Senior engineers used to build things. At a growing number of companies, they now spend a growing share of their time fixing code that AI wrote.
A New Relic report surveying 200 US technology decision-makers found that 94% of leaders rate AI-generated code as higher quality than what their own engineers produce at the time of review. Leaders cite reasons such as cleaner formatting, more consistent style, and fewer obvious bugs. Sixty-seven percent said AI generates or significantly refactors between 51% and 75% of the code their organizations ship each week.
Then the code goes live.
- 78% of respondents reported more production incidents after deploying AI-generated code
- 86% said senior engineer "firefighting" has increased, with more time spent debugging and fixing machine-generated code
- 82% experienced at least one production failure tied to AI-generated code in the past six months
AI-generated code introduces roughly 1.7 times more critical runtime issues than peer-reviewed human code. The reason is almost boring.
AI writes code that works perfectly in controlled conditions but, as New Relic puts it, is "entirely blind to the trace"—it has no visibility into how your specific system behaves when real users interact with it. The weird edge cases, the unexpected interactions between services, the stuff that actually breaks in the real world, none of it shows up at the review stage. And because the code reads so well, many teams ship it without thorough review.
Amazon found out what that looks like at scale. In March, an AI-assisted change contributed to an outage on its North American retail platform that, according to B17News, cost millions of orders. According to The Safety Magazine, Amazon later said the root cause was really about an engineer having broader access permissions than intended, not the AI itself. But the company's response told its own story. According to reports, Dave Treadwell, Amazon's SVP of e-commerce services, ordered a 90-day safety reset and started requiring what he called "controlled friction" on changes to critical systems. Steve Tarcza, a Director of Amazon Stores, told The Register the policy is now non-negotiable: "Nothing ships without someone looking at it and validating it."
A separate Lightrun survey backs up the pattern beyond Amazon. 43% of AI-generated code changes need debugging after they're already deployed. "The volume of change is overwhelming human validation," Or Maimon, Lightrun's Chief Business Officer, told VentureBeat, "while the generated code itself frequently does not behave as expected when deployed."
Nobody disputes the speed gains. Microsoft says AI generates 30% of its code. Google says 75%. But Kent Beck, creator of Extreme Programming, identified the trade-off nobody planned for: "We're accumulating code faster than we are accumulating trust."
Tarcza seemed aware of the longer-term consequences too. "We can't get to the point where we don't have more junior engineers coming in," he said. "We have to continue to grow the talent. We can't end up in a spot where there are not folks to maintain these systems."
AI coding tools were sold as a way to free up the most experienced engineers for higher-level work. What survey data suggests they've actually created at a growing number of companies is a de facto new role nobody planned for, where the most experienced people on the team spend their weeks auditing machine output instead of building. The organizations that figure out how to capture the speed without the downstream cleanup will have a real advantage. Everyone else is quietly optimizing for the wrong end of the pipeline, measuring success by how fast code ships while the cost of keeping it running grows in the background.
