Coding Was Never the Bottleneck
Is AI truly boosting productivity in software development?
1AI and Software Development
The integration of AI tools into software development processes has created significant excitement in the industry. However, initial research shows that AI usage hasn't delivered the expected productivity gains. METR's early 2025 study revealed that engineers using AI tools worked 19% slower.
2The Perception-Reality Gap
Interestingly, while engineers felt 20% faster, data contradicted this perception. This demonstrated a significant gap between how productive engineers felt versus actual performance. This perception-reality divide became a crucial learning point for the industry.
3Follow-up Study Revelations
METR's August 2025 follow-up study uncovered an even more striking reality: 30-50% of participants refused to submit tasks they didn't want to do without AI. This made controlled studies nearly impossible as the most AI-enthusiastic developers wouldn't participate in control groups.
4Changing Developer Attitudes
Late 2025 data showed some improvement in AI tool effectiveness. However, this was limited by selection bias as the most AI-enthusiastic developers now refused to work without it. METR concluded controlled measurements had become nearly impossible.
5Anthropic's Internal Findings
Anthropic's late 2025 internal study reported 50% productivity gains, with Claude used in 60% of daily work and 67% more pull requests per engineer. However, these optimistic figures contradicted other industry data, suggesting potential measurement biases.
6CircleCI's Large-scale Data
Analysis of 28 million CI workflows showed 59% more workflow throughput but 7% decline in main branch throughput for median teams. Build success rates fell to 70.8%, a five-year low. More code was being written, but less reached production.
7Harvard's Expertise Study
Harvard research found AI helped non-experts brainstorm as well as experts, but in execution they underperformed by 13%. The gap AI appeared to close in planning reemerged during delivery, showing AI doesn't fully address knowledge inequalities.
8METR's 2026 Survey
METR's survey of 349 technical workers reported 1.4x-2x self-assessed productivity gains from AI. Notably, METR's own researchers -most aware of perception bias- reported the lowest gains of any subgroup, questioning subjective assessments' reliability.
9What This Looks Like in Practice
Engineering metrics appear great superficially: more pull requests, more commits, faster task completion. But items marked 'in progress' six weeks ago remain unfinished. AI reduces starting costs but production-ready polish remains challenging.
10The Real Bottlenecks
AI has made first drafts, boilerplate, and test writing easier. But the real bottlenecks - product decisions, design reviews, QA, compliance - remain unchanged. Faster coding just piles more WIP items against the same downstream constraints, as shown in CircleCI's data.
11Key Takeaway
Data proves coding was never software development's true bottleneck. AI might be solving the wrong problem. Real productivity gains require addressing constraints across the entire software lifecycle. This represents a significant paradigm shift for the industry.
Original article: O'Reilly Radar — Kod Yazmak Asla Darboğaz Değildi