All the enterprise AI roadmaps I’ve reviewed this year are essentially identical. They cover vector databases, hybrid cloud movement, agentic workflows, and data governance for large language models. All of that matters. None of it mentions the systems still processing the majority of the world’s core financial transactions.
Calling the mainframe “legacy” misses the point. The word gets used dismissively, but the mainframe is still the system of record for payments, claims, and account balances at most large banks and insurers, and it’s been quietly absent from almost every AI infrastructure conversation happening right now.
That omission isn’t accidental. It’s a skills gap dressed up as a strategic choice.
Why the Gap Exists
AI teams build against the data they can see. Cloud-native platforms expose clean APIs, structured logs, and metrics pipelines that plug straight into a RAG stack or a fine-tuning job. Mainframe data doesn’t offer any of that by default. SMF records come out in a binary format shaped by decisions made decades ago. Turning raw SMF into something an LLM can reason about takes someone who understands both the platform and the model, and that combination is rare enough that most teams route around the problem instead of solving it.
The result is a strange kind of blindness. A bank can run world-class fraud detection on its cloud transaction data and have almost nothing running against the mainframe transactions that generated the account activity in the first place. The infrastructure carrying the most risk gets the least AI attention, simply because it’s harder to reach.
What This Actually Looks Like in Practice
I ran into this directly while making z/OS SMF data queryable in plain English, instead of relying on hand-built SAS jobs or a Splunk license most teams can’t justify for a mainframe use case. The technical part wasn’t the hard part. Retrieval-augmented generation is well understood at this point. The hard part was everything upstream of it: parsing EBCDIC-encoded, densely packed SMF records into something with enough semantic structure for an embedding model to make sense of, without losing the operational context that makes the data useful in the first place.
Once that pipeline existed, the payoff was immediate. An operations engineer could ask “what changed on this LPAR in the last hour” in plain English and get an answer grounded in actual SMF activity, not a canned dashboard built for a different question. That’s a small example, but it points at something bigger: the mainframe has decades of extraordinarily rich operational data sitting behind a format barrier, not a value barrier.
Where the Real Risk Sits
Skip the productivity angle for a second and look at security. Enterprises are racing to inventory their AI exposure and their cryptographic posture at the same time, especially with NIST’s post-quantum standards (FIPS 203, 204, and 205) now in play. Most of that inventory work happens against cloud and distributed systems, because that’s where the tooling points by default.
Nobody’s cryptographic inventory is complete if it stops at the network edge and never touches AT-TLS policies, RACF-protected key material, or the certificate stores living on the mainframe. The same blind spot that’s slowing AI adoption is also slowing PQC readiness, for the same underlying reason: the platform is unfamiliar, so it gets skipped rather than solved.
What Actually Needs to Happen
This isn’t a call for a mainframe-specific AI product category. It’s simpler than that. Enterprises need to stop treating “AI infrastructure” and “mainframe infrastructure” as separate conversations run by separate teams. The organizations that get this right will be the ones that put people in the room who understand both sides, not the ones that wait for a vendor to build a turnkey connector.
That’s a smaller ask than it sounds. Most of the hard problems here, EBCDIC translation, SMF record parsing, RACF-aware access control, have already been solved in other contexts. What’s missing isn’t the technology. It’s the decision to point it at this part of the estate instead of working around it.
The mainframe isn’t going anywhere in the next decade, no matter how many cloud migration slides say otherwise. Every enterprise AI strategy that doesn’t account for it is planning around a gap instead of closing one. The organizations that figure this out first won’t just have better observability. They’ll have a more complete picture of their own risk, which is worth more than another dashboard.
