GitHub published a peer-reviewed controlled experiment in 2023 that has become one of the most cited pieces of research in the AI-assisted development space. The methodology was straightforward: 95 professional developers were randomly assigned to work on a defined task either with or without GitHub Copilot. The Copilot group completed the task in an average of 1 hour 11 minutes. The control group took 2 hours 41 minutes. That is a 55.8% speed improvement, statistically significant with a p-value of 0.0017. The task completion rate was 78% with Copilot versus 70% without. Separately, 73% of developers reported staying in a flow state, and 87% reported reduced mental effort on repetitive tasks.
A follow-up study in 2024 looked at code quality rather than speed: code written with Copilot showed improvements in readability (3.62%), reliability (2.94%), maintainability (2.47%), and conciseness (4.16%). These are small but consistent signals in the right direction.
Then the 2024 DORA State of DevOps report landed, surveying 39,000 professionals, and presented a more complicated picture. Teams with higher AI adoption showed a 1.5% decrease in delivery throughput and a 7.2% reduction in delivery stability. This was the second consecutive year DORA found this correlation. The report's hypothesis for the mechanism: AI tools tend to increase batch size (developers produce larger changesets) and DORA's decade of data consistently shows that larger batch sizes introduce more risk and reduce stability.
GitClear's independent research added a further data point: a 2025 report found a 4x growth in code clones (copy-pasted duplicates) in repositories with high AI assistant usage, alongside increased "churn code", code written and then reverted within two weeks.
How do you reconcile a 55.8% speed gain in a controlled experiment with degraded delivery stability at the portfolio level? The answer is that individual productivity and system-level outcomes are different things, and conflating them is a category error. A developer producing code 55% faster is producing more code. If that code is in larger, less reviewable changesets, if it introduces subtle clones that become maintenance liabilities, and if reviewers are approving it faster because "the AI wrote it so it must be fine," the downstream effects on stability are predictable.
The pattern that resolves this tension in practice is to use AI tooling at the individual level (autocomplete, suggestion, boilerplate generation) while tightening the system-level controls that govern how code moves from individual to production. That means enforcing small PR size limits (under 400 lines of net change as a starting point), requiring that AI-assisted PRs explicitly note which sections were AI-generated so reviewers can calibrate their attention, and treating the static analysis and architecture conformance checks in your pipeline as non-negotiable rather than advisory.
On Azure DevOps specifically, the documented pattern for integrating Azure OpenAI into CI pipelines, described in detail by John Lokerse in January 2026, adds an AI-powered review task that runs against the PR diff before human review. The task calls an Azure OpenAI endpoint with context about the changed files, the associated work items, and a system prompt encoding your architectural principles. This narrows the surface area a human reviewer needs to cover and ensures architectural concerns surface before approval rather than in a post-release incident.
Sources: GitHub Blog, "Research: Quantifying GitHub Copilot's Impact on Developer Productivity and Happiness" (2023); GitHub Blog, "Does GitHub Copilot Improve Code Quality?" (2024); GitClear, "AI Copilot Code Quality: 2025 Data" (gitclear.com); Google Cloud / DORA, "Accelerate State of DevOps 2024" (dora.dev); John Lokerse, "Automated Code Reviews in Azure DevOps using Azure OpenAI" (johnlokerse.dev, January 2026).