Where GenAI Actually Helps Mainframe Modernization

Generative AI is not equally effective across all aspects of mainframe modernization. Today, its strongest applications are code understanding, documentation and IT operations such as incident triage, anomaly detection and capacity forecasting. While AI can assist with code conversion, insurance carriers should approach automated transformation cautiously, applying rigorous validation and governance before moving converted code into production.

Generative AI’s value on the mainframe isn’t evenly distributed. Understanding where it’s already earning its keep — and where it still needs heavy guardrails — helps insurance carriers avoid both the hype and the overcorrection.

The April 2026 Gartner® research titled “Too Big to Fail: Why Mainframe Exit Projects Are Likely to Fail in the Age of Generative AI” identifies three separate domains where generative AI affects mainframe work: code understanding, code conversion/completion, and IT system operations. The maturity and effectiveness of GenAI tooling vary across these three areas.

That variance is the whole story. Treating all three as equally mature is how carriers end up disappointed with AI, or worse, talked out of using it at all.

Three Domains, Three Very Different Maturity Curves

Domain one: Code understanding — Advancing quickly, with evidence to show for it

This is where generative AI is genuinely strong today, and it maps directly into the legacy versus heritage conversation. Large language models are demonstrably good at summarizing what a COBOL program does, extracting business rules buried in decades of change requests and generating readable documentation for systems that were never properly documented in the first place.

This isn’t theoretical. In one recent engagement, Ensono helped UK insurance carrier Markerstudy apply AI to accelerate analysis and modernization across 6 million lines of legacy code, surfacing an equivalent of over an $11 million annual savings opportunity in a fraction of the time a manual code review would have taken. For a heritage claims or policy system, this is often the single highest-value application of AI available right now: not converting the system, but finally understanding it well enough to know what must be preserved.

Domain two: Code conversion/completion — Still requires heavy validation and governance

This is the domain the market hype is mostly about, and it’s also the domain to be the most cautious about. Generative AI can accelerate a first-pass code conversion, but it can’t certify that it’s captured every edge case, undocumented workaround or regulatory rule embedded in fifty years of maintenance.

Current tooling has, according to Garter, “significant limitations when it comes to automated conversion and migration of legacy code. It also does not account for the unique capabilities that the mainframe offers.” This includes whether the same throughput and performance can be achieved once a workload has moved. I agree and would go further for a regulated industry like insurance by asking if AI-assisted conversion output should be treated as a draft requiring the same underwriting rigor the business applies to any other high-consequence decision, not as a finished product.

Domain three: IT system operations — The most practical near-term win, and where the market is moving fastest

This is the domain we think is currently underrated in the mainframe and AI conversation, and it’s also the one we hear the most genuine appetite for in client conversations right now. Insurance carriers aren’t just asking how AI can help them leave the mainframe. They’re asking how AI and machine learning can make the platform they already run more observable, more efficient and less dependent on a handful of specialists who understand its quirks.

The Gartner report points to this domain directly, citing the emergence of “specialized CICS and JES agents for an aging workforce” as one of the more concrete near-term capabilities. Our own operations experience backs this up directly: Ensono’s teams have used AI to resolve IT incidents up to 70% faster, and we’re seeing the same pattern play out across code analysis, documentation, incident investigation, capacity planning and operational automation more broadly. In practice, the use cases cluster into a few categories:

• Anomaly detection and log analysis: Machine learning models trained on historical system logs can flag unusual patterns in CICS regions, JES queues or DB2 performance well before they become incidents, rather than after an operator notices a problem.
• Incident triage and prioritization: Generative AI can correlate alerts across subsystems, summarize likely root causes in plain language and help a smaller or less specialized operations team prioritize what actually needs attention first — the direct driver behind the incident-resolution gains referenced above.
• Capacity and performance optimization: Machine learning-based forecasting of MIPS consumption and transaction volumes can inform specialty processor offload decisions and batch scheduling, often surfacing cost and performance opportunities a manual review would miss.
• Automated runbook execution, routine remediation steps: Restarting a subsystem, clearing a queue or reallocating a resource can be codified and executed by an agent under human supervision, reducing mean time to resolution.
• Hybrid observability: As mainframe workloads sit increasingly inside a hybrid stack of cloud services, APIs and distributed systems, AI-assisted correlation is becoming necessary to trace a failure that starts in a cloud service and surfaces as a mainframe problem, or the reverse.

