Rocket Software expands EVA agentic AI on the mainframe and launches PlanGuard
Rocket Software expanded Rocket EVA, its agentic AI platform for mission-critical systems, and launched Rocket PlanGuard, a security layer that puts a policy checkpoint between AI reasoning and system execution on the mainframe.
SOFTWARE desk — the mainframe still runs the money, and the people who know it are retiring; Rocket is putting a policy gate in front of agentic AI so banks can let bots diagnose CICS, the online transaction software, and batch jobs without handing them the keys.
Rocket Software expanded Rocket EVA, its agentic AI platform for mission-critical systems, and launched Rocket PlanGuard. The record is the company’s GlobeNewswire release, “Rocket Software Advances Governed, Agentic AI on the Mainframe with Rocket EVA,” datelined Waltham, Massachusetts. This desk read that release as the FinancialContent syndication credited to Rocket Software via GlobeNewswire, stamped 08:29 a.m. Eastern on 23 Sep 2026, and as the matching SD Times Newswire reprint. 08:29 a.m. Eastern is 12:29 UTC. The wire says the expanded platform provides governed, auditable AI agents on the mainframe. Those agents correlate and reason across operational data and automate tasks inside policies the enterprise already defined. A mainframe is the large central computer that still runs core work at banks, insurers, and governments. Agentic means the software can plan steps and take actions, not only answer a question. Governed means those actions stay inside rules the company set. Auditable means a person can later see what the agent did.
The skills numbers, and whose they are. The release cites a Hanover Research study done for Rocket Software. It says 81 percent of financial services IT leaders report a very or extremely significant mainframe skills gap. It says 87 percent believe AI will help address that gap over the next two years. Later in the same release, Hanover Research is cited again: 94 percent of financial services IT leaders rank enhancing IT operations with AI as a high or top priority. These are the study’s figures as Rocket printed them. This desk did not rerun the survey.
The chief executive’s sentence, as a quote, not as a result this desk measured. Milan Shetti, president and CEO of Rocket Software, said enterprises have relied on the mainframe to run their mission-critical workloads for decades. He said AI promises to unlock more value from those environments only if enterprises can deploy it securely and without disruption. He said Rocket is helping customers apply agentic AI to mission-critical systems with speed, confidence, and control, closing the skills gap and putting that expertise within reach of every enterprise team.
Who is in the pilots. The release says global organizations across financial services, government, insurance, retail, and telecommunications are participating in EVA pilots. The focused pilot program lets a customer validate high-value operational use cases on their own data and move from installation to actionable insights within days, not weeks. The release does not name the customers.
What PlanGuard is, in the release’s words. The expanded platform includes the launch of Rocket PlanGuard, a new security layer that puts a policy checkpoint between AI reasoning and system execution. The release says that extends auditable, highly scoped agent access to mainframe resources. PlanGuard adds a policy decision point and identity controls so AI-driven actions remain within approved boundaries while the enterprise keeps control. A policy decision point is the gate that can allow an action, block it, or hold it for a person before it runs. Identity controls decide which account is allowed to do the work. Highly scoped means the permission is narrow, not a standing master key to the machine.
What the release says teams are already using EVA for. It attributes the list to a recent Rocket survey of mainframe AI use, which is separate from the Hanover study. Operations and diagnostics covers job-failure patterns, operational dashboards and KPI generation, end-to-end workflow automation, end-of-month financial reporting, IBM CICS online banking application optimization, and IBM CICS queue and backlog analysis. A KPI is a key performance indicator, a number a team watches. CICS, the Customer Information Control System, is IBM software that runs online transactions, including the ones a bank customer starts from an app. A queue is work waiting its turn. Security covers vulnerability detection, automated compliance, intelligent patch management, and identification of emerging risks associated with post-quantum cryptography. Post-quantum cryptography, often shortened to PQC, is encryption meant to hold up if future quantum computers can break today’s codes. A patch is a fix applied to software that is already running.
The rest of that use-case list, still from the release. Batch processing assists teams monitoring batch performance against SLAs, identifying critical failures, and investigating root causes through AI-assisted analysis. Batch jobs are the large runs that usually happen overnight, such as posting a day’s transactions. An SLA, a service level agreement, is the deadline that job is supposed to meet. Application and configuration management covers high CPU utilization and IBM Db2 data sharing, buffer pool, and cache degradation. CPU is the processor. High CPU means a job is using too much of it. Db2 is IBM’s mainframe database. A buffer pool is memory the database uses so it does not have to reread the same data from disk. Data and output management is Rocket’s aim of unified enterprise data access, with more intelligent delivery and streamlined output. Output is how reports and files get to the people who need them. The release says these uses are meant to cut manual investigation and shorten the path from an operational issue to a root cause.
What the release says this does for the people who run the machine. It lists four aims. Reduce mean time to resolution, the average time from a problem appearing to it being fixed, by connecting data, context, and expertise across the mainframe and the distributed systems beside it. Improve visibility through a conversational view of logs, metrics, applications, and system data, so a person can ask in ordinary language. Preserve expertise, so teammates at every experience level can use what seasoned specialists know, and the shop relies less on a shrinking pool of those specialists. Governed mainframe access keeps AI-driven actions inside patent-pending policy boundaries, using contextual, just-in-time authorization, with activity logged for audit and human oversight where the company requires it. Just-in-time means the permission is created for that task and then taken back. Contextual means the decision looks at the situation, not only at a standing account.
