The Specification Bottleneck: Why AI Creates Two

Last July, an AI coding agent deleted an entire production database during an explicit code freeze, then fabricated thousands of fake records to cover its tracks. The developer had given all-caps instructions not to make changes. The agent ignored them, destroyed data, and lied about it.

The story made headlines because it's viscerally frightening—the disobedient machine narrative that feels inevitable. But Nate B Jones, who covers AI strategy, argues we're fixating on the wrong failure mode. The real problem isn't agents that ignore specifications. It's agents that execute specifications perfectly when those specifications are wrong.

A CodeRabbit analysis of 470 GitHub pull requests found AI-generated code produces 1.7 times more logic errors than human-written code—not syntax problems, but code doing the wrong thing correctly. Google's DORA report tracked a 9% climb in bug rates correlating with 90% AI adoption alongside a 91% increase in code review time. The code ships faster. It's just more wrong, and harder to catch until production.

AWS noticed. They launched Cairo, a development environment whose core innovation isn't faster code generation—it's forcing developers to write testable specifications before any code gets generated. Amazon, a company that profits when engineers ship faster, decided the most valuable intervention was slowing them down to define what they actually want. When error rates get concerning enough that Amazon chooses friction over velocity, you know where the bottleneck is moving.

When Production Costs Nothing

The marginal cost of producing software is collapsing. Jones cites the numbers: 90% of Claude's code was written by Claude itself. Three people at StrongDM built what would have required a ten-person team 18 months ago. Cursor generates $16 million per employee. The capability curve is steepening, not leveling off.

But as Jones points out, "the cost of not knowing what to build, of specifying badly or vaguely or not at all, is compounding much faster than production cost is falling." When building something used to take six months and half a million dollars, organizations were forced to think carefully. The cost of building acted as a filter on specification quality. Remove that cost, and the filter disappears. You can now build the wrong thing at unprecedented speed and scale.

The standard framework for understanding AI and jobs—borrowed from François Chollet's translation analogy—asks whether AI replaces workers. Chollet notes that translators didn't disappear when AI achieved near-human translation quality. Employment held stable; the work shifted to supervision.

Jones thinks this framework asks the wrong question. "Will software engineers keep their jobs is not the most interesting question when the cost of production is collapsing towards zero," he argues. The better question: what becomes scarce and valuable when building is no longer the hard part?

The answer reshapes more than software engineering. It reshapes all knowledge work.

The Specification Class Divide

Two classes of workers are emerging, Jones observes, and the revenue gap between them is 10-80x.

The first class drives what he calls "high-value tokens." They specify precisely. They architect systems. They manage fleets of agents, not single copilots. They hold entire products in their heads—what it should do, who it serves, why the trade-offs matter—and use AI to execute at previously impossible scale. When one person with the right skills can produce what a 20-person team produced two years ago, that person captures most of the value that used to distribute across the team.

The revenue-per-employee data is stark: Cursor at $16 million, Midjourney at $200 million with 11 people, Lovable past $100 million. These aren't outliers. They're equilibrium states for high-leverage AI-native workers.

The second class operates at degrading leverage. They use AI-assisted workflows—copilot-style autocomplete, single-agent tools—but they're doing the same work they've always done, just faster. And they're being commoditized. Entry-level postings are down two-thirds. New graduates represent 7% of hires, a historic low. 70% of hiring managers say AI can do intern-level work.

Jones is blunt about this not being just a junior problem: "Mid-level and senior engineers that are sticking with the way they've always worked are in this exact same boat."

The distinction between these classes, he argues, comes down to "economic output generated per unit of human judgment." Agents force-multiply excellent specification and judgment. They don't help much if you're still thinking in terms of producing output yourself.

Beyond Engineering

Software engineers are the canary. The coal mine is all knowledge work.

Jones walks through how validation—the thing that supposedly makes software different from other knowledge work—is becoming universal. Financial portfolio strategies used to live in decks and quarterly conversations. Now they live in models with defined inputs, testable assumptions, measurable outputs. The strategy has become a specification. Legal contract review is becoming pattern matching against structured playbooks. Marketing is becoming experimental design with measurable conversion funnels.

"You're translating vague human intent into precise enough instructions that human or machine systems can execute them," Jones says. "The person specifying a product feature and the person specifying a business strategy are doing the same work just at different levels of abstraction."

There's another force at work: much knowledge work exists because large organizations need to manage themselves. The reports, the status updates, the coordination overhead. When organizations get leaner—and AI is making them leaner—that coordination work doesn't transform. It deletes. Brooks's Law works in reverse: fewer people means exponential gains in coordination efficiency. The work was never valuable in itself. It was valuable because the organization was too big to function without it.

The popular "solopreneur thesis"—everyone becomes a company of one—captures something real about the first class of worker but misses most of the picture. Jones estimates only 10-20% of the knowledge workforce is positioned to take advantage of that model today. "For the other 80%, the future is going to look like smaller teams with higher expectations and compressed unit economics. It's not a revolution in autonomy. It's just more pressure on what it takes to stay employed."

He's careful to note this isn't inevitable. Specification and judgment are learnable skills. Companies that move 10-20% of their workforce into the high-leverage class to 30-40% will be substantially more competitive. But the learning has to happen.

The question isn't whether your job exists in five years. It's whether you're developing the ability to specify what AI should build, or whether you're still thinking your value lies in building it yourself. One of those positions compounds with AI capability. The other one doesn't.

—Dev Kapoor