AI coding tools are taking over a substantially larger share of software production, but the result is not necessarily fewer hours of engineering work. Instead, developers are moving “a layer above the code,” spending more time reviewing, debugging and securing machine-generated output, according to a new year-over-year survey from BairesDev.

The nearshore software development company’s Q3 2026 Dev Barometer surveyed 705 developers across more than 60 countries and 41 enterprise CTOs.

It finds that 42% of developers now say AI writes at least half their code, compared with just 12% in the same period last year, Q3 2025.

At the same time, developers report that AI saves them 13 hours of coding per week, nearly double the roughly seven hours reported a year ago.

Those figures would seem to imply a dramatic increase in spare engineering capacity. But BairesDev’s data points in another direction.

Only 21% of developers now spend more than half their workweek writing new code from scratch. Meanwhile, 67% say they are spending more time reviewing AI-generated code than they did a year ago, and 52% report spending more time debugging problems introduced by AI. Developers also report spending an average of nine hours per week learning AI tools and other new technologies, up from four hours a year ago.

“Developers nearly doubled the time AI saves them in coding, to 13 hours a week. Not one of those hours came back,” BairesDev CEO Darren Shimkus said in the release. “A year ago we read the first seven hours as capacity, and we got that wrong.”

For enterprise technology leaders, the findings suggest that the bottleneck in AI-assisted development may be shifting from code generation to code validation.

The developer becomes the accountability layer

BairesDev’s findings fit a wider pattern emerging around AI coding tools: adoption is rising considerably faster than trust.

Stack Overflow’s 2025 Developer Survey, which drew more than 49,000 responses, found that 80% of developers were using AI tools in their workflows, while only 29% trusted their accuracy. Sixty-six percent said they were spending more time fixing AI-generated code that was “almost right,” while 75% said they would still turn to another person when they did not trust an AI-generated answer.

More recent research suggests AI agents themselves are rapidly becoming routine development tools. JetBrains’ 2026 Developer Ecosystem Survey findings on AI coding agents, based on a survey of more than 15,000 professional developers, found that 90% were using AI coding agents at work at least weekly during May through July 2026, with 68% using them daily.

BairesDev’s data suggests enterprises are responding by putting additional resources behind the human checkpoint. Among the 41 CTOs surveyed, 78% said they had increased spending on code review, quality assurance and validation to support AI-generated work. Just 7% of developers said the decision to ship code had been entirely delegated to AI without their input.

That distinction becomes especially important in production systems.

In an interview with VentureBeat, Shimkus said he expects developers’ personal accountability for code to move toward 100%, even as the proportion actually generated by AI continues increasing.

The refrain he hears from sophisticated enterprise customers, he said, is essentially that if an engineer cannot explain what AI-generated code does and why it works that way, the organization will not permit that engineer to check it in.

“In a world where we don’t have AI explainability, I need developer explainability,” Shimkus said.

He drew a sharp distinction between AI-generated prototypes, where companies can tolerate considerably more autonomy, and software entering production environments serving thousands of companies or millions of consumers.

That code, he said, “will not happen without review, at least not anytime soon.”

From writing code to supervising systems

The shift does not appear to be making developers less satisfied with their work. BairesDev found 86% now say AI makes their role more fulfilling, up from 76% in Q3 2025.

Shimkus compares the transformation to what computer-aided design did to architecture. CAD automated much of the manual drafting without eliminating architects; instead, it allowed them to spend more time on other parts of the design process.

“The job hasn’t gotten smaller; it’s moving up a level,” he told VentureBeat.

For software engineers, that higher level increasingly encompasses architecture, problem definition, code review, security, compliance and understanding how generated code interacts with larger systems.

There is evidence this transition was already underway before the latest survey. BairesDev’s inaugural Q3 2025 Dev Barometer found developers saving an average of 7.3 hours per week on coding with AI. The accompanying Q3 2025 analysis reported that 76% said AI made their work more fulfilling and that developers were spending nearly four hours per week upskilling.

By the following quarter, the anticipated change was becoming more explicit. BairesDev’s Q4 2025 Dev Barometer found that 65% of surveyed developers expected their role to be redefined in 2026. Among that group, 74% expected to spend less time coding and more time designing solutions, while 50% foresaw greater emphasis on architecture and strategy.

BairesDev maintains the full Dev Barometer research archive, including previous editions and underlying research materials.

There is an important methodological qualification in interpreting the newest numbers. The 705 developers in the Q3 2026 survey are not simply a sample of BairesDev’s own engineers. According to the company, the majority are applicants participating in BairesDev’s screening process and are not employed by the company. The 41-person CTO sample is also substantially smaller than the developer sample.

Security and judgment become more valuable

The changing workload is also beginning to show up in compensation.

Among surveyed developers who received raises in 2026, 29% identified AI tool fluency as the driver, followed by system design and architecture at 20% and human skills — including communication, mentorship and cross-functional collaboration — at 15%. Yet only one in four CTOs surveyed said they were actively investing in human-skills development.

Shimkus argues developers should therefore resist focusing exclusively on learning the latest coding model. His recommendation is to become highly proficient with one AI tool rather than spreading attention thinly across many similar products, while investing heavily in architecture, review and especially security.

“If I had to advise anybody, should you go even deeper in your code development infrastructure, or to go deeper in security, go deeper in security,” he said.

That emphasis has become particularly timely as agentic systems move beyond code generation toward executing actions themselves. Recent research on security and governance in agentic software engineering reported by ITPro found a significant gap between adoption and the controls engineering organizations have put around agents, including limited use of deployment gates and other governance mechanisms.

The implication for enterprise engineering organizations is that AI adoption does not eliminate the cost of expertise; it changes where that expertise is required. Generating another thousand lines of code is becoming cheap. Determining whether those thousand lines belong in a production system remains comparatively expensive.

As Shimkus put it, companies can purchase AI tools. What they cannot necessarily purchase off the shelf is “the judgment to be able to apply that in a way that accomplishes the business goal but keeps the company safe.”

If BairesDev’s year-over-year trend continues, that judgment — rather than the ability to manually produce code fastest — may increasingly define the value of an enterprise software engineer.