AI in Software: Adoption Surges, Impact Lags
AI is universally adopted in software engineering, yet real impact and productivity gains remain elusive due to bottlenecks in the SDLC.
Key Insights
-
Insight
AI adoption is nearing universality, with 88.3% of developers using AI regularly, but this high adoption rate does not yet translate into proportionate productivity impact or improved delivery outcomes.
Impact
This highlights a critical disconnect where organizations might be investing in AI without realizing the expected efficiency gains, leading to potential misallocation of resources and unmet expectations.
-
Insight
AI-generated pull requests (PRs) are significantly larger (over 400 lines of code at P75) and take more than five times longer to be picked up for review compared to unassisted PRs, leading to a much lower merge rate (32.7% vs. 84.5%).
Impact
This indicates that AI-generated code is introducing new bottlenecks in the SDLC, particularly in the review process, and may be creating more work or friction downstream rather than streamlining it.
-
Insight
AI primarily generates new code rather than refactoring existing code, with a near-zero refactor rate for AI-assisted PRs, potentially exacerbating technical debt instead of resolving it.
Impact
This trend could lead to a rapid increase in unoptimized, redundant, or complex codebases, making future maintenance, debugging, and feature development more challenging and costly.
-
Insight
A significant majority (65%) of organizations report lacking dependable data quality, which is identified as the biggest challenge for successful AI adoption across all business functions.
Impact
Poor data quality directly impacts AI model performance, leading to hallucinations, incorrect outputs, and erosion of trust, ultimately hindering AI's ability to deliver reliable and valuable solutions.
-
Insight
AI policies are sharply polarized, with about 60% of companies having clear policies while 26% lack them, indicating a foundational governance gap in many organizations regarding AI usage and best practices.
Impact
Without clear AI policies, organizations risk inconsistent AI tool usage, security vulnerabilities, compliance issues, and a fragmented approach that undermines collective productivity and innovation.
Key Quotes
"AI adoption is effectively maximized at this point. It's almost universal. So it in our findings, 88.3% of developers now use AI regularly, so that's you know, at least multiple times a week, which is up from 70 just under 72% when we last surveyed this back in early 2024."
"AI is accelerating code generation. I I think we all know that at this point, but it's also exposing bottlenecks everywhere else in the SDLC, primarily with things like reviews, testing, governance, organizational readiness."
"If you have all of those things, AI will amplify the benefits of them. And if you have problems with them, AI will amplify the problems that they create. So it will make the worse worse and it will make the best better."
Summary
The AI Paradox: Universal Adoption vs. Elusive Impact in Software Engineering
Artificial Intelligence has permeated nearly every aspect of software engineering, with widespread adoption across development teams globally. However, new data from the 2026 Engineering Benchmark Report reveals a critical disparity: while AI usage is almost universal, its tangible impact on productivity and delivery outcomes remains significantly constrained.
AI-Generated Code: Larger, Slower, Less Accepted
The report, based on an extensive analysis of 8.1 million pull requests (PRs) from 4800 engineering teams across 42 countries, highlights distinct behavioral patterns in AI-generated code. AI-assisted PRs are typically 2.5 times larger than unassisted ones, often exceeding the optimal 300 lines of code threshold. More strikingly, AI-generated PRs wait over five times longer for review and merge at less than half the rate of human-authored code. This stark difference underscores that merely generating more code does not equate to efficient delivery.
Bottlenecks Shifted, Not Solved
While AI accelerates code generation, it is exposing and amplifying bottlenecks in other critical areas of the Software Development Life Cycle (SDLC). The review process is particularly strained, with AI-generated PRs experiencing significantly longer pickup times. Qualitative feedback from engineering leaders indicates a reluctance to review AI-authored code due to concerns about trust, potential errors, missing context, and the mental load of grappling with verbose or off-scope changes. This often leads to a phenomenon of “rubber stamping," where large volumes of AI code slip into production without thorough human scrutiny, impacting quality and security.
The Technical Debt Dilemma: New Code vs. Refactoring
AI predominantly creates new code paths rather than improving existing legacy code. The report indicates an almost negligible refactor rate for AI-assisted PRs, suggesting AI is not actively helping to address technical debt and may, in fact, be contributing to its accumulation. This tendency is attributed to AI's "free" code generation capability and a lack of explicit instruction to prioritize leveraging existing components. The emerging concept of
Action Items
Implement robust metrics to measure the actual impact of AI on software delivery, focusing on acceptance rates of AI-generated code into production, review patterns, and overall delivery outcomes rather than just raw usage statistics.
Impact: This will provide actionable insights into AI's real value, identify specific bottlenecks, and enable data-driven adjustments to AI strategies for optimizing productivity and ROI.
Prioritize fixing foundational issues such as ensuring high-quality data, establishing clear and communicated AI policies, and improving internal tooling and platform reliability before broadly scaling AI adoption.
Impact: Addressing these foundational elements will create an environment where AI can amplify positive outcomes, prevent the exacerbation of existing problems, and build a reliable framework for future AI success.
Focus on "context engineering" by providing AI models with comprehensive organizational context, including established policies, coding standards, and existing library usage, to guide AI towards generating more relevant and maintainable code.
Impact: This approach will reduce the generation of off-scope or redundant code, minimize technical debt accumulation, and increase the likelihood of AI-generated contributions being integrated efficiently and securely.
Mentioned Companies
Linear B
5.0Linear B is the co-producer of the 2026 Engineering Benchmark Report, provides the data and framework (Apex), and is presented as a solution provider.
DORA
4.0DORA research is referenced positively as a source of inspiration for the AI readiness matrix and success criteria, indicating its value and influence.