AI is writing more code than ever, but is it good enough?

In 2024, a DORA State of DevOps Report revealed that AI adoption surged by 25%, but this surge came with a 1.5% drop in software delivery throughput and a 7.2% decline in delivery stability. The report suggested that AI doesn't inherently enhance software delivery. By 2025, the same report indicated that when teams effectively utilized AI, the throughput issue reversed, but stability continued to decline.

In June 2026, a GitClear analysis of 623 million code changes from 2023-2026 showed that refactoring decreased by 70%, while code duplication surged by 81% and error-masking constructs rose by 47%. The implication is that AI may not write poor code, but the maintenance that keeps codebases robust is being overlooked.

Augustine Tumi Mogashoa, an IT and Business Continuity Management Specialist, noted that many companies rushing to adopt AI coding assistance are inadvertently creating a new form of debt. This concern is particularly pertinent to South African companies, given the widespread use of coding assistants across development teams. However, volume isn't the key metric for evaluating code quality. The crucial metric is whether the code can be trusted, tested, and maintained once it moves from the assistant to a live system.

Mogashoa emphasized that while coding assistants can boost output, faster code generation doesn't guarantee better software. The issue lies not in the tools but in how companies use them. Some teams are leveraging AI to boost code volume without increasing review, test automation, or architectural discipline, leading to a deceptive speed in development and a burgeoning backlog of defects, integration failures, and remediation work.

This pattern of rapid progress followed by a backlog of issues often surfaces at the worst possible time and scale. Mogashoa cited a well-known failure pattern: companies are adding AI capabilities to old and unstable systems, catching them off guard when the old and new systems interact. AI is being placed atop outdated systems, fragmented data, and undocumented rules, assuming that the new interface has modernized what lies beneath.

The root of the problem is not the AI tools themselves but how companies employ them. The fix isn't to halt AI adoption but to ensure that speed and assurance grow in tandem. Companies need automated pipelines that validate every small change, deeper reviews for higher-risk code, and ongoing monitoring after release. "Companies can succeed with AI if they establish robust foundations for trusting how their AI models are used," Mogashoa concluded.