One thing I have learned from watching technology evolve over the years is that we tend to underestimate the ability of established systems to change. Every new wave brings predictions about what will become obsolete, and yet the technologies that continue to matter often find ways to evolve alongside what comes next.
I have seen that pattern through client-server computing, distributed systems, the internet, cloud, and hybrid cloud. And now AI. Each changed how enterprises operate, sometimes quite dramatically. But those changes rarely unfolded as neatly as the predictions suggested. More often, the technologies that mattered most adapted to new expectations and found new ways to remain essential.
The mainframe may be one of the best examples.
The mainframe has adapted through each of these shifts. Client-server changed how applications interacted with it, while hybrid cloud expanded its connection to the broader enterprise. What interests me about AI is that the change now reaches beyond how systems connect. It is beginning to influence how people understand their environments, make decisions, and carry out their work.
That has led me to think differently about what it means to Agentify the Mainframe. We have already seen what AI can do to help people find information and understand complex problems. The next question is how we give AI the expertise, context, and appropriate controls to take part in the work itself.
As AI agents become more capable of investigating problems and advancing tasks, that question becomes increasingly important. We are entering a new era of agentic AI for the mainframe, where AI can participate in real mainframe work. We need to think carefully about what it takes to do that responsibly.
The Mainframe Has Already Proven Its Relevance
I find it interesting that conversations about AI and the mainframe still sometimes begin with the question of modernization, as though the platform needs AI to justify its place in the enterprise. That overlooks what is actually happening in organizations today, and our latest research offers a useful reminder.
In the 2026 BMC Mainframe Survey, 94% of respondents said they view the mainframe as a long-term platform for new workloads, and the same percentage said they continue to invest in it. Sixty-nine percent reported growing general-purpose mainframe capacity. Transaction volumes, data volumes, and database activity are growing as well.
What I find particularly telling is what these numbers reveal about the role the mainframe continues to play. Organizations are placing greater demands on these systems, not fewer. They are handling more workloads, more transactions, and more data, often while taking on additional responsibilities across the enterprise.
That leads me to a different question.
Rather than asking how AI might make the mainframe relevant again, shouldn’t we be asking whether AI is ready to meet the expectations of the mainframe?
When a system is responsible for some of the most consequential work in banking, insurance, government, healthcare, travel, retail, and global commerce, intelligence alone is not enough. Knowing something about the mainframe is not the same thing as understanding the mainframe you are responsible for.
And understanding matters a great deal once AI begins doing more than answering questions.
Knowing Mainframes Is One Thing. Understanding Yours Is Another
There is something we can learn from the way an experienced mainframe professional approaches a problem in their own environment. They rarely start with one piece of information.
They bring years of product and domain knowledge. They know the architecture. They know the application. They know what happened the last time something similar occurred. They know the operating practices of their organization, the runbooks, the exceptions to the runbooks, and sometimes the reason those exceptions exist in the first place.
And then they look at what is happening right now. That combination is what produces judgment. I think we sometimes expect AI to take a shortcut around that.
A general-purpose AI model can know an enormous amount about mainframe technology. But it does not inherently know your operating practices, your history, your applications, your workloads, your databases, or what changed twenty minutes ago. For AI to participate meaningfully in mainframe work, it needs context.
For us, that context comes from three places: specialized mainframe expertise, the institutional knowledge an organization has accumulated over time, and live operational information about what is happening now.
When we bring those three sources of knowledge and context together, AI can begin to understand not only how mainframes work, but how your mainframe works, including the conditions and practices that make your environment unique.
I believe that distinction is fundamental to making AI truly mainframe-native.
Expertise Shouldn’t Stay With the Few Who Have It
One thing I have come to appreciate is just how much an organization depends on the experience of its people. In complex environments like the mainframe, some of the most valuable expertise has been accumulated over decades, and much of it has never been formally documented.
It lives in people. In experience. In habits. In the way someone recognizes a pattern because they have seen it three times before. In knowing that the obvious answer is not always the right one. In understanding why a system was designed a particular way ten or fifteen years ago. Some of that knowledge gets written down. A lot of it does not.
That creates an interesting challenge as generations of practitioners change. We often describe it as a skills shortage, but I think that description is too narrow. The bigger question is how accumulated expertise travels from one generation to another without losing what made it valuable in the first place.
The opportunity with AI is not simply to fill a skills gap. It is to preserve mainframe expertise, make it accessible, distribute it into the flow of work, and eventually help more people put it to use. That is an idea we have been spending a great deal of time on at BMC.
The goal is not to replace the expert. It is to multiply the impact of the expert.
Today, expertise is often something you have to find. Who understands this system? Who has seen this issue before? Who knows why this job behaves differently at month-end? Who remembers what happened the last time we made this change?
Now imagine that much of that expertise can increasingly be available at the moment someone needs it, informed by the condition of the system in front of them.
