A longitudinal study from developer experience firm DX (GetDX) is challenging the productivity narratives that have dominated AI vendor marketing for the past several years. Tracking a random sample of 400 companies between November 2024 and February 2026, the study found that a 65% average increase in AI tool usage translated to only a 9.97% improvement in pull request throughput — used as a proxy for developer output. The figure was specifically hardened against gaming effects by excluding teams that had set individual PR throughput targets, making it one of the more methodologically careful data points in an otherwise anecdote-heavy conversation. Authors Justin Reock and Abi Noda, writing in the Engineering Enablement newsletter, note that most engineering leaders they speak with report landing in the 8–12% range, consistent with the study's central finding.
The explanation DX heard most consistently from developers points to a fundamental misdiagnosis in how AI tools have been marketed. Writing code was never the primary bottleneck in software delivery. Planning, alignment, scoping, code review, and handoffs — the coordination-heavy, human-centric portions of the software development lifecycle — <a href="/news/2026-03-14-ai-did-not-simplify-software-engineering">remain largely unaddressed by current AI tooling</a>. As one developer told the researchers: "A four-day task might take three. But that doesn't mean I'm shipping 3x more PRs." André Meyer, a researcher at FlowLabs studying AI adoption in Swiss engineering organizations, reached the same conclusion: self-reported time savings of roughly four hours per week align with the ~10% throughput figure. Meyer also flagged a redistribution effect: faster code generation <a href="/news/2026-03-14-amazon-employees-say-ai-is-just-increasing-workload-study-confirms">shifts effort downstream toward review, validation, and managing accumulating technical debt</a> rather than eliminating work altogether.
The results directly contradict high-profile vendor claims, including GitHub's reported 55.8% faster task completion for Copilot users and Nvidia's assertion of 3x code production with Cursor. A separate study from METR, an independent AI evaluation organization, has added further skepticism around self-reported productivity gains — the organization even had to revise its study design after developers began refusing to complete randomly assigned no-AI control tasks, introducing bias that likely understates true AI impact. BCG has separately found that only 26% of companies generate tangible AI value at scale, a figure that rhymes with the DX data.
Hacker News commenters pushed back on the disappointment framing. One observed that a genuine, sustained 10% productivity improvement across an entire industry would itself be historically significant — a "once-in-a-lifetime" economic event by conventional measures. The letdown narrative, that argument goes, reflects vendor-inflated expectations rather than a sober economic baseline. DX says the full longitudinal study is ongoing, with future work focused on why some teams capture more upside than others — and whether the answer lies in the parts of the workflow AI hasn't touched yet: review cycles, handoffs, and coordination overhead.