Coding Was Never a Bottleneck

· Source: AI & ML – Radar · Field: Technology & Digital — Software Development & Engineering, Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, long

Summary

AI productivity research indicates that while AI tools have taken software development by storm, the expected productivity gains are often misperceived or misdirected. A METR study in early 2025 found engineers using AI took 19% longer, despite believing they were 20% faster. Subsequent METR research in late 2025 showed an 18% speed improvement for returning developers and 4% for new ones, but noted severe selection bias. Anthropic engineers reported a 50% productivity boost and 67% increase in pull requests per engineer. However, CircleCI data across 28 million workflows revealed a 59% increase in throughput but a 7% decline in main branch throughput and a five-year low 70.8% build success rate. Harvard Business School research showed AI helps brainstorming but doesn't close the 13% execution gap between novices and experts. The core issue is that AI reduces the cost of starting work, shifting bottlenecks downstream to product decisions, design reviews, QA, compliance, infrastructure, and release processes, rather than eliminating them.

Key takeaway

For engineering leaders evaluating AI's impact on development, relying solely on increased pull requests or self-reported velocity is misleading. Your teams may feel more productive, but AI often shifts bottlenecks to code review, CI/CD, and quality assurance, leading to stagnant feature timelines and lower build success rates. You must redefine productivity metrics to focus on end-to-end cycle time from conception to production, and invest in rigorous quality gates like "reasonable doubt" code reviews and advanced integration tests to ensure actual delivery.

Key insights

AI tools primarily reduce the cost of initiating work, shifting development bottlenecks downstream rather than eliminating them.

Principles

Method

Implement "reasonable doubt" code reviews, adapt CI pipelines with integration and architecture fitness tests, deploy with feature flags and aggressive monitoring, require human-written tests for AI-assisted code, and protect dedicated knowledge-sharing time.

In practice

Topics

Best for: CTO, VP of Engineering/Data, Executive, Software Engineer, Director of AI/ML, MLOps Engineer

Related on AIssential

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by AI & ML – Radar.