In 2024, the DORA State of DevOps Report found that a 25% increase in AI adoption was met with a corresponding 1.5% decrease in software delivery throughput and a 7.2% decrease in delivery stability.[1] AI, said the report, does "not automatically improve software delivery". In 2025, the same report found that when teams used AI more effectively, the throughput issue flipped from negative to positive, but that delivery stability remained in the negative.[2] In June 2026, the GitClear longitudinal analysis of 623 million code changes from 2023-2026 found that refactoring was down by 70%, but that code block duplication rose by 81%, and error-masking constructs rose by 47%. Translated, it means that while AI doesn’t write bad code necessarily, the maintenance that keeps codebases healthy is being neglected.[3]
“Inside many companies rushing to adopt AI coding assistance, there is a new category of debt building,” says Augustine Tumi Mogashoa, IT and Business Continuity Management Specialist at ASQE. “It is a concern that South African companies can’t afford to miss, especially as coding assistants are increasingly becoming standard issue across development teams and are valued for how much code they can generate in a day.”
However, volume isn’t the right metric against which code quality should be measured. The real metric is whether or not the code can be trusted, tested and maintained once it leaves the assistant and enters a live system. As Mogashoa highlights, while coding assistants can increase output, faster code generation doesn’t necessarily create better software.
“This is not an argument for avoiding AI tools; it is rather asking that the acceleration taking place without automated testing, disciplined version control, small changes and fast feedback simply moves the bottleneck and the risk downstream,” she says.
The problem is not the tools but what companies are doing (or not doing) around them. Some teams are using AI to increase code volume without increasing review, test automation or architectural discipline, and the result is deceptively fast progress at the front of the delivery pipeline and a growing backlog of defects, integration failures and remediation work in the back end.
This backlog is also often a risk that takes the business unawares. AI-assisted development can look like genuine progress for months before the debt surfaces, usually at the worst possible time and at scale. Mogashoa points to a familiar failure pattern: “Companies are layering new AI capabilities onto old and unstable foundations, and they are caught out where old and new meet. AI is placed on top of ageing systems, fragmented data and undocumented rules while leaders assume the new interface has modernised what lies beneath.”
Software written faster than it can be reviewed, tested and understood is shipped, integrated and then used by companies to make decisions about customers and pricing. If a company cannot say how much of its codebase was generated by AI or how much of that code has undergone the same review and testing discipline as everything else, it is carrying a risk. The fix isn’t to slow AI adoption but instead to insist that speed and assurance grow together.
“Companies need automated pipelines that validate every small change, deeper reviews for higher-risk code and monitoring that continues long after release,” she says. “Companies can thrive with AI if they have the strongest foundations for trusting how their AI models are used.”
[1] https://dora.dev/research/2024/dora-report/
[2] https://cloud.google.com/resources/content/2025-dora-ai-assisted-software-development-report
[3] https://www.gitclear.com/the_ai_code_quality_maintainability_gap
Share
