Rocket Software adds PlanGuard to EVA 2.0 for mainframe AI

Rocket Software's pitch is aimed at a specific tension in enterprise AI: agents that can investigate and act are useful, but mainframes run payroll, transactions and other systems that cannot tolerate unsupervised changes. EVA already handles multistep investigations after a person submits a natural-language request, selecting connected tools and data sources, collecting operational context and returning findings with supporting evidence. The current emphasis, the company says, is analysis rather than unsupervised system changes.

PlanGuard is meant to mediate the step from analysis to action. It operates as a policy decision point at invocation time, evaluating the caller, request, session, tool and environmental conditions, and it can permit, deny or require another approval. When it grants permission, it creates a temporary execution identity limited to the approved task and revokes it when the work is complete. The system records who initiated a request, what an agent proposed, which policy applied, whether approval was necessary, the execution identity and the resulting action, and EVA adds a tamper-evident, hash-chained audit trail.

Rocket is testing EVA across financial services, government, insurance, retail and telecommunications, and cites two pilots: a large South American financial institution that spent about three weeks investigating a production problem before EVA identified a probable root cause and supporting evidence in less than a day, and a major retailer whose production CICS region stopped after exhausting temporary storage resources, where EVA concluded the event was a localized, application-driven issue. The open question is whether customers will move beyond AI-assisted investigation to governed execution. PlanGuard supplies policy and identity controls, but organizations must still decide which actions agents may take, when people must approve them and how to validate conclusions before changes reach production.