Why AI Has Not Replaced Software Engineers

AI is Not Driving Mass Software Engineering Layoffs

Contrary to corporate narratives, there is little empirical evidence that AI is causing mass layoffs in software engineering. While CEOs often cite AI as a reason for headcount reductions to satisfy stakeholders, data suggests these cuts are typically driven by financial pressure, activist investors, or corporate restructuring.

The Phenomenon of "AI Washing"

Many companies use AI as a convenient explanation for layoffs to avoid admitting to financial instability or poor management. Evidence of this "AI washing" includes:

  • Executive Disconnect: A Harvard Business Review survey of over 1,000 global executives found that while 60% had made headcount reductions in anticipation of AI, only 2% had made large reductions based on actual AI implementation.
  • Lack of Ready Tooling: Forrester analyst J.P. Gownder reports that 90% of companies claiming to prepare for AI-driven layoffs do not have a mature, vetted AI application ready to fill those roles.
  • Regulatory Data: In New York State, where the WARN Act added an AI disclosure checkbox in March 2025, nearly zero companies checked the AI box for mass layoffs in the first year, despite thousands of workers being affected.

Employment Trends vs. Layoffs

Layoffs are a poor signal for AI's productivity impact because firing experienced workers destroys the tacit knowledge required to operate AI effectively. Instead, AI's impact manifests as slower hiring. Federal Reserve economists found that employment in the U.S. is still growing, though it is growing roughly 3 percentage points slower per year post-ChatGPT than it would have in a no-AI counterfactual.

The "Decide-Execute-Deliver" Sandwich Model

Software engineering is not merely the act of writing code; it is a three-layer process consisting of decision-making, execution, and delivery. AI has significantly compressed the middle layer but has left the outer layers largely untouched.

1. The Decide Layer (Problem Framing and Specification)

This layer involves deciding what to build based on user needs, market signals, and regulatory constraints. This is a primary bottleneck because requirements specification is complex and time-consuming. As AI automates lower-level decisions, the value of human decision-making simply migrates upward to more complex problems.

2. The Execute Layer (Design and Implementation)

This is where AI has had the most impact. AI agents can generate vast amounts of code rapidly. However, research on 100,000 GitHub developers shows that while AI agents led to an eight-fold increase in lines of code written, this only resulted in a 30% increase in actual releases. This confirms that writing code was never the primary bottleneck in software production.

3. The Deliver Layer (Verification and Accountability)

Human teams must remain accountable for what is delivered. Because AI remains unreliable for mission-critical systems, the need for human verification, testing, and integration remains constant. Liability laws and professional norms ensure that humans remain the "fall guy" or the accountable party for production software.

Agentic Engineering vs. Vibe Coding

There is a critical distinction between using AI as a tool and relying on it blindly, a difference that impacts job security and software quality.

  • Vibe Coding: The user tells an agent what to do without supervising the process, reviewing the code, or possessing the skills to evaluate the output. This approach is high-risk and often results in vulnerabilities.
  • Agentic Engineering: The engineer uses AI agents as a force multiplier while remaining in control and accountable for the output. This process is often mentally exhausting due to the high level of supervision required.

Data from the SWE-chat dataset indicates that only 44% of agent-produced code survives into final user commits, and "vibe-coded" commits introduce vulnerabilities at nine times the rate of human-only code.

Future Demand: Jevons' Paradox and Software Elasticity

Rather than replacing engineers, AI may increase the total demand for software engineering through a phenomenon similar to Jevons' Paradox: as a resource (code) becomes cheaper to produce, the demand for that resource increases.

  • Price Elasticity: Software is highly price-elastic. When the cost of creation drops, more software is created—including one-off utilities and complex enterprise systems that were previously too expensive to build.
  • Historical Precedent: Programmer employment in the U.S. grew from near-zero in 1950 to millions today, despite decades of continuous automation (e.g., the shift from assembly to high-level languages).

Community Perspectives and Counterpoints

While the "sandwich model" suggests overall stability, community discussion highlights several nuances:

"If you have a narrow skill set and are focused in particular areas (e.g. front-end web development), then I would worry more, because even if AI cannot replace software engineers in general, it’s quite likely to be able to completely consume specific domains."

Other practitioners argue that the ratio of engineers to projects will change, potentially leading to "one-person teams" for greenfield projects, which would reduce the total number of jobs even if the overall amount of software produced increases.

Additionally, some argue that the "Delivery" layer is actually the domain of DevOps and SREs, and that once deployment scaffolding is fully automated by agents, the barrier for non-technical users to ship production software will vanish, further displacing traditional engineers.

Sources