9.97%. That's the actual productivity increase developers got from AI coding tools across 40 companies over 16 months, according to a new longitudinal study from DX. Not 2x. Not 10x. Roughly ten percent.
The DX Longitudinal AI Impact Study tracked pull request throughput from November 2024 through February 2026. During that period, AI tool usage jumped 65% across the organizations studied. Most teams landed somewhere in the 8-12% improvement range for PR output. The study specifically excluded teams with individual PR throughput targets to filter out metric gaming.
Writing Code Was Never the Bottleneck
This finding will surprise exactly zero developers who've spent a day wrestling with unclear requirements before writing a single line of code. AI coding assistants like Copilot, Cursor, and Claude Code are genuinely good at generating code faster. The problem is that writing code represents a relatively small slice of what software engineers actually do all day.
Planning. Alignment meetings. Scoping ambiguous problems. Code review. Handoffs between teams. These are the parts of the software development lifecycle that eat the most time, and AI tools barely touch them. You can generate a function in 30 seconds instead of 10 minutes, but if the three-day process of figuring out what function to build hasn't changed, the net gain is modest.
This is the core insight Justin Reock, Deputy CTO of DX, highlights in the research: accelerating one step in a multi-step process only helps as much as that step was actually the constraint.
The Vendor Marketing Problem
AI tool companies have been throwing around 2-3x productivity claims in their marketing for the past two years. These numbers typically come from narrow benchmarks - timed coding exercises, isolated task completions, self-reported developer surveys taken right after trying a shiny new tool.
The DX study measured something different: actual organizational output over more than a year, across dozens of companies. That's a much harder metric to inflate. And at ~10%, it tells a very different story than the pitch decks.
This doesn't mean AI coding tools are useless. A consistent 10% throughput increase across an engineering organization is genuinely valuable. For a 100-person engineering team, that's roughly the equivalent output of 10 additional engineers without the hiring, onboarding, and management overhead. That math works out to real money.
But it does mean that companies who bought AI coding tool licenses expecting to cut headcount by half, or ship twice as fast with the same team, are working from bad assumptions.
Where the Real Gains Might Be
The study points toward an uncomfortable truth for the AI industry: the biggest productivity drains in software development are coordination problems, not coding speed problems. If you want a 2x improvement, you need tools that help teams align faster, scope work more precisely, and reduce the back-and-forth in code review.
Some AI tools are starting to move in this direction - automated PR summaries, AI-assisted code review, meeting summarization for planning sessions. But these are much harder problems than code completion, and the tooling is far less mature.
For now, 10% is the honest number. Teams making purchasing decisions should plan around that reality, not the marketing slide.