This number should worry CIOs who signed off on deploying AI coding in hopes of seeing a productivity windfall.
According to Stack Overflow, approximately 84% of developers currently use or plan to use AI tools. Recently published research Over 49,000 developers are participating. However, a 2026 study from developer intelligence platform DX Research found that even though 93% of the 121,000 developers surveyed had reached AI, productivity gains plateaued at about 10%.
Productivity numbers are only part of the story. These are signs of deeper changes. AI is not just changing the speed at which software is built. It’s changing the way developers work, the way teams are structured, and most importantly the way the next generation of engineers learn technology.
Kai Chuang, CIO at Circles, said he has seen this firsthand as developer work moves from hands-on coding to design and system architecture. He said developers at workplace hospitality service providers now spend less time “literally programming” and more time specifying what to build and testing whether it works. The speed of change caught him off guard. Once developers started trusting the output, “the switch to near-full AI code generation happened quickly and automatically” without top-down directives, he said.
The missing skill is no longer writing code. “It’s about knowing what to build, how to design it, whether it’s safe, and whether it’s actually going to improve business outcomes,” explained Eric Brown, senior partner at management and technology consulting firm West Monroe.
Rather than simply writing more code, Brown said, “companies that get this right will redesign their software development lifecycle around AI. Companies that just hand tools to developers will get more activity, but not necessarily better results.”
Developers become designers and reviewers
At UiPath, an enterprise automation software company, a new division of labor is already the norm: “The majority of the code that goes into production is already written by coding agents,” says Raghu Malpani, chief technology and product officer.
“Developers are transforming from coders to reviewers to system architects. Rather than writing every line of code, developers are defining intent, validating output, and shipping more code faster,” Malpani said. He calls this “one of the core parts of a developer’s identity changing.”
Once coding is no longer a slow step, the bottleneck moves upstream to the design, and the AI re-architects as well. That puts new demands on business analysts and product managers for “out-of-the-box” concepts, Circles’ Chuang said. By using AI to explore use cases and mock up interfaces before developers get involved, he said, “we can deliver better, more polished designs.”
Raghu Malpani, Chief Technology and Product Officer, UiPath
Beyond productivity metrics
If productivity appears to be leveling off, CIOs should first assess whether they’re measuring the wrong thing. Consider what Cornerstone Research, an economic and financial consulting firm that supports high-stakes litigation, discovered in its data. After reviewing more than 1 million billable time records, “so far the answer is essentially the same,” said Phil Leslie, chief technology and innovation officer. But while that conclusion is accurate, it is also misleading.
“The use of AI has not measurably reduced the percentage of analyst work hours,” Leslie said. “But what it has done is a change in composition. Analysts report that they now spend less time coding and debugging and more time interpreting, methodology, and thinking. Even though the time hasn’t changed, the work feels different.”
Some organizations report significant productivity gains with AI-assisted coding. But even there, technology executives argue that increased productivity is not the most important change.
For example, Bank of America, which invests approximately $14 billion annually in technology, says its more than 18,000 developers are using AI-powered coding assistance to improve efficiency by more than 20%. But raw speed isn’t the key, says Hari Gopalkrishnan, the bank’s chief technology and information officer.
“The need for talented people who can solve complex problems, use judgment and build relationships will continue to be important,” Gopalkrishnan said.
Hari Gopalkrishnan, Chief Technology and Information Officer, Bank of America
Comparing activity metrics and business outcomes
Most standard AI coding metrics still count effort rather than results. That is, number of sheets expanded, tokens consumed, lines of code generated, self-reported time saved, etc. “These are indicators of activity,” said Brown, of West Monroe. “The better question is whether there has been a change in business and engineering outcomes.”
Brown recommended a dashboard that doesn’t read like a token counter but tracks idea-to-production cycle time, deployment frequency, change failure rate, defects avoided, security vulnerabilities, and the percentage of AI-generated code that requires critical human fixes. “The goal is not to add more code,” he said. “Faster, safer, and higher quality delivery leads to business outcomes.”
The data shows why quality matters. Developer tools maker GitClear analyzed 211 million lines of code and found that code churn nearly doubled between 2020 and 2024, while refactoring fell from 25% to less than 10%. Software distribution company Opsera’s 2026 benchmarks found that AI-generated pull requests take 4.6 times longer to review and contain 15-18% more security vulnerabilities than human-written code. Time saved by writing code often shows up again later in review queues and security fixes.
Junior developer time bomb
However, the most serious risks will not appear in this year’s indicators. It’s lurking for two or three years. The routine tasks that AI now absorbs (bug fixing, documentation, test coverage) were exactly how young developers honed their skills. “Unless there is a new apprenticeship model to replace it, stripping it away will create a talent shortage for companies in two or three years,” Brown warned.
Cutting entry-level roles on the theory that AI will replace juniors is “almost like a one-way door,” Cornerstone Research’s Leslie warns. “The apprenticeship model is a way for nearly all professions to develop the judgment that seniors ultimately rely on.”
The solution is not to stop hiring young developers, but to redefine their roles. A good early developer “doesn’t just know how to write code; they know how to ask the right questions, understand the business intent behind the software, and evaluate whether the output produced by AI actually solves the problem,” Brown said.
Recruitment calculations will also change. Chuang said he currently prefers developers who are “more interdisciplinary and interested in solving underlying business problems.” UiPath’s Malpani added that as the cost of coding becomes cheaper, “the judgment of what and how to code becomes a valuable asset,” giving that premium to developers who understand system design and can keep automation “secure, compliant, and maintainable over time.”
Governance moves to the center
As AI-generated code proliferates, surveillance is moving from corner to corner at the heart of work. “The focus has shifted from reviewing every line of handwritten code to managing the entire software lifecycle, including testing, deployment, permissions, auditability, and runtime behavior,” Malpani said.
“Enterprises will need a platform that provides consistent monitoring, control, and traceability regardless of which coding agent writes the code,” he said, stressing that agents “need guardrails and experienced reviewers.”
That’s a counterintuitive lesson. Coding agents “do not eliminate the need for low-code or enterprise development platforms,” Malpani said. “They made [those platforms] Something more valuable. Faster code generation increases the need for review, judgment, governance, and collaboration. ”
Speed is part of the reward. But the real value of AI coding tools won’t come until the way we approach software development changes: how we build teams, how we measure work, and how the next generation learns to judge what machines produce.
