Rob Versaw, bridging innovation and business impact, a leader in technology, product and strategy.
Generative AI tools like GitHub Copilot and Amazon CodeWhisperer are changing how software is written. As a former front-line developer, I remember spending days writing and refining code that, today, can be generated in minutes.
These tools offer real-time suggestions, reduce boilerplate and significantly accelerate development workflows. They represent a breakthrough in how engineers interact with code—but they also reflect a strategic misalignment.
AI’s Growing Use In Software’s Initial Development
Having consulted several tech companies, I can confidently say that current AI usage in software tooling is heavily focused on the initial development phase—the most visible, but the least costly, stage of the software life cycle. Decades of industry research make this clear: Only 20% to 30% of a software system’s total cost is incurred during development, while 70% to 80% is spent on maintenance, operations and evolution over the system’s lifetime.
This cost distribution has been validated across a range of studies. Analyses from PwC and Carnegie Mellon University estimate that approximately 70% of software costs emerge post-deployment. The European Commission’s Conference on Software Maintenance places the figure closer to 79%, while both IDC and Forrester Research have cited maintenance and operations consuming as much as 70% to 80% of the total cost of ownership. Although these studies were conducted prior to the widespread use of AI in development, their consensus remains clear: The vast majority of long-term cost stems from what happens after code is shipped.
Yet, today’s generative AI landscape remains overwhelmingly focused on optimizing the initial 20%. Recently, Satya Nadella said that 30% of Microsoft’s new code is written by AI and increasing!
Maintenance: The Most Costly Part Of Software Development
This disconnect represents more than a missed opportunity. If the true financial and operational burden lies in maintaining, updating and operating systems, then the most strategically valuable role for generative AI is not just in the development environment, but throughout the entire life cycle. From infrastructure management and incident response to compliance automation and system monitoring, this is where AI has the potential to radically improve software economics—and where the next wave of innovation should be focused.
Software maintenance includes far more than bug fixes. It spans enhancements, performance tuning, adapting to new platforms and compliance with evolving regulatory regimes. The IEEE Standard 1219 defines four maintenance categories: corrective (fixing defects), adaptive (responding to environmental changes), perfective (enhancing usability/performance) and preventive (reducing future risk). Of these, perfective and adaptive changes often consume the majority of maintenance budgets—surpassing 50% of total effort, according to IEEE data.
Why is maintenance so costly?
• Environmental Drift: Operating systems, APIs, libraries and hardware evolve. Code written even five years ago may rely on deprecated dependencies or unsupported services.
• Security And Compliance: Frameworks and configurations must adapt to new security vulnerabilities and changing compliance requirements (e.g., HIPAA, SOC 2, etc.). These updates often require systemwide audits and recertification.
• Complex Integration Layers: As systems become more modular—using microservices, cloud functions and third-party SaaS components—the coordination burden increases. Diagnosing an issue in a distributed system is harder than in a monolith.
• Technical Debt: Trade-offs made during fast-moving development cycles accumulate interest. Eventually they must be repaid—often at great cost.
These forces mean that the cost of maintaining and evolving software typically dwarfs the cost of building it.
Making The Most Of AI’s Efficiency Gains
Generative AI is excelling at the most structured part of the engineering workflow: writing new code. GitHub Copilot, trained on billions of lines of code, helps developers complete functions, generate test cases and write documentation. In one Microsoft study, developers using Copilot completed tasks up to 55% faster.
This efficiency gain is real—but it addresses the smallest slice of the long-term cost structure. The current generation of generative AI tooling does little to address issues such as legacy system modernization; security patch prioritization; config drift detection and correction; monitoring, alerting and incident response; or compliance updates and evidence collection.
In short, AI has made engineers faster—but it hasn’t yet made systems easier to maintain. And that’s the battleground where margins, uptime and user trust are truly won.
Even infrastructure-as-code tools are adopting AI to reduce maintenance friction. Generative models trained on Terraform and Kubernetes configurations are increasingly able to auto-generate compliant infrastructure code, detect violations of policy-as-code frameworks and forecast the blast radius of a proposed change.
This is where generative AI begins to shift from a productivity play to a risk management and cost optimization strategy. If 80% of software cost stems from maintenance, then technology leaders must rethink how they measure the ROI of AI. While boosting developer productivity is valuable, the far greater return lies in making systems easier to maintain and scale.
This shift demands investing in AI that works on production telemetry, not just source code. It also requires aligning AI efforts with infrastructure, compliance and observability teams and treating post-deployment operations as core use cases, rather than peripheral concerns.
Generative AI is changing how we build software. But to move the needle on cost, risk and resilience, the real opportunity is in how we maintain it. The future of AI in engineering lies not in code completion, but in system continuity and maintenance.
Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?
