Enterprise AI coding is cheap to start and expensive to sustain. A study of 30,000 enterprise systems released by Software Improvement Group on Wednesday shows that productivity gains from AI-assisted development disappear once a codebase exceeds 100,000 lines, while agentic coding runs can consume up to 1,000 times more tokens than standard code chat, turning modest-seeming projects into very large invoice items.
The numbers from the State of Software 2026 report are striking in their specificity. In one documented case study, autonomous AI agents assembled an entire system in a week. The token bill came to between €10 million and €15 million, and the resulting code was described as nearly unmaintainable. AI-generated code now accounts for 1.9% of enterprise production code in SIG's benchmark, but it carries roughly double the security risk violations of human-written code, with over half containing at least one vulnerability.
The picture across enterprise codebases more broadly is not encouraging. SIG's benchmark, which spans over 400 billion lines of code, found that 86% of code falls below its recommended maintainability rating, 71% has a low degree of security controls, and 50% scores below the recommended architecture threshold. Enterprises with stronger architecture resolve issues 30% faster; those that reduce code-level technical debt stand to save roughly €870,000 per system per year in developer time.
The central argument is that AI does not add or subtract from an organisation's engineering discipline. It multiplies it. A team with strong governance and measurement in place will find AI accelerates delivery. A team without those foundations will find AI accelerates the accumulation of debt, cost, and security exposure.
"Nothing in this report is an argument against AI. The productivity gains are real, and the organisations that fail to embrace it risk falling behind those that learn to use it effectively. But you cannot manage what you do not measure, and you cannot sustain speed on a foundation you do not understand. When code generation outruns governance, technical debt accumulates faster, security exposure widens, and the systems a business depends on become harder to maintain and evolve," said Luc Brandts, CEO at Software Improvement Group.
The report also flags that AI systems themselves have a quality problem: 72% of AI systems in production score below SIG's recommended build-quality rating. The full State of Software 2026 report draws on systems analysed over the past year within SIG's benchmark of over 30,000 systems and 400 billion lines of code.
To stay across the latest in cloud, AI and enterprise tech analysis from Compare the Cloud, subscribe to our weekly newsletter at https://www.comparethecloud.net/newsletter
