The retirement narrative keeps not happening
Every few years, a wave of vendor marketing declares the mainframe finally dead, and every few years the installed base quietly proves that declaration premature. BMC's latest annual survey of 1,300 mainframe professionals puts a hard number on that persistence: 94 percent expect long-term investment in the platform to continue, and only a small minority anticipate meaningful reduction in mainframe reliance over the coming years. That is not a nostalgia number driven by sunk-cost thinking among aging staff. It reflects a practical reality for CIOs at banks, insurers, retailers, and logistics firms who run core transaction volume through mainframe systems that remain cheaper to modernize incrementally than to replace outright.
That reality holds especially now that AI tooling has started closing the skills and interface gap that made mainframe careers feel like a dead end to younger engineers entering the workforce. A full rip-and-replace of a core banking or claims platform can run into the hundreds of millions of dollars and take years, a timeline few boards are willing to fund without a clearer payoff than modernization in place already offers on a much shorter horizon, and with far less risk of the multi-year outages that failed replacement projects tend to produce when a fixed deadline slips repeatedly and the original business case quietly stops making sense to the executives who originally signed off on it.
AI is the modernization strategy, not a bolt-on afterthought
45 percent of respondents named AI tool adoption a top priority for their mainframe environment this year, and the survey shows real budget commitment sitting behind that answer: more than a third plan to invest in building internal AI agents for mainframe management, while 32 percent plan to bring in third-party agents to handle the same functions instead of building in-house capability from scratch. The split between build and buy roughly mirrors how enterprises are approaching agentic AI everywhere else, which suggests mainframe teams are not treating this as a niche problem requiring a bespoke playbook.
John McKenny, BMC's senior vice president and general manager, described this shift as an industry moving beyond AI experimentation toward trusted, operational adoption. That distinction matters for CIOs weighing pilot budgets against production deployment commitments: the mainframe world appears to have moved past the pilot-purgatory phase that Gartner and other analysts have flagged this year as the norm across most other enterprise AI initiatives still stuck in evaluation, unable to clear the bar into funded production work.
Securing agent access is the harder problem
The survey findings point toward a specific technical focus that shifted from prior years: mainframe leaders are increasingly using AI to secure connections for both human and application access, a change from the AI translating or rewriting legacy COBOL code use case that dominated discussion throughout 2024 and 2025. That earlier code-translation focus still continues in many shops today, though it has clearly lost its position as the headline use case driving new budget requests.
That shift matters because access governance, not code translation, is where most mainframe security incidents actually originate in practice. An AI agent granted broad access to a core banking or insurance claims system represents a materially different risk profile than a chatbot summarizing customer support tickets, and this survey suggests mainframe teams are now treating agent access with a correspondingly higher level of scrutiny and control than most enterprise cloud environments currently apply to comparable agent deployments.
Trust has to be earned before autonomy is granted
Steven Dickens, CEO of HyperFRAME Research, described the shift by saying the platform is being actively architected to lead in the era of AI, well beyond simply surviving alongside newer infrastructure choices. McKenny added that the path to greater AI autonomy on the mainframe will be earned through trust, language that captures the sequencing most enterprise IT leaders are now applying broadly across agentic AI deployments. That sequencing means proving reliability on narrow, auditable tasks first, before expanding any given agent's scope of action into higher-stakes territory once the earlier, smaller wins have held up under real operating conditions for a meaningful stretch of time, typically measured in quarters rather than weeks.
For CIOs managing core systems that cannot tolerate downtime, this cautious sequencing is the practical path regulators and boards are willing to accept today. The mainframe world's decades of experience with formal change control processes may end up becoming a template other parts of the enterprise quietly borrow from as they scale their own agentic AI programs past the pilot stage in the coming year, since those change control disciplines were built for exactly this kind of high-stakes, low-tolerance-for-error environment where a single bad release can take down a core system for hours.
The takeaway for enterprise architecture roadmaps
For CIOs still treating the mainframe as a line item to eliminate rather than an asset worth modernizing in place, this survey is a data point worth taking seriously. Ninety-four percent of the people who actually operate these systems day to day do not see full retirement as the realistic plan; they see AI-assisted operation as the path their organizations are actually funding and executing against right now, with budget already committed rather than sitting in next year's wish-list column.
That has direct implications for platform strategy conversations happening alongside ERP and core system modernization decisions elsewhere in the enterprise. The choice most organizations face is rarely mainframe versus cloud framed as an absolute either-or decision. It is closer to identifying which specific workloads genuinely benefit from migration, and which are better served by AI-augmented operation on infrastructure that already runs reliably at scale and carries decades of institutional tuning that a fresh cloud rebuild would take years to replicate, if it can be replicated fully at all without introducing new and unfamiliar failure modes that nobody on the current team has ever had to diagnose before.