I think the greater opportunity is to make that hard-earned expertise available wherever it is needed, so its value is no longer limited to the few people who carry it.
Thanks For the Answers, AI. Now Let’s Get Some Work Done.
One of the most valuable things generative AI has given us is the ability to make complex information easier to understand and act upon. That is no small thing.
Understanding unfamiliar code faster, explaining a complex operational condition, finding knowledge buried across documentation, or helping someone make sense of an unfamiliar system can save enormous amounts of time. But eventually there is another question. What happens next?
Someone still has to investigate. Gather evidence. Move between systems. Decide whether the recommendation makes sense. Determine what is permitted. And, in many cases, carry out the next step.
That is where the shift toward agentic AI for the mainframe becomes interesting.
An agent can move beyond simply providing information and begin participating in a defined task: investigate something, assemble evidence, analyze it, recommend what should happen next, and, where appropriate, move the work forward within established boundaries.
At BMC, we have been building toward this progression deliberately.
BMC AMI Assistant helps people understand, investigate, and decide by bringing intelligence and expertise into the flow of work.
BMC AMI Agentics takes the next step by putting specialized digital workers to work on defined mainframe tasks.
We are beginning with BMC AMI Agentics for Ops and BMC AMI Agentics for Data, bringing specialized AI experts to the kinds of problems these teams deal with every day. In operations, that might mean investigating a CPU issue, explaining unfamiliar REXX, or helping someone work through an incident. For a data professional, it could mean analyzing a SQL workload or investigating what is happening across Db2 or IMS.
What matters to me is the specialization. These are not general-purpose agents trying to understand every mainframe problem. They are purpose-built AI agents, grounded in the expertise appropriate to the work in front of them. For operations teams, that means AI agents for mainframe operations are working from the context of the problem rather than approaching it as a generic AI task.
And the idea extends beyond Ops and Data. BMC AMI Assistant is now integrated with Microsoft Teams, bringing trusted mainframe knowledge into a collaboration environment people already use. For mainframe development teams, AI-assisted hybrid CI/CD connects AI coding assistants to governed mainframe build, promote, and deploy workflows.
But the product names are not really the most important part of the story. The larger change is that intelligence is starting to participate in work.
There is another aspect of this that I believe matters just as much. Organizations have spent years building their mainframe environments around technologies, processes, and expertise they trust. Our approach is to bring these new AI capabilities into that existing BMC AMI environment, allowing organizations to build on what they already have rather than introducing another disconnected layer of technology.
And once that happens, trust becomes far more important.
Trust Is What Gives AI the Right to Act
The growing capabilities of autonomous AI agents are genuinely exciting, and I can understand why organizations are eager to explore what they might make possible. But experience has taught me that being able to do something and being ready to take responsibility for it are two very different things.
Just because an AI system can take an action does not mean it should.
Anyone who has led a large team has probably experienced some version of this with people as well. You do not usually give someone unlimited authority on their first day. Responsibility grows as context grows. As experience grows. As trust grows.
AI should not be fundamentally different. As systems move from assisting people to participating in execution, we have to answer some fairly basic questions.
What is the AI allowed to access? What is it allowed to do? Under which circumstances? What evidence should be visible? When should a person review a decision? When should the system stop and ask? And afterward, can we understand exactly what happened?
This is why I believe trust has to precede autonomy.
The path is not simply from assistance to full automation. It is more nuanced: assist, act, greater autonomy where it makes sense, and ultimately innovation. Human involvement does not disappear along that path. It changes.
Sometimes the human is making the decision. Sometimes the human is approving it. Sometimes the human has already established the policy and boundaries within which the system is allowed to operate.
Accountability remains. That is why governance for AI on the mainframe cannot be something we bolt on after the interesting part is finished. It has to be part of how the system operates from the beginning.
For organizations adopting agentic AI, that means building oversight into the environment from the beginning, including visibility into agent performance, resource utilization, policy compliance, and auditable activity. It may not sound as exciting as the promise of unrestricted autonomy, but I have come to believe that this is what makes meaningful autonomy possible in the first place.
There Will Not Be One Agent
One other result from our mainframe survey stayed with me. Organizations are not choosing between creating their own agents and adopting agents from technology providers. They are planning to do both. Thirty-six percent told us they plan to create agents to help manage the mainframe, while 32% plan to use agents from others.
That may be a more consequential finding than it first appears. Because if that direction continues, the future enterprise will not have one AI agent. It will have many.
Agents built internally. Agents from software providers. Different models. Different tools. Different domains of expertise. Different levels of authority.
And probably a few things nobody has thought of yet. Once you accept that, the problem changes again. The challenge is no longer simply, “How do I build an agent?” It becomes, “How do all of these agents work together without creating a new kind of complexity?”
That is why I do not think agentifying the mainframe looks like one enormous general-purpose agent sitting above everything.