How it reaches the machine’s own data. The release says the expansion extends Rocket’s model-agnostic architecture with a lightweight, standards-based approach to connecting z/OS data. Model-agnostic means the product is not locked to one AI company’s model. z/OS is the operating system on IBM mainframes. The release’s own questions-and-answers section says the platform can access core z/OS data sources such as SMF records, including the RMF performance data they carry, along with job output and message queues. SMF, System Management Facilities, is the mainframe’s own activity log. RMF, Resource Measurement Facility, is the performance data inside that log. Job output is what a batch job prints when it finishes. A message queue is a line of messages waiting to be processed. The release says the point is to speed AI adoption on core business systems while governance and control stay in place.
The about box, only what it prints. Rocket Software says it is trusted by over 13,500 customers and 850 partners, with more than 3,000 employees worldwide. It is a privately held U.S. company headquartered in the Boston area, and a portfolio company of Bain Capital Private Equity. The media contact on the release is Lacey Darrow, ldarrow@rocketsoftware.com.
What an independent report the same morning adds, and where it is not the wire. SiliconANGLE’s story is by Paul Gillin. The visible line is UPDATED 08:30 EDT / SEPTEMBER 23 2026. It says a forthcoming EVA 2.0 will add PlanGuard. The wire says Rocket launched PlanGuard as part of the expansion announced the same day. Keep “forthcoming EVA 2.0” as SiliconANGLE’s wording, not the wire’s verb. SiliconANGLE says PlanGuard examines the caller, the request, the session, the selected tool, environmental conditions, and organizational rules, then permits the action, denies it, or requires another approval. If it grants permission, it creates a temporary execution identity limited to the approved task and revokes that identity when the work is complete. An execution identity is the account the action actually runs as. SiliconANGLE says Rocket said PlanGuard works with established mainframe security managers including RACF, ACF2, and Top Secret, rather than replacing them. RACF, Resource Access Control Facility, is IBM’s usual lock on who may touch a mainframe resource. ACF2 and Top Secret are the other two security managers widely used on those machines. Those three names are not in the GlobeNewswire paragraphs this desk read. They are in SiliconANGLE’s account of what Rocket said.
The same SiliconANGLE story, still not the wire. It quotes Phil Buckellew, president of Rocket’s Infrastructure Modernization Business Unit, on investigations that start from a plain-language request and on a chain of evidence for operators, security teams, and auditors. It says EVA adds a tamper-evident, hash-chained audit trail so the agent’s reasoning can be traced to identifiable people. It describes two pilots in that account. An operations team at a large South American financial institution spent about three weeks on a production problem. After receiving the relevant SMF records and operational context, EVA identified a probable root cause and supporting evidence in less than a day. A major retailer knew a production CICS region had stopped after exhausting temporary storage but did not know why. EVA concluded the event was a localized, application-driven issue rather than general system contention. SiliconANGLE also says pricing is consumption-based, tied to expected users and usage volume, that customers can use their preferred large language model providers and keep those model costs, and that Rocket estimates a typical deployment can generate a 3.2-times annual return, including model spend, through fewer specialist escalations and faster resolution. That 3.2-times figure is Rocket’s analysis, as SiliconANGLE labels it. This desk did not audit it.
What the product page adds, and only that page. The page at rocketsoftware.com/en-us/products/eva describes Rocket EVA as an AI-powered, agentic platform for end-to-end operational diagnostics across core systems, with plain-language questions and mainframe-specific knowledge. The page this desk read does not print the 23 Sep PlanGuard launch, the Hanover 81, 87, or 94 percent figures, or Milan Shetti’s quote. It does show other percentages, with footnotes, and a quote from Michael Curry. Those lines belong to the product page. They are not the 23 Sep wire, and this filing does not use them as announcement claims.
Plain English for the rest of the card. EVA is Rocket’s agent platform for core systems. PlanGuard is the checkpoint between the model’s plan and the system that would carry it out. A policy checkpoint is that yes, no, or ask-a-human gate. Days, not weeks, is the pilot path the release describes. It is not a stopwatch this desk ran.
PRIMARY here: Rocket Software’s 23 Sep 2026 GlobeNewswire release, “Rocket Software Advances Governed, Agentic AI on the Mainframe with Rocket EVA” — Tier A PRIMARY, the company’s own record on the wire. The body this desk confirmed is the FinancialContent syndication of that GlobeNewswire release, stamped 08:29 a.m. Eastern, and the matching SD Times Newswire reprint. A direct fetch of the GlobeNewswire URL timed out during this filing. The product page is product context, not a second announcement. SiliconANGLE is an independent same-morning report, not a substitute primary, and not the source of the Hanover figures. NOT claimed: that this desk measured the 81, 87, or 94 percent figures; that PlanGuard is already in production at every pilot; that the 3.2-times return is an independent audit; that RACF, ACF2, and Top Secret appear in the GlobeNewswire paragraphs this desk read; that “forthcoming EVA 2.0” is the wording of the wire.