This is your last article that you can read this month before you need to register a free LeadDev.com account.
Estimated reading time: 3 minutes
Key takeaways:
- Report finds 81% of leaders had an AI-linked production incident in six months.
- Faster code isn’t faster releases: engineers spend more time debugging than writing.
- Measure defect rates andtime to resolution , notlines of code .
AI-coding tools are meant to help engineering teams ship software faster. New research suggests that the speed they can generate code means engineers are struggling to understand and debug the end product before it reaches production.
A new report from debugging company Undo found 81% of engineering leaders say their organization has experienced a production incident or service outage linked to its use of AI-coding tools in the last six months.
More than nine in 10 had encountered serious defects, poorly optimized code, or test escapes that reached production. A further 93% said an AI hallucination had led to the root cause of an issue being incorrectly diagnosed.
Your inbox, upgraded. #
Receive weekly engineering insights to level up your leadership approach.
The research asked 300 senior engineering leaders at large US and UK companies developing mission-critical software on complex codebases to describe how they worked and what they encountered, with the questions posted during July and August 2026.
The results point towards a growing challenge for engineering leaders adopting AI. Their teams are generating more code, but that’s not necessarily feeding through to delivering working software more quickly.
Has AI progress stalled? #
Some 79% of engineering leaders surveyed say their release cycles are no faster than before the advent of AI. One reason for that may be that the time spent on different jobs has differed.
The engineers the managers oversee now spend an average of 16.9 hours a week debugging, compared with just 9.8 hours producing code. That’s largely because 35% of AI-generated code is reaching production before engineering teams have fully comprehended it.
The data points suggest that AI has shifted the bottleneck rather than removed it, says Mark Williamson, CTO at Undo. “We’ve kind of unbottlenecked the bugging phase,” he says. AI can produce code extremely quickly, but it remains vulnerable to misunderstanding requirements and edge cases. “There’s definitely going to be bugs,” he explains.
The preponderance of bugs becomes particularly problematic inside long-lived systems with large codebases. Existing enterprise software often contains years of accumulated assumptions and undocumented interdependencies. How all that works together is often locked inside the heads of experienced engineers.
Undo’s report argues that the problem is particularly acute in mission critical systems like trading platforms and networking infrastructure because even the smallest mistakes can have consequences far beyond a failed test.
More like this #
Changing the job of a developer #
Williamson also thinks more AI use changes the relationship developers have with their own code.
The scale and speed at which it works means if an agent generates a large chunk of an application in seconds, the engineer responsible for maintaining it can likely never end up with the detailed knowledge that those who wrote it in previous generations would.
This is an issue that is vexing the industry as it tries to build up the knowledge junior developers need but can’t get, thanks to the convenience of ChatGPT and other tools.
Williamson calls it a “code machine gun” at times – something capable of producing huge amounts of software without a linked improvement in the ability to verify or maintain it. It means engineers end up spending most of their time saved writing code reviewing AI output and fixing it – as well as just getting up to speed with a code base.
Berlin • November 9 & 10, 2026
Close the gap between what leadership expects and what’s actually possible at LeadDev Berlin.
The report argues that measures such as defect rates and mean time to resolution provide a better indication of whether AI is actually accelerating software delivery than the number of lines of code produced, mainly because the pipeline might get bunged up otherwise.
“AI has helped the up end of the pipeline a lot,” Williamson says. “The output end of the pipeline, not so much.”