AI code still needs production debugging, report finds

Lightrun has published its State of AI-Powered Engineering Report 2026, which found that 43% of AI-generated code requires manual debugging in production.

The report is based on an independent poll of 200 SREs and DevOps leaders in the US, UK and EU, including directors, vice-presidents and C-level executives at large companies. It examines how engineering teams use AI coding assistants and AI site reliability engineering tools, and where those tools fall short.

One of the clearest findings is how much human intervention is still needed after AI-written code reaches live systems. The research found that 88% of companies require two to three manual redeploy cycles to confirm that an AI-generated fix works in production. On average, respondents said three manual redeploys are needed to verify a single AI-suggested code fix.

The study also points to a broader drain on engineering time. Developers spend an average of 38% of their week on debugging, verification and troubleshooting, or roughly two days. That suggests AI-generated output is not removing as much operational work as some teams may have expected, particularly once software is running in live environments rather than test systems.

Visibility Problem

Respondents linked many of these problems to a lack of visibility into how software behaves at runtime. Sixty per cent of SRE and DevOps leaders identified this as the main bottleneck in resolving incidents. In 44% of cases where investigations by AI SRE tools or application performance monitoring systems failed, respondents said the necessary execution-level data had not been captured.

That gap appears to be shaping how engineering leaders view the current generation of AI operations tools. The report found that 77% lack confidence in existing observability stacks to support automated root cause analysis and remediation. It also found that 97% believe AI SREs operate without significant visibility into what is happening in production.

Even during serious outages, the research suggests many organisations still rely on internal experience rather than automated diagnosis. More than half of high-severity incident resolutions, 54%, still depend on tribal knowledge instead of diagnostic evidence from AI SRE tools or application performance monitoring systems.

The findings reflect a tension in software engineering as companies increase their use of AI coding assistants while also trying to automate reliability work. Coding models can suggest patches and changes quickly, but the report indicates that validating those changes in production remains labour-intensive. For DevOps and SRE teams, the bottleneck may be shifting from writing code to proving that code behaves as intended under live conditions.

Lightrun argues that the issue is not simply code generation, but whether AI tools can observe real systems at the level needed to identify causes and test remedies. Without live runtime data, AI agents are left to infer what happened rather than confirm it directly.

This sits within a wider industry debate about observability and automation. Many engineering organisations have invested heavily in logs, metrics and traces, yet the survey suggests senior leaders still see gaps when trying to use those tools for automated diagnosis and incident response. If the findings are representative, vendors offering AI assistants for operations may face pressure to show not only that they can propose fixes, but that they can verify them against live execution data.

Lightrun commissioned the research with independent firm Global Surveyz. The report examines the AI-powered software development lifecycle and the role of runtime data in validating AI-generated code and AI-led incident response.

In a statement accompanying the findings, Ilan Peleg, chief executive of Lightrun, set out the company's view of the challenge facing engineering teams as AI use expands.

"Engineering organizations need runtime visibility to embrace the possibilities offered by AI-accelerated engineering. Without this grounding, we aren't slowed by writing code anymore, but by our inability to trust it," said Peleg.

"When almost half of AI-generated changes still need debugging in production, we need to fundamentally rethink how we expect our AI agents to solve complex challenges," he added.