Abstract: AI-assisted programming tools are increasingly adopted to boost developer productivity by automating routine tasks and accelerating software development. While prior evidence emphasizes short-term productivity gains, we show that heavy reliance on AI-assisted coding can accelerate technical debt accumulation, slowing software development, increasing maintenance costs, and eroding revenues over time. We develop a dynamic simulation model grounded in the literature and in semi-structured interviews we conducted with practitioners across a wide range of industries and organizational levels. The simulation traces how performance evolves over time, while an analytical characterization of the model’s equilibria establishes when and why tipping points arise. We use both to examine the long-run implications of AI assistance across project horizons, levels of legacy-system interaction, developer skill, and rates of market change.
Our results indicate an inverted-U relationship between AI assistance and performance, with a “sweet spot” in which AI improves productivity and profitability, beyond which compounding technical debt undermines system maintainability and the sustainability of service offerings. This decline can take the form of a tipping point under a wide range of conditions. Once AI assistance pushes debt past this point, performance eventually collapses, often after a delay that masks the risk. Reversing the collapse requires cutting AI assistance well below the level that triggered it or a substantial investment in debt retirement.
We further find that environments with highly skilled developers, limited legacy-system interaction, and rapidly evolving markets can sustain higher levels of AI assistance, whereas less skilled developers, extensive legacy-system interaction, and slowly evolving markets increase vulnerability to long-run decline. We also show that proactive technical-debt retirement serves as a strategic guardrail that raises the level of AI assistance a firm can sustain. Finally, the results carry implications for job design. Replacing senior developers with AI-assisted junior developers can backfire. Reducing the workforce to capture AI productivity gains can lower profits at moderate levels of AI assistance while also weakening the pipeline of future senior developers. Taken together, these findings suggest that AI-assisted development is not merely a productivity tool but a strategic choice shaped by system longevity, skill composition, legacy constraints, and market conditions.
Operations and Information Management Department