None of this requires touching a single line of application code. That’s precisely why we think it’s the most practical entry point for AI on the mainframe: the risk profile is fundamentally different from code conversion, the time to value is shorter, and the operational data needed to train and validate these models already exists on the platform.

What This Means for a Claims or Policy System

Bringing this back to a specific heritage system — a claims adjudication engine or a policy rating system — the sequencing matters. AI is most valuable first as a way to improve understanding and decision-making: mapping what the system does, documenting the business rules it enforces and giving operations teams better visibility into how it behaves under load. Only once that understanding exists does it make sense to talk about where automated conversion, refactoring or migration might responsibly apply — and even then, only for the legacy portions of the estate that warrant it.

Attempting to jump straight to automated transformation of a claims or policy system, before that foundation of understanding and operational visibility exists, risks mistaking a tool for a strategy.

A Practical Way to Sequence This

Domain	Current Maturity	Where to Start
Code understanding & documentation	Advancing quickly, with proven client results	Use now to map heritage systems and capture undocumented business logic
IT operations & performance	Practical near-term win; fastest-moving domain	Use now for anomaly detection, incident triage, capacity forecasting, runbook automation and hybrid observability
Code conversion & completion	Requires heavy validation and governance	Use selectively, on legacy (not heritage) applications, with rigorous validation before production use

Where this leaves the AI conversation

The honest version of the GenAI story on the mainframe isn’t “not yet” and it isn’t “finally, a magic converter.” It’s that AI is already earning its keep in code understanding and, increasingly, in IT operations — while code conversion remains a capability to use carefully, on the right workloads, with governance that matches the stakes. That isn’t a hedge. It’s the difference between adopting AI where the evidence supports it and adopting it because the marketing promised it would replace the hardest work.

This is the third article in a multi-part series exploring how insurance technology leaders should think about their mainframe strategy in the age of generative AI. The first article argued that insurers are asking the wrong questions around the “mainframe exit” conversation. The second gave insurance executives a filter for telling legacy applications from heritage ones.

Frequently Asked Questions

Where does generative AI provide the most immediate value in mainframe modernization?

Code understanding and documentation, and increasingly IT operations — anomaly detection, incident triage, capacity forecasting, and automated runbook execution. Both require no changes to application code and can be validated against existing operational data.

Why is automated code conversion still risky for regulated industries like insurance?

Current tooling can accelerate a first-pass conversion but can’t certify it has captured every edge case, undocumented workaround or regulatory rule built up over decades. For claims, policy, and billing systems, an incomplete conversion is a compliance and policyholder-trust risk, not just a technical bug.

Can AI improve mainframe operations without touching application code?

Yes. Anomaly detection, incident correlation, capacity planning and hybrid observability all run on operational data — logs, metrics, transaction volumes — rather than application source code, which is why this domain carries a fundamentally lower risk profile than code conversion.

What’s the safest way to sequence AI adoption across a mainframe estate?

Start with code understanding to map heritage systems and IT operations to improve day-to-day visibility and efficiency. Reserve code conversion for legacy applications that have already been identified as genuine candidates for retirement or replacement, with rigorous validation before anything reaches production.

Gartner, Too Big to Fail: Why Mainframe Exit Projects Are Likely to Fail in the Age of Generative AI, By Dennis Smith, Alessandro Galimberti, Tobi Bet, 8 April 2026. GARTNER is a trademark of Gartner, Inc. and/or its affiliates.