Writing code was never the bottleneck!

Leaders keep hearing that AI assistants will turbo-charge coding, but the real insight is that most of the promised gains are illusory. Studies from JPMorgan, Atlassian and METR show mixed or negative impacts on velocity, and the hype masks deeper workflow frictions that AI can actually worsen if applied to the wrong problems.

The data is stark: JPMorgan reports up to 20% efficiency gains, Atlassian says 99% of developers notice time savings, yet the METR study found developers were 19% slower when using AI. Developers enjoy the tools, but the productivity lift comes from feeling happier, not from delivering code faster. The disconnect shows that focusing on code generation misses the real bottlenecks-manual approvals, documentation gaps, and search pain.

Experts like Ben Lloyd Pearson of LinearB and Seth Rosenbauer of Joggr argue that AI's highest value lies in automating the "boring" work: test scaffolding, doc creation, and routine templates. Jared Coyle at SAP notes that the most impactful use case is scanning expense receipts, not writing novel algorithms. When AI removes repetitive toil, engineers can spend time on strategic, high-impact tasks.

AI-generated code also introduces hidden costs. The State of Software Delivery 2025 found 67% of developers spend more time debugging AI code, and 68% wrestle with new security vulnerabilities. DORA metrics show a 7.2% drop in delivery stability after a 25% rise in AI adoption. Leaders must measure impact beyond lines of code, tracking stability, throughput, and code quality.

The practical takeaway for technical leaders is to reorient AI investments toward reducing friction: improve internal knowledge bases, automate approval pipelines, and train teams to use AI as a productivity aid, not a replacement for human review. Measure outcomes with reliable metrics, invest in upskilling, and keep a human lens on code quality to avoid the illusion of speed that masks deeper delays.