Businesses are rapidly adopting AI coding tools, and the results are being seen in the speed at which software is shipped.
This is the central finding of a comprehensive new benchmark study by engineering intelligence platform Jellyfish, which analyzed data from more than 700 companies, 200,000 engineers, and 20 million pull requests.
Adoption is already mainstream. Across companies in our dataset, the median AI tool adoption rate is 63%, with 64% of companies generating the majority of their code with AI assistance. According to data from Jellyfish, weekly usage has steadily increased over the past year, with an increasing proportion of engineers using AI coding tools multiple days per week.
The biggest impact is on production. Companies with the highest levels of AI adoption (defined as 75% to 100% of engineers using AI coding tools at least 3 days a week) merged an average of 2.2 pull requests per engineer per week. That’s nearly double the 1.12 pull requests per week in low-adoption companies (a pull request (PR) is a proposal to add or change code in a shared software project so that others can review and approve it before publishing).
“A lot of people are accepting that AI is going to make coding faster, and they’re starting to think about all the problems that come after that,” Nicolas Arcolano, head of research at Jellyfish, told me in an interview. “Are we actually getting a good ROI from these benefits? How does that impact quality?”
“Claude Christmas”
He said he stopped writing the code himself in the fall of 2025 and turned the job over to AI tools. Around that time, significant improvements in models led to significant improvements in AI coding tools and their use began in earnest. Arcolano calls the moment “Claude Christmas” when many software engineers discovered the true power of Anthropic’s Claude Code service and started experimenting with it over the holidays.
“Last fall was around the time I stopped writing my own code altogether,” Arcolano told me. “So, since, say, October, I haven’t written any code myself or even actually looked at any code.”
Other common AI coding tools used by engineers include OpenAI’s Codex, Cursor, and GitHub Copilot.
code quality
Importantly, according to Jellyfish data, code quality does not appear to suffer under the weight of output acceleration. As AI adoption increases among software engineers, the revert rate (code that must be rolled back after deployment) increases only slightly, from 0.61% at low adoption companies to 0.65% at the highest adoption rates.
“We didn’t see much of an impact on quality,” Arcolano said. “That’s where we need to focus as we move forward. Quality concerns and trying to hold the line under increasing pressure to do things faster and the ability to validate AI code are bottlenecks.”
autonomous coding
Meanwhile, more advanced uses of AI are also emerging. Autonomous agent activity (pull requests opened or committed by AI agents) is still a small portion of overall work, but is growing rapidly, especially among top adopters.
What struck Arcolano most was the difference between organizations that have fully adopted AI agents for coding and those that have done little or nothing in this new area.
“The separation is accelerating,” he told me. “So the people at the bottom are not moving, the people in the middle are gradually moving, and the people at the top are on a rocket ship and running away with it.”
“This is the story of AI adoption,” he added. “This is why there’s so much excitement around autonomous agents.”
Sign up for BI’s Tech Memo newsletter here. Please contact us by email. abarr@businessinsider.com.
