AI coding assistants were supposed to help software teams ship faster. A new survey of engineering leaders suggests many teams are instead spending their saved time cleaning up after the code.
In a report from debugging-software company Undo, highlighted this week by the developer news site LeadDev, 81% of 300 senior engineering leaders said their organization had a production incident or service outage linked to AI coding tools in the past six months. The respondents work at large U.S. and UK companies that build mission-critical software.
The Key Numbers
- 93% said an AI “hallucination” had led to the wrong diagnosis of a problem’s root cause.
- More than 90% had serious defects, poorly optimized code or test escapes reach production.
- 79% said their release cycles are no faster than before AI.
- 16.9 hours a week: the average time their engineers spend debugging, versus 9.8 hours writing code.
- 35% of AI-generated code reaches production before the team fully understands it.
Undo’s chief technology officer, Mark Williamson, told LeadDev that AI has sped up the front of the software pipeline far more than the end of it. “The output end of the pipeline, not so much,” he said.
A Second Survey Points the Same Way
Undo isn’t the only company reporting this. In August, a Wakefield Research survey of 400 business and engineering executives, done for testing company Sauce Labs, found that 80% had traced a production incident, outage or customer-facing defect to AI-generated code in the previous 12 months. Two-thirds in that survey admitted cutting corners on quality or testing to hit deadlines, and 53% said they had shipped with known testing problems.
Yet the same executives were upbeat: 89% reported a positive return on their AI tools, and 91% said software quality had improved. Put the two surveys side by side and the picture is less “AI code is broken” and more “AI makes so much code that checking it has become the bottleneck.”
Why It Matters Beyond Tech
You don’t have to write code for this to touch you. The banking apps, airline systems and online stores Americans rely on are built by exactly these kinds of teams. More code, reviewed less carefully, raises the odds of glitches and outages, and both surveys point to a fix that is unglamorous: more testing and more human review, not less.
For small businesses using AI to build websites, plugins or simple tools, the lesson scales down. Test changes before they go live, keep backups, and make sure someone understands what the code does before customers depend on it.
The Catch
Both surveys come from companies that sell debugging or testing products, so they have an interest in the problem looking big. Neither is a random sample, and the reports don’t explain in detail how responses were collected or what counted as an “AI-linked” incident. Treat the exact percentages with caution; the consistent direction is the more meaningful signal.
More on Contoh
- Just 10 Minutes of AI Help Made People Quicker to Give Up, Study Finds
- AI Still Struggles With Small Business Bookkeeping, a New Benchmark Finds
- Microsoft’s Satya Nadella Says AI Needs an “Emergency Brake” Humans Control