability is already here; the human adjustment is years behind, and it only closes as fast as people change, which is to say slowly.

But only the speed is ever in question, never the direction. The traditional roles are going, and the new ones will be fewer and higher up. The bottleneck was never whether the machine could do the work — it can, today, in my company; it’s how fast the people move.

So the real question isn’t whether AI takes the jobs. It does. It’s whether you make it into the tier that’s left.