85% of Developers Say the Real AI Coding Bottleneck Isn't Writing Code Anymore — It's Reviewing It

AI is making developers write code faster — but it isn't making software ship faster, according to a GitLab-commissioned Harris Poll survey of 1,528 developers and technology buyers. The headline split: 78% said AI increased their code output, but 85% said the real constraint has moved from writing code to reviewing and validating it.

That's a genuinely important distinction for anyone measuring AI adoption by how much code gets generated. More implementation work doesn't automatically mean faster delivery — and ITPro's coverage of the same research makes that gap concrete: 79% of respondents saw higher individual developer productivity, but delivery speed hasn't improved to match. A patch isn't a production change until it's been tested, reviewed, integrated, and released, and that's exactly where GitLab's data says the pressure is now building.

This points to something more useful than a simple "AI good" or "AI bad" verdict: AI appears to be relocating engineering effort, not eliminating it. Faster drafting just shifts the bottleneck downstream — more pull requests, more generated tests, more refactors, all needing human and automated capacity to confirm a change actually works, fits the system, and doesn't introduce security or maintenance risk.

A cited arXiv paper on one enterprise's "2x mandate" adds a striking data point: per-capita throughput reached 2.09x baseline, work per reviewer roughly doubled, and automated review actually surpassed human review volume in that environment. Worth flagging though — this is one company's experience, with no public detail on defect rates, security outcomes, or whether automated review caught everything that mattered. Treat it as a supporting data point, not evidence of a universal pattern.

Bottom line: GitLab's own survey data doesn't claim AI-generated code is worse — just that review and validation is where the strain is concentrating for many teams. For engineering leaders, that means the metrics worth watching now are review wait times, change failure rates, deployment frequency, and time-to-production — not just how much code gets written. Code volume alone is an incomplete signal of how fast a team is actually shipping.