It looks more like a collection of specialized AI agents and capabilities across operations, data, application development, security, cloud, storage, and other domains, each grounded in the expertise relevant to that work and connected through a common foundation for context, multi-agent orchestration, identity, governance, and audit.
And it needs to remain open.
Organizations are already making significant investments in AI, from the models and platforms they have chosen to the specialized agents they are developing themselves. The answer cannot be to create another island. The mainframe has to participate in the enterprise AI ecosystem.
Organizations have already invested heavily in the technologies, workflows, and expertise that support their mainframe environments. I believe the next generation of AI should build on those investments, not require an entirely new approach. The opportunity is to make AI part of the architecture and operating practices already in place, rather than introducing another disconnected layer of technology.
What It Really Takes to Agentify the Mainframe
When I think about what it will take to Agentify the Mainframe, I find myself coming back to three things. AI needs to understand the mainframe environment in which it operates, draw on specialized expertise to participate meaningfully in the work, and operate within the controls and boundaries that organizations already depend on.
There is a technical side to that, of course. It takes more than connecting a general-purpose model to mainframe data. AI needs access to the right knowledge and live operational context. Specialized agents need the skills and approved capabilities required for particular kinds of work. And agentic workflows need orchestration when a problem crosses domains or requires more than one kind of expertise. And all of that has to operate within identity, permissions, policies, auditability, and human oversight.
The model is only one part of the system.
What matters is everything around it that allows intelligence to become useful in the environment where the work is actually happening.
BMC AMI Assistant provides the intelligent experience for understanding, investigating, and deciding. BMC AMI Agentics adds specialized digital workers for defined Operations and Data work. Underneath both are shared services for knowledge, approved models, agent discovery and orchestration, and governed access to the tools and data those agents need.
It means connecting what the organization already knows with what is happening now. It means taking expertise accumulated over decades and making more of it available to more people. It means specialized agents working on specific problems rather than assuming one general-purpose model will somehow understand every operational nuance.
And it means allowing the mainframe to participate in the larger enterprise AI strategy rather than becoming a separate AI destination.
Just as important, our approach to mainframe AI is being built into the BMC AMI Platform and the portfolio customers already use. The intention is to extend existing investments—in BMC technology as well as the models, tools, platforms, and enterprise AI systems customers have chosen—not require a separate AI stack.
At BMC, we think about this through four complementary strengths: mainframe-native expertise, integrated intelligence across the BMC AMI portfolio, enterprise openness, and governed execution.
None of those ideas is particularly interesting in isolation. That is important to say. Other companies will have expertise. Others will have agents. Others will talk about openness and governance. What matters is how those things come together.
What does the AI know? How does that intelligence connect? And how does it act when the consequences matter?
What I find compelling is the possibility of bringing these capabilities together in a way that allows AI to contribute to mission-critical work with the understanding, coordination, and accountability those environments demand.
The Real Goal Isn’t Autonomy. It’s Innovation
Whenever a new technology captures our attention, we sometimes become more focused on what the technology can do than on what we actually want to accomplish with it. We saw this with cloud and automation, and I think we’re beginning to see something similar with AI agents and the growing conversation around autonomy. But organizations do not invest in autonomy because autonomy itself has business value. They invest because they want to solve problems faster, increase capacity, improve resilience, develop applications more quickly, help people make better decisions, and create something new.
That last point matters.
Innovation is the destination.
Our conversations with industry analysts have reinforced this for me. Autonomy can help scale execution, but the real business outcome comes when people have more capacity to create new value. Human judgment remains central to that.
AI can identify patterns, assemble evidence, explain, recommend, and increasingly execute within defined boundaries. But deciding where an organization should go, which problems are worth solving, what risks are worth taking, and what new value should be created still requires human judgment.
I have come to believe that the greater opportunity with agentic AI is not to diminish the role of human expertise, but to make that expertise more valuable. When intelligent systems can take on more of the work, people have more room to apply their judgment, explore new possibilities, and focus on the kind of innovation that moves organizations forward.
Evolving Without Losing What Matters
If there is one thing the mainframe’s history has demonstrated, it is that evolution does not have to come at the expense of what makes a technology dependable. Through decades of change, the platform has continued to adapt while preserving the reliability, resilience, security, control, and expertise that organizations have come to expect.
I think the same principle should guide the way we introduce AI. We should absolutely take advantage of what these technologies make possible. We should help AI do more than answer questions, give specialized agents meaningful work, and connect the mainframe to the broader enterprise AI environment.
But we should not confuse moving quickly with abandoning the things that earned trust in the first place.
The agents are coming, and in many organizations they are already here. The more interesting question is what kind of intelligence we want participating in the systems the world depends on.
It should be intelligence that understands the environment in which it operates, learns from the expertise accumulated over decades, and has the operational context to recognize what is happening in the moment. Just as importantly, it must work within boundaries that people can understand and trust, earning greater responsibility as confidence in its capabilities grows.
That is what I mean by Agentifying the Mainframe.
