As AI tools proliferate in software upkeep, experts emphasise the importance of defining clear boundaries and disciplined processes to truly harness its cost-saving potential, moving beyond superficial automation.
AI is being sold as a direct answer to application maintenance costs, but the economics are more selective than many vendors admit. In 2026, the strongest case for AI is not that it writes more code faster. It is that it reduces the time engineers spend on repetitive diagnosis, triage, documentation and remediation, while leaving judgment, security and architecture in human hands. That model matters because maintenance spend is driven as much by investigation, testing, dependency management and technical debt as by the act of coding itself. According to ISHIR’s argument, the companies seeing the best results are those that define exactly which tasks AI can handle, which tasks it can draft and which remain strictly human-owned.
The clearest savings tend to come from operational work with predictable patterns. AI can scan logs, correlate incidents across services, classify tickets, generate starter documentation and help build tests for legacy systems that were never properly covered in the first place. It can also analyse dependency trees, surface vulnerable packages and assist with repetitive bug fixes or version migrations. The promise is practical rather than transformative: shorter diagnosis times, faster onboarding for undocumented systems and less engineering effort spent on low-risk maintenance work. That is consistent with other industry examples. AliceLabs says AI predictive maintenance in industrial settings can cut unplanned downtime by up to half and lower maintenance costs by a quarter, while emphasising the role of trained models and sensor data. Mycom describes similar benefits in network operations, where anomaly detection and root-cause analysis are used to move from reactive to predictive maintenance.
The caveat is that AI only lowers cost when the surrounding process is disciplined. Snowman Labs argues that organisations should measure the real cost of each system, remove unnecessary work and modernise where maintenance dominates, rather than assuming AI alone will fix an expensive estate. That warning is especially relevant for legacy software. The U.S. Government Accountability Office has found that federal agencies spend about 80% of their IT budgets on operations and maintenance of existing systems, which underlines how quickly minor inefficiencies can become structural cost burdens. In that context, AI can be useful as an accelerant, but it is not a substitute for governance, code quality or a deliberate modernisation strategy.
The bigger risk comes when AI is allowed to act independently on tasks that require context, judgement or business history. Changes affecting payments, authentication or regulated systems need human approval before release. So do architectural decisions such as refactoring a module, splitting a monolith or changing a data model. The same applies to security-sensitive patches and legacy business rules that may not be written down anywhere. An AI model can read code, but it cannot know why a specific discount formula exists because of a contract negotiated years ago, or why a compliance exception was encoded in an old workflow. ISHIR’s position is that the failure mode here is not dramatic; it is deferred. The code may look plausible, but the real cost appears later in production incidents, audit findings, customer complaints or rework.
That is why the question is not whether to use AI in maintenance, but how narrowly to define its role. The more mature approach is task-by-task orchestration: AI handles volume and repetition, engineers own ambiguity and risk. In that model, the right performance measures are operational rather than cosmetic. Mean time to diagnose, mean time to resolve, repeat incident rate, change failure rate, technical debt trend and total cost of ownership all matter more than the number of lines generated by a model. For many organisations, the best outcome is not fewer engineers. It is more engineering time returned to product work, with fewer incidents and a system that becomes easier to change rather than harder.
Disclaimer: This content is intended for informational purposes only. Readers are advised to exercise their own judgement, conduct due diligence, or consult a qualified expert before acting on any information provided.
