Can domain-specific AI accelerate mainframe learning while keeping human expertise and judgment central? IBM’s Meredith Stowell weighs in.
Artificial intelligence (AI) can help newcomers learn jobs quicker and contribute faster. A study by the National Bureau of Economic Research found that generative AI increased productivity 34% among novice and lower-skilled workers.
It’s unsurprising, then, that AI is now a tool for teaching mainframe skills. Meredith Stowell, Vice President of IBM Z & LinuxONE Ecosystem at IBM, calls AI a “mentor in your pocket.”
“You can ask them a question anytime you want,” said Stowell. “That to me is what is truly going to change the speed with which you can learn and the speed with which you can be productive on the platform.”
That immediacy matters on a platform where applications may contain decades of business logic and support the world’s banks, airlines, governments, and retailers. Yes, training can teach the system, and documentation can explain a process. But neither approach guarantees answers to sudden questions that pop up mid-task.
“You can ask [AI] a question whenever you want.”
AI can help. It can explain unfamiliar concepts, guide users through common activities, and reduce the time between learning something and applying it.
Stowell doesn’t see AI replacing experienced mainframers or the communities that sustain mainframe careers. Rather, she sees it augmenting them. “AI isn’t about replacement,” she said. “It’s not an ‘or.’ It’s an ‘and.’”
That distinction may determine whether AI benefits the mainframe skills gap — or simply produces answers faster.
From Training to Productivity
The mainframe skills conversation often focuses on how many people enter the field. With AI, Meredith suggests, companies can ask a different question: How quickly can new-to-Z employees contribute once they arrive?
A newcomer may understand IBM Z concepts without knowing how a particular employer has configured its systems. Mainframe standards, dependencies, exceptions, and institutional knowledge rarely fit neatly into a training module.
Because of that reality, newcomers usually pair with experienced colleagues. That remains essential, but the right expert may work in another office or time zone. Senior employees may not have time to answer every routine question. New employees may also hesitate to interrupt someone or admit what they don’t understand.
But an AI assistant can provide an immediate starting point. It can interpret terminology, walk through a process, or suggest the next step while the work stays in front of someone. AI agents can also handle recurring activities, giving newer professionals more time to understand why the system works the way it does.
The ‘Right’ Mainframe Knowledge
Calling AI a “mentor” invites an obvious challenge: What happens when it gives the wrong answer? Stowell’s response starts with a principle from her background in business intelligence and data science: “The output is only as good as the input.”
A general-purpose large language model (LLM) may produce an answer that sounds convincing but doesn’t reflect the realities of IBM Z. In a mission-critical environment, plausibility is not enough. The most useful AI must draw from relevant, trustworthy information and operate within a clearly defined domain.
“The output is only as good as the input.”
Stowell sees value in smaller, more specific models that can improve accuracy while using computing resources more efficiently. She pointed to watsonx Assistant for Z for operations and IBM Bob Premium Package for Z for application development, tools trained explicitly for IBM Z rather than every possible technology.
“You might be able to get the answer from other models,” she said. “But the accuracy may not be there based on how it was trained.”
Domain focus does not guarantee perfection. Organizations still need to ground AI in approved enterprise information, test its responses, refine it, and validate its recommendations. For instance, retrieval-augmented generation can connect a model with current documentation and organization-specific knowledge, but only if that source material remains accurate and well governed.
In short, AI fluency must include knowing when not to trust the first answer. Users need to examine the source, test the recommendation, and recognize when a question requires an experienced human.
AI fluency must include knowing when not to trust the answer.
Responsible AI Boundaries
The risks increase when AI moves from answering questions to taking action.
An AI agent might automate an operational step or help complete a development task. That capability can accelerate work, but it shouldn’t give an inexperienced employee authority they wouldn’t otherwise have.
“There are certain permissions and certain things that they’ll be able to do,” Stowell said. “And there are other things that they won’t, just like you would have limits whether you’re using AI or not.”
Security controls should limit what an agent can reach and do. Teams should verify consequential outputs rather than assume the technology got them right. Responsible AI adoption requires organizations to:
- Ground AI in relevant, approved information.
- Test outputs before using them in production work.
- Restrict agent permissions according to the user’s authority.
- Keep people accountable for consequential decisions.
- Escalate uncertain answers to an expert.
Don't miss these other great articles
Human Mentors Provide What AI Cannot
When Stowell meets people who are new to IBM Z, she asks how they found the platform and what keeps them there.
“Almost always it comes down to the community. It’s about mentors. It’s the people that they work with,” she said. Nothing replaces human connections.
AI can answer a technical question at 2 a.m., but it cannot fully explain why an organization made an unusual architectural decision 15 years ago. It cannot sense when a newcomer has lost confidence, or introduce them to a useful professional contact to help their career.
Experienced mainframers provide the necessary context and judgment. They identify exceptions, challenge AI-generated recommendations, and show how general guidance applies to a particular environment.
A newcomer can use AI to build an initial understanding, then bring a sharper question or proposed answer to a colleague. That makes mentorship more efficient without making it less important.
As Stowell put it, “You have your mentors there and you have the people there, and they can help you to validate.”
Lowering the Barrier Without Lowering the Standard
AI will prove most useful when it makes IBM Z easier to enter without reducing the standard for working on it. Stowell argues that any AI platform should meet developers at the level they work at today. Familiar tools and languages can help newcomers begin contributing while they acquire specialized IBM Z skills.
For example, COBOL remains the right choice for many business workloads, but IBM Z also supports Java, Python, Go, Rust, and other languages. AI can help developers understand unfamiliar code and move between languages, but the business purpose should still determine the language.
“It’s not COBOL or Python; COBOL or Java,” Stowell said. “It’s an ‘‘and.’’ You should use the right language for the right job.”
The assumption that developers cannot or will not learn COBOL also misses how programming education has changed. Students increasingly learn how languages work rather than preparing to use one language forever.
The Opportunity and its Limits
AI could shorten the path from IBM Z newcomer to productive contributor. It can make specialized knowledge easier to reach, explain unfamiliar systems, automate repetitive tasks, and support learning while in the flow of work.
But it won’t solve the skills gap. Success depends on the knowledge behind the model, the controls around the agent, and the people surrounding the learner. AI can explain syntax and code, but it can’t effectively decide whether a proposed change preserves the business logic inside a production application.
To succeed with trustworthy AI as a skills-development tool, organizations need:
- Domain-specific information
- Rigorous validation
- Limited permissions
- Human accountability
- A culture where questions lead to conversations
The strongest approach will combine AI assistance with human mentorship, specialized knowledge with familiar tools, and experienced judgment with new perspectives.
AI may give every mainframer a mentor in their pocket. But only people can ensure an AI mentor deserves – and maintains – trust.
Next
- Refresh yourself on the many charms of COBOL, with Uwe Graf
- Reflect on why human judgment matters in the AI Era, with Allan Zander
0 Comments