Somewhere right now, a mainframe leader is sitting in a meeting where an executive is asking about the company’s AI strategy. And that leader (who manages a platform processing millions of transactions daily on the most reliable infrastructure in the enterprise) is expected to have an answer that sounds exciting and forward-thinking. The problem? The person asking the question usually has no understanding of what AI would actually mean in a mainframe environment.

This is playing out across industries. Leadership pushes down AI mandates without understanding the platform, the security implications, or even what specific problem they’re trying to solve. VPs of IT are fielding executive teams who are “hyped up” about AI but can’t articulate what the business needs. In some organizations, senior leaders have privately admitted they just need to buy some AI software to check a box. There’s nothing strategic about it. The pressure is purely about the optics, but the pressure is real none-the-less.

There is a light-hearted term for this: “management by In-Flight Magazine,” where executives read about the latest trend at 30,000 feet, and by the time they land, they’ve defined directives without understanding what’s technically realistic. Others describe the phenomenon as simply FOMO, with leaders afraid of being left behind rather than first understanding what they’re chasing. The irony is hard to miss: an experienced DBA would first thoroughly plan the expected outcome of releasing the newest version of a database platform into a production environment containing mission-critical data. But somehow organizations are willing to turn over that same mission-critical data to freshly released AI models without first understanding the outcome they’re seeking to achieve or how it achieves the organization’s strategy.

The security perspective is especially concerning, but it’s the part of the conversation that tends to get the least amount of attention in executive sessions. What happens when employees start feeding proprietary code, customer data, or internal business processes over to AI tools? In industries like financial services and insurance, where regulations abound, these are tangible risks. For years, a company’s proprietary data has been considered as priceless as real capital. The intellectual property that differentiates one company from another could be exposed without anyone fully realizing it happened.

Then there’s the practical reality of AI on the mainframe itself. AI tools perform reasonably well for widely used languages like JavaScript and Python, where training data is abundant. But mainframe-specific languages like high-level assembler or deeply customized COBOL don’t have the same foundation. The AI simply hasn’t been trained on enough of that code to produce useful results. On top of that, the quality of existing mainframe code varies wildly. Training a model on decades of legacy code with inconsistent standards and old logic doesn’t produce better code. Many times, it simply reproduces the problems faster.

None of this means AI is useless for mainframe teams. There are real productivity gains from using AI on everyday office tasks like email, documentation, and spreadsheets, freeing up more time to focus on actual mainframe work. That’s a legitimate and low-risk application. The trouble starts when the mandate expands beyond that into territory where the costs, risks, and technical limitations haven’t been thought through.

The cost dimension alone deserves more scrutiny. CPU time on the mainframe is expensive, and training AI models is extraordinarily compute-intensive. IBM’s own hardware strategy for the Z17 reflects this: the platform is designed for running pre-trained models and doing inferencing close to the data, not for building the models themselves. That’s a meaningful distinction that most executive-built AI mandates don’t account for.

So what should mainframe leaders do with all this pressure? Here’s a few principles worth considering. For starters, lead with the data story. The mainframe is often the system of record housing some of the most business-critical data in the organization. Help leadership understand the value that already exists before layering on new tools. Build governance and security guardrails before deploying anything. And respond to vague mandates with specific questions: What problem are we solving? What data is involved? Who owns the risk?

The mainframe has outlasted every hype cycle the industry has thrown at it. AI won’t be the exception. Instead, it will be the genesis of true evolution for how the mainframe is leveraged within every organization’s ecosystem. And the leaders who manage these platforms and understand them well need a seat at the table to guide the conversation with the expertise that a true AI strategy requires.