AI coding tools caused outages at 81% of large firms, survey of engineering leaders finds

AI coding tools were sold as a way to ship software faster. A new survey suggests many companies are getting AI code production outages instead, along with more bugs and more debugging. In a report from the debugging company Undo, 81% of senior engineering leaders said their organization had a production incident or service outage linked to AI coding tools in the past six months. The findings drew fresh attention this week after LeadDev covered them on October 12, and they land at a moment when nearly every software team is under pressure to use AI for more of its code.

What the survey found

According to LeadDev, Undo surveyed 300 senior engineering leaders at large US and UK companies that build mission critical software on complex codebases. The questions ran during July and August 2026. VMblog, which reported the release in late September, said the full report is titled "Overcoming the limitations of coding agents in complex software systems."

The headline numbers are blunt. Over the past six months, because of AI coding tools:

• 81% of organizations had a production incident or outage that affected internal users or customers.
• 93% had a problem's root cause misdiagnosed because of an AI hallucination.
• 91% saw test escapes, serious defects or poorly optimized code reach production.

The survey also asked about speed. 79% of leaders said their release cycles are no faster than they were before AI. Engineers on their teams spend an average of 16.9 hours a week debugging, which VMblog notes is about 42% of a working week. LeadDev reported that those same engineers spend 9.8 hours a week actually producing code. And 35% of AI generated code reaches production before the team fully understands it.

Why AI code production outages happen

The pattern in the data is not that AI writes useless code. It is that AI writes a lot of code, quickly, and the work of checking it has not kept up.

Mark Williamson, Undo's chief technology officer, put it this way to LeadDev: AI "has helped the up end of the pipeline a lot," but the output end "not so much." He has also called AI a "code machine gun," a tool that produces huge amounts of software without a matching improvement in how that software is verified or maintained. In his view, the bottleneck in software delivery has moved from writing code to finding and fixing what went wrong.

Greg Law, Undo's founder and CEO, framed the issue differently in the company's release. "AI doesn't have an intelligence problem; it has an evidence problem," he said, according to VMblog. His point is that coding agents often guess about how a program behaves instead of looking at what it actually does when it runs.

Old systems hold hidden rules

LeadDev's report highlights a problem that anyone who has worked on a large codebase will recognize. Enterprise systems carry years of undocumented assumptions and dependencies. Often, the only record of why something works the way it does lives in the heads of a few senior engineers. An AI model has no access to that history, so it can write code that looks correct and still breaks a rule nobody wrote down.

That risk is highest in systems where small mistakes spread fast, such as trading platforms and networking infrastructure. A misread edge case in a hobby app is an annoyance. In a payment system it can be an outage.

Hallucinations make debugging harder

The 93% figure may be the most worrying number in the report. It suggests AI is not only introducing bugs but also sending engineers in the wrong direction when they try to fix them. If an assistant confidently names the wrong cause, a team can lose hours chasing it. This echoes problems we have seen elsewhere, including when Google paused its open source bug bounty after a flood of AI generated reports that took time to sort through.

Who is affected

Engineering managers are the obvious group. If release cycles are not getting faster, the business case for paying for AI coding tools gets harder to make. Undo's survey also found that 82% of leaders expect coding agent costs to rise sharply, according to VMblog.

Developers are affected too, and not only through longer debugging sessions. LeadDev notes that engineers who maintain AI generated code may never build the deep knowledge the original author would have had. That raises a long term question about how junior developers learn when much of their work is reviewing machine output. It connects to recent research on how AI help affects people, such as the Berkeley study on AI and persistence, which found that short periods of AI assistance made people give up on hard tasks sooner.

Customers are the last link in the chain. When 81% of surveyed organizations report an incident affecting users, some of those users are people trying to log in, pay a bill or place an order.

What engineering teams can do

The report and the experts quoted in it point to a few practical steps.

Measure what matters

Williamson's advice, as reported by LeadDev, is to stop judging AI by how much code it writes. Track defect rates and mean time to resolution instead. If those numbers are getting worse, the tool is not saving time, whatever the line count says.

Make understanding a release gate

With 35% of AI code shipping before teams fully understand it, a simple rule helps: nobody merges code they cannot explain. Code review for AI output should be at least as strict as review for a new hire's first pull request.

Give agents real evidence

Undo's survey found that 82% of leaders believe coding agents would be far more useful for understanding and debugging if they were grounded in runtime context, meaning data about how the program actually behaves. Whatever tools a team uses, feeding agents logs, traces and test results tends to beat asking them to reason from source code alone. Our guide to AI agents covers where these tools work well and where they still fail.

Keep a human brake

Microsoft CEO Satya Nadella recently argued that AI systems need pause controls and audit logs, which we covered in Nadella's emergency brake proposal. The same idea applies to code: keep rollback paths fast and keep a person accountable for every change that reaches production.

The limits of this data

It is worth reading the survey with some care. Undo sells debugging tools, so it has an interest in showing that debugging is a growing problem. The respondents were senior leaders at large companies with complex codebases, not a cross section of all developers, and the public coverage does not describe how the sample was chosen. The numbers are self reported, and "linked to AI coding tools" is a judgment call that different leaders may make differently.

Still, the findings line up with what many teams say privately: AI makes the first draft cheap, and everything after the first draft is where the real cost shows up.

What to watch next

Expect more vendors and research groups to publish their own numbers on AI code quality, and expect engineering leaders to push AI vendors for better ways to verify output. The next big question is whether coding agents that can run and inspect programs, rather than only read them, will bring defect rates down. For more coverage, visit our AI news page.

Frequently asked questions

What percentage of companies had outages from AI code?

In Undo's survey of 300 senior engineering leaders, 81% said their organization had a production incident or outage linked to AI coding tools in the past six months.

Does AI make software releases faster?

Not for most respondents. 79% said their release cycles are no faster than before AI, largely because debugging time has grown.

Why does AI code cause bugs in large systems?

Large codebases contain undocumented rules and dependencies that AI models cannot see. Code can look correct and still break assumptions only experienced engineers know about.

How can teams reduce AI coding risks?

Track defect rates and time to resolution, require engineers to understand code before merging it, and give AI tools runtime data such as logs and test results.