When ‘Legacy’ Is an AI Advantage
AI needs quality data, and that plays right into the mainframe’s hands
By Andrew Wig
When ‘Legacy’ Is an AI Advantage
AI needs quality data, and that plays right into the mainframe’s hands
By Andrew Wig
No one wants that six-letter word in front of their name, whether they’re a “legacy act,” “legacy media” or a “legacy platform.”
That last one pokes at a tender spot for the mainframe community. “I’ve been very conditioned not to use the word ‘legacy,’” says Dave Dalton, director of product management for BMC’s AMI Data business.
The mainframe community knows the mainframe itself is not a “legacy platform,” per se. But what if the mainframe’s legacy—its decades of high-quality data and unmatched reputation for stability—gave it an advantage in the ongoing race to operationalize AI at enterprise scale?
An Old Insult
“People have been discussing legacy computers since the early ’90s when people like Stuart Alsop said the mainframe is dead,” says Paul Robichaux, CEO and co-founder of New Era Software. Yet the mainframe remains. There have been nine generations of the mainframe since Alsop, the tech journalist, uttered those words in 1991.
Over that time, the mainframe has been building up vast stores of high-quality data, the lifeblood of AI. “Back in the day, we used to say analytics is only as good as the data. Today, we say AI is only as good as the data,” says Pete McCaffrey, principal product management lead, IBM Z.
And the mainframe is a faithful steward of valuable data. “The data that lives on the mainframe is high-value, it’s trusted, it’s real-time, and so therefore it is a very logical place to exploit AI,” McCaffrey says.
It is the system of record, after all. “The mainframe has huge amounts of curated information going back years and years and years,” Robichaux notes. That makes the retrieval knowledge base for a retrieval-augmented generation (RAG) implementation all the richer, he says.
Practical AI for Real z/OS Modernization
Stand Alone Environment (SAE) ?
Image Focus (IFO) ?
The Control Editor/AI (TCE/AI) ?
AI Family ?
Mainframe Data: Structurally Sound
It also helps that mainframe data tends to be structured, Whitbourne points out. That means it’s easier to access, whether it’s in DB2, IMS or even VSAM, he explains.
Compared to modern unstructured formats like MongoDB, these structured data formats lend themselves well to data integrity, notes Phil Young, director of mainframer penetration testing at NetSPI.
In unstructured databases, “you can just dump in whatever garbage you can give it, and there’s no tables, there’s no nothing,” Young explains. “And then you look at a SQL or DB2 database and it has rigid columns. The mainframe is even more rigid because even the files aren’t freeform on the mainframe. The files themselves have column limits. And so you can’t have bad data be introduced, say, into your ACH (automated clearing house) system because that’s been checked and triple checked and made sure it’s the exact right format for 30, 40 years.”
The Longer the Record, the Smarter the Model
And what about that legacy? “AI can leverage that history to discover hidden patterns,” McCaffrey says. More data means more edge cases being captured. For instance, in the case of fraud detection, a particular transaction may not trigger any flags on its own, but an AI model trained on years of mainframe data may recognize a pattern that would have otherwise gone undetected.
“There may be this one fraudulent transaction, but hey, we’ve seen that before,” McCaffrey explains. “We saw that 12 months ago, we saw that 18 months ago, and it tends to show up at this frequency in this way.”
Since anomalies don’t occur on a predictable schedule, the longer the training data is collected, the more likely it is to capture those nuances. Just because they aren’t happening currently doesn’t mean they won’t happen again. “The further you go back, the more refined the models become,” Robichaux says.
Data longevity can be especially important in cases such as mortgages, Whitbourne notes. That data has a natural lifespan of decades, giving AI models a fuller picture spanning varying economic conditions and interest rate environments. “That’s stuff that having access to archives from years ago is actually really, really important,” Whitbourne says.
60 Years of Trust
The mainframe has had plenty of time to develop a rock-solid reputation. Framing that reputation as a “legacy” may cause the platform’s future-looking evangelists to wince, but decades of stability as the world’s system of record have provided the mainframe with a critical asset for operationalizing AI—trust.
Trust is key, because if humans can’t trust the output of AI, they can’t do anything with it. And the mainframe has spent over 60 years building that trust, one transaction at a time. The technologists who witness this every day first-hand have an almost-instinctive level of confidence in the data processed by the system, Whitbourne says.
That trust, he explains, is “absolutely essential” to making AI work. “And I’d say it gets even more extreme,” he says, “especially the closer you go to the mission-criticality of the applications and the sensitivity of the data.”
Trust is also one reason the mainframe’s data isn’t going anywhere, Robichaux argues. The stakes are just too high for any rational executive to move years’ worth of vital records to a public cloud, he says.
“And if you attempted to do it, there would be regulators all over the place that would be sounding alarms. … What if it’s personal information? In some European countries, losing control of personally identifiable information is essentially life-threatening for a corporation,” Robichaux says.
Putting the AI Next to the Data
To help enterprises keep their data on-premises, IBM has spent the last two Z generations bolstering the system’s abilities to process AI workloads directly on the machine. This enables low-latency inference for workloads like real-time fraud detection.
Conversely, separating the mainframe’s data from its applications—like putting the application on the mainframe and the data in the cloud, or vice versa—introduces latency and new security boundaries while stretching transactional context across systems. This can cause a cascade of problems, Dalton says.
“Most people are keeping the transactional applications and data on the mainframe whilst building frontend experiences in the cloud," he says.
IBM’s On-Premises AI Hardware
Helping enterprises keep AI inferencing on-premises, IBM introduced the Telum processor as an AI workhorse with the z16, following that up in 2025 with the introduction of the Telum II and the new Spyre AI accelerator card for the z17 release.
The on-premises AI inferencing enabled by this hardware may help organizations overcome one of the key challenges in implementing AI at scale—governance. “Having a localized LLM running on your LPAR that you control is probably a far better solution than giving it to Anthropic,” Young says.
‘What’s the Problem?’
IBM’s embrace of on-premises AI is one more sign the mainframe is not the backwards platform that the “legacy” label implies. Meanwhile, a new generation of mainframers aren’t sure how their platform got that reputation in the first place.
As those who grew accustomed to the “legacy” narrative retire, younger mainframers are viewing the server as a forward-looking system. In BMC’s 2025 State of the Mainframe Report, 99% of Gen Z and 98% of millennial respondents viewed the mainframe as a long-term platform for new workloads.
Young Mainframers Don’t Buy the ‘Legacy’ Label
The mainframe’s power, its level of security and data integrity—“When they map that against a Linux server or a Dell server or a Windows server ... they generally come up thinking, ‘Well, the mainframe looks pretty good here; what’s the problem?’” Sigler says.
