Research we cite.
Every figure on this site links to the original study, or to coverage of it where the report itself is not openly available. Here is the research behind our points of view, what it found, and what we think it means for engineering leaders.
What it found
Every 25% increase in AI adoption was associated with an estimated 7.2% drop in delivery stability and a 1.5% drop in throughput.
Based on a large global survey of technology professionals. Individual productivity and satisfaction rose with AI adoption while delivery performance did not.
Our take
AI speeds up one step of delivery. Without changes to the flow around it, the gain is absorbed and stability suffers.
What it found
Experienced developers took 19% longer to complete real tasks with AI tools, while believing they had been about 20% faster.
16 experienced open-source developers completed 246 real tasks on their own mature projects, randomly assigned with or without AI tools.
Our take
The feeling of speed is an unreliable guide. Estimates and staffing decisions need measured evidence.
What it found
45% of AI-generated code samples failed security tests, across more than 100 language models.
Tested across 80 coding tasks. Java had the highest failure rate, above 70%, and 86% of samples failed to defend against cross-site scripting.
Our take
AI writes insecure code fluently. Security checks have to live inside every delivery loop, ranked by real risk.
What it found
Duplicated code blocks increased eightfold during 2024, while moved (refactored) code fell by almost 40%.
Analysis of 211 million changed lines of code. 2024 was the first year copied and pasted lines outnumbered moved lines.
Our take
AI makes adding code easy and reshaping it rare. Technical debt now compounds in months, not years.
What it found
66% of developers say their biggest frustration is AI output that is almost right, but not quite. 46% distrust the accuracy of AI tools.
More than 48,000 developers in 177 countries. Distrust of AI accuracy rose from 31% the previous year, and favorable views fell from 72% to 60%.
Our take
Almost right is the most expensive kind of wrong. Review has to check intent, not just whether code runs.
What it found
95% of enterprise generative AI pilots fail to deliver meaningful financial returns.
Research into enterprise generative AI initiatives found that a small minority produced measurable value, while most stalled at the pilot stage. Linked here through press coverage of the report.
Our take
Tools alone do not produce a return. The operating model around them does. The headline figure has been debated, so we treat it as a signal of the gap rather than a precise measure.
Figures are summarized as reported by each source at the time of writing. See our disclaimer for how we use third-party research.