Developers Feel 24% Faster with AI. They're Actually 19% Slower.

Here's a number that should make every engineering manager uncomfortable: 32.7%.

That's the acceptance rate for AI-generated pull requests, according to Faros AI's research report aggregating multiple 2026 studies. Human-written PRs? 84.4% acceptance. A 52-percentage-point gap. And the kicker: developers self-report feeling 24% more productive while measuring 19% slower by actual output metrics.

I've been using Claude Code daily for over a year. I believe it's changed how I build software. But these numbers demand honesty about what's happening. The studies found AI-generated code carries 1.7x more bugs and 15-18% more security vulnerabilities than human-written code. Review times increase 91% at the team level because reviewers can't skim AI output the way they skim a colleague's familiar patterns. Developers spend 45% of their time debugging AI output rather than writing new code.

The feeling of speed is real. I get it. When Claude generates 200 lines of working code in 30 seconds, the dopamine hit is enormous. But that's measuring input speed, not output quality. The studies converge on a sweet spot: teams see 10-15% genuine productivity gains when 25-40% of their code is AI-generated. Go above 40%, and review overhead eats the gains.

Right now, 41% of all code across surveyed organizations is AI-generated. We're already past the sweet spot on average. And 21% of AI coding licenses remain completely unused, meaning the teams that are using AI tools are pushing well past that 40% line.

This connects directly to the pricing story below. Vendors are shifting to usage-based billing because agentic workflows burn 5-10x more compute. If that compute is producing code with a 32.7% acceptance rate, the ROI math gets uncomfortable fast.

What I'd tell my team: track your AI PR acceptance rate separately from your human rate. If the gap is wider than 20 points, you're generating code faster than you can verify it. Pull back to the 25-40% range and spend the saved compute budget on longer context windows and better prompts rather than more raw generation. The bottleneck was never writing speed. It's review quality.