The Job Market Split Nobody's Talking About (It's Already Started). Here's What to Do About It.

AI is making code production virtually free, but the cost of knowing what to build (specification) is becoming paramount. An AI coding agent deleted a production database, highlighting that while AI can follow instructions, its output can be wrong if the specifications are flawed. Studies show AI-generated code leads to more logic issues and increased code review time, even as adoption rises. AWS's Cairo forces developers to write testable specifications before code generation, indicating the critical importance of defining intent. The marginal cost of software production is collapsing, leading to a focus on new bottlenecks: human intent and precise specification. The typical question of whether AI replaces jobs is irrelevant; the real question is where new bottlenecks and valuable human work will emerge.

François Chollet's framework suggests that while AI can perform core tasks (like translation), jobs don't disappear but transform. Translators now supervise AI output, though pay rates and new hiring have been affected. Similarly, software engineers' roles will transform, not vanish. The demand for software will explode as production costs drop, similar to how desktop publishing, mobile cameras, and mobile development expanded their respective markets. Custom software will become accessible to sectors currently reliant on manual processes. While total software employment may grow, individual jobs are not guaranteed, especially as the constraint shifts from production to precise specification.

Most software projects fail due to poor specification, not bad engineering. Historically, high building costs forced careful specification, a filter now disappearing with AI. This accelerates the cost of bad specification, allowing incorrect things to be built at unprecedented speed. The ability to define what code should do, translating vague business needs into precise instructions, is the new center of gravity. Two classes of engineers are emerging: those who drive high-value tokens by precisely specifying systems, orchestrating agents, and evaluating output against intent (demonstrated by companies like Cursor and Midjourney with very high revenue per employee), and those who operate at low leverage with single-agent workflows, doing similar work faster but becoming commoditized. Entry-level positions are declining, and even mid-level and senior engineers face this if they don't adapt.

The 'solopreneur thesis' suggests individuals can unlock tremendous value, which is true for a select 10-20% who possess entrepreneurial instincts, deep domain expertise, risk tolerance, and quick adoption of AI tools. For the remaining 80%, the future will involve smaller teams, higher expectations, and compressed unit economics, increasing pressure to stay employed. The key distinction lies in the economic output generated per unit of human judgment. Agents amplify excellent human specification and judgment, widening the gap between these two classes. This skill is learnable, and organizations that empower their workforce to develop it will gain a significant competitive advantage. Software engineers are the canary in the coal mine; this trend will extend to all knowledge work.

Knowledge work, including analysis, consulting, and project management, is transforming as AI makes organizations leaner. Much coordination work within large organizations (reports, slide decks) will be deleted as teams shrink, as its value stemmed from organizational complexity, not inherent worth. The remaining knowledge work will become more verifiable. Financial services now define portfolio strategies as models with testable assumptions and measurable outputs, essentially specifications. Legal, compliance, and marketing are following suit by structuring outputs into testable claims. This convergence means knowledge work will be subject to the same quality signals as software, making the distinction between engineering and other knowledge work less pronounced. We are all now in the same boat, working with AI agents.

To adapt, knowledge workers should adopt an 'engineering mindset,' not necessarily learn to code. Key steps include learning to specify work with acceptance criteria, understanding both the capabilities and limitations of AI tools, making outputs verifiable (e.g., with data sources and measurable milestones), learning to think in systems rather than documents, and auditing one's role for coordination overhead. Roles focused on organizational complexity will be vulnerable as AI makes organizations leaner. The focus should shift to creating direct value and contributing to revenue-generating products or business direction. This requires precision, testability, and understanding tools well enough to identify errors.

We are currently in the 'trough' of a J-curve of technology adoption, where AI deployment initially reduces productivity before surging. Some manufacturing firms experienced drops of up to 60 percentage points. Experienced developers, despite believing they were faster, were 19% slower with AI tools. However, companies that have mastered AI (like Midjourney and Cursor) demonstrate massive productivity gains and revenue per employee. The adoption cost will compress, with early adopters already moving past the dip. This will lead to a significant gap in productivity ratios between companies that leverage AI effectively and traditional organizations. The historical parallel is not ATMs or calculators, but telephone operators in the 1920s, whose jobs disappeared or shifted to lower-paying roles, despite overall employment growth. Leaders must support their teams to navigate this transition.

The economy will create unprecedented amounts of software and rely more on computers. This is structurally optimistic, as compute creates leverage and abundance. However, individual job security depends on adapting. The bifurcation in the job market is evident, with AI-native companies rapidly gaining market share. The gap between engineers who can drive high-value tokens and others is significant. Leaders must intentionally prepare their teams to develop skills in agent fluency and precise specification. This is a learnable skill, and the window for adaptation is closing as AI capabilities accelerate. Individuals and organizations need to lean in, understand how to work with agents, and embrace the challenge of 'boiling the ocean' by thinking big and producing more value, leveraging AI to define clear intent and constraints rather than merely performing the work.