Estimated reading time: 3 minutes
Key takeaways:
- AI is generating code faster than teams can trust it. 35% won’t ship AI-generated code because they can’t assess the risk.
- The bottleneck has moved from generation to verification: AI writes at machine speed, review runs at human speed.
- Speed without understanding creates an ownership crisis. Developers are merging code they can’t explain, and accountability becomes dangerously ambiguous when production breaks.
AI-generated code is rapidly becoming standard practice, yet more than a third of teams are still hesitant to move it into production, wary of risks they lack visibility into.
A new report from the engineering platform Flux paints a paradoxical picture of AI-assisted coding. Flux carried out a study of 309 engineering leaders and practitioners across five continents. It found that 35% of teams are using AI to write code but won’t ship it because they lack the confidence to do so safely.
The use of AI-coding assistants is increasingly common across engineering teams. LeadDev’s 2026 State of AI-Driven Software Releases report found that most organizations are already using AI-generated code to some extent, although for many teams it still represents less than 25% of new code.
“Teams celebrate the productivity gains while flying blind on what’s changing in their code, but you can’t manage what you can’t see,” says Ted Julian, CEO and founder of Flux.
Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
The AI coding visibility gap
The root cause of the issue is a visibility gap. Teams often lack the insight needed to fully understand what AI has generated, making it difficult to assess risk with confidence.
As a result, a significant number of teams are holding back code they aren’t comfortable shipping, while just 3.6% report that AI-generated issues never make it into production, Flux found.
The visibility gap becomes even more pronounced as AI accelerates code production. Many organizations don’t yet have the tools or processes required to review, validate, and govern AI-generated code at the speed and scale it’s being created.
LeadDev reports that AI-generated pull requests (PRs) contain roughly 1.7× more issues than human-generated ones, making code review a growing bottleneck.
“Today, the bottleneck has shifted entirely to everything that happens after generation: code review, quality control, security verification, and cost governance,” says Kris Kang, chief product officer, agent systems at JetBrains. “We are effectively running code generation at machine speed, while the downstream verification and organizational processes are still dragging along at human speed.”
When teams generate volume, developers frequently end up merging code they didn’t personally write, he adds.
“Over time, the deep, structural understanding of how the system functions fades, leaving teams hesitant to ship because nobody can explain what the code actually does or why,” Kang says.
Jeff Watkins, chief AI officer at consultancy NorthStar Intelligence, agrees. “The biggest bottleneck today is no longer typing code into an editor. It is verification. More specifically, it is the combination of visibility, comprehension, and confidence. Can we see what changed? Do we understand why it changed? Can we prove it behaves correctly? Can we support it in production? Can we explain it to auditors, users, or incident responders if something goes wrong?” he says.
More like this
Accountability in AI coding
AI has accelerated code creation, but not code understanding, creating a growing ownership gap.
“Without the right infrastructure, AI makes ownership vastly more ambiguous,” Kang says. “It creates a highly uncomfortable tension where developers are pressured to move faster, yet they remain entirely on the hook if production breaks down the line. It becomes dangerously easy to accept a PR filled with code that no single human fully comprehends.”
This uncertainty is compounded by a “triple debt,” Kang says: rising AI costs, workforce disruption, and growing system complexity that becomes difficult for human engineers to maintain.
Berlin • November 9 & 10, 2026
AI. Burnout. Big decisions.
The pressure is real. Find what works at LeadDev Berlin.
“Relying blindly on any single ‘frontier’ model provider as a long-term advantage is highly unwise and invites vendor lock-in. To clear up this ambiguity, we look at the industry through a skyscraper framework: while powerful models sit on the ground floor and basic chat interfaces sit in the penthouse, there is a total lack of infrastructure in the middle,” Kang says.
Engineering leaders should remain vendor-neutral and focus on value over hype, using the right infrastructure to scale AI safely while preserving developer effectiveness and business impact.
