CICS User Group | December 2025
AI and CICS
In today’s fast-paced digital landscape, businesses demand seamless, efficient, and secure transaction processing to keep up with evolving customer expectations and market dynamics. On IBM Z, a platform renowned for its reliability and performance, CICS Transaction Server has been at the forefront of enabling robust application serving for decades.
Read the Transcription
[00:00:00] – Amanda Hendley, Planet Mainframe
We haven’t met yet—my name is Amanda Hendley. It’s great to see you, and thank you for joining us today. We’ve got a strong group and a really timely topic, so I’m excited for today’s session.
[00:00:21] – Amanda Hendley, Planet Mainframe
I’d like to thank our partners, Broadcom and DataKinetics, for sponsoring today’s program.
[00:00:33] – Amanda Hendley, Planet Mainframe
To get us started, please drop your location and job title into the chat so we can get to know each other a bit and see where everyone is joining from. Steve and I were talking earlier about how busy this time of year is, so we really appreciate you being here.
[00:01:01] – Amanda Hendley, Planet Mainframe
I also want to mention that we have an open call for sessions for our user groups—CICS, Db2, and IMS. We’re also always looking for articles for Planet Mainframe. January is our Workforce & Training Month, and if you have content you’d like us to consider, we’d love to see it. My email is on the screen: amanda@planetmainframe.com.
[00:01:33] – Amanda Hendley, Planet Mainframe
For Q&A today, feel free to drop questions into the chat as we go. Steve will keep an eye on them, and I’ll help moderate so no one misses anything. I’m already seeing some great locations—Atlanta, Paris—very cool.
Before we dive in, I want to remind everyone that today and tomorrow are the final days to complete the 2026 Arcati Mainframe User Survey. The results will be published in February with the Arcati Mainframe Navigator Report. If you complete the full survey, you’ll receive a $10 incentive and can submit an early nomination for the Influential Mainframers program.
[00:03:05] – Amanda Hendley, Planet Mainframe
If you’re interested in brand visibility, the Arcati Navigator also includes a directory of mainframe companies, which you can explore online. I’ll drop the relevant links in the chat shortly.
With that, we’re here for AI and CICS, so I’ll stop sharing and turn it over to Steve.
[00:04:16] – Steve Wallin, IBM
Thanks, Amanda—and hello everyone. Depending on where you’re joining from, good morning, good afternoon, or good evening. It’s great to be here, and please feel free to ask questions as we go.
What I want to focus on today is how AI is impacting CICS and the broader CICS ecosystem, and what we’re doing to continue driving innovation across the platform.
This is very much a full-stack story. When we think about IBM Z, it’s not about any single component—we’re integrating hardware, operating system, middleware, and application enablement to support some of the most critical workloads in the world.
[00:05:27] – Steve Wallin, IBM
With the IBM z16, we introduced the TELUM processor, with AI Accelerator Units (AIUs) embedded directly on the chip. That allowed us to perform real-time AI inference at scale, for the first time on a commercial, production-grade infrastructure.
With the z17, we’ve taken that further with TELUM II, continuing to push the limits of what’s possible on a vertically scaled platform that supports the transaction volumes our clients depend on.
[00:06:44] – Steve Wallin, IBM
We’ve expanded AIU capacity, increased throughput, and reduced power consumption—making it an extremely efficient architecture for high-volume workloads.
There are two core AI technologies in the z17:
TELUM II, which excels at real-time, in-transaction inference—use cases like fraud detection, AML, and anomaly detection.
SPire, which builds on this by extracting AIUs onto a PCIe-based accelerator card, enabling generative AI workloads to run on-platform while keeping data co-located and secure.
[00:08:03] – Steve Wallin, IBM
We also support Spyre Accelerator cards, which extend AI capacity beyond on-chip acceleration. Together, these components allow both predictive and generative AI workloads to run in the same environment as your transaction processing.
This builds on the strong momentum we’ve seen with the z platform. The z17 has been one of our most successful GA launches, driven by resilience, security, hybrid cloud integration, and now AI.
[00:09:17] – Steve Wallin, IBM
CICS is a core part of this story. Whether it’s predictive AI using AIUs or generative AI performing textual analysis, we can now support both directly on the platform.
These capabilities primarily support two scenarios:
Infusing AI into applications and data, either in real time or as part of broader workflows.
Operational excellence, spanning DevOps, diagnostics, automation, and system management.
[00:10:31] – Steve Wallin, IBM
From there, we see four main AI use cases:
In-transaction inference
Application workflow integration
Developer productivity (Code Assist for Z)
Operations automation (Assistant for Z)
[00:11:55] – Steve Wallin, IBM
This year we shipped CICS Transaction Server 6.3, and if you’re still on a version 5 release, I strongly recommend upgrading. It’s one of the smoothest upgrade experiences we’ve delivered, and it enables continuous delivery across the ecosystem.
Key enhancements include OpenTelemetry support, which provides end-to-end tracing across APIs, MQ, CTG, Db2, IMS, and back again—giving visibility into hybrid application flows.
[00:14:44] – Steve Wallin, IBM
We’ve also continued modernizing the developer experience with VS Code integration, configuration-as-code, and YAML-based definitions stored in Git.
From a security perspective, we’re seeing increased focus on crypto agility, quantum readiness, and zero trust principles, all of which are increasingly relevant to CICS environments.
[00:16:09] – Steve Wallin, IBM
We’ve enhanced support across Java, Node.js, and supported versions of Spring Boot, and expanded the CICS Java APIs as part of our broader language-level support.
Finally, we’ve made CICS ready to interact naturally with AI agents and LLMs by introducing an MCP (Model Context Protocol) server in CICS TS 6.3.
This is becoming increasingly standard across model providers, allowing models to interact directly with CICS for operational management of regions or an entire CICSPlex.
[00:17:21] – Steve Wallin, IBM
Before moving on, I want to take a quick detour from the slides and show you something practical.
One of the easiest ways to start interacting with AI capabilities in CICS today is actually through the CICS documentation. If you go to IBM.com and search something like “Explain CICS events,” you’ll still see the traditional links to relevant documentation—but at the top, we now also provide an AI-generated summary.
That summary pulls from multiple documentation sources and presents a concise, contextual explanation, along with links to the documents used to generate it. It’s a simple but powerful way to get to the information you need more quickly.
[00:18:39] – Steve Wallin, IBM
What’s especially interesting is how this can be tailored by persona. For example, if I ask “Explain CICS events to a cloud architect with no prior CICS knowledge,” the system restructures the explanation—removing CICS-specific jargon and reframing the concepts in more familiar cloud terminology.
This kind of adaptive explanation is extremely powerful when you’re trying to communicate how CICS works to audiences outside the traditional mainframe team.
[00:19:54] – Steve Wallin, IBM
With that, let’s return to the slides and move into application modernization.
One of the key modernization patterns we see is enhancing and extending existing applications rather than replacing them outright. A good example of this is how many organizations have historically embedded business rules directly in COBOL code—things like tax rates, eligibility thresholds, or decision logic.
[00:21:30] – Steve Wallin, IBM
Products like IBM Business Automation Decision Services were designed to extract those rules into a rules engine, allowing business users to modify behavior without changing, recompiling, and redeploying code.
Once rules are externalized, you get a natural progression: instead of manually defining rules, you can start generating decisions using data-driven inference models. This creates a clean path from hard-coded logic to predictive, AI-driven decision-making.
[00:22:51] – Steve Wallin, IBM
With Decision Runtime for z/OS, CICS applications can easily call into either traditional rule-based decisions or inference-based models. From there, organizations often move to Machine Learning for z/OS, where data scientists build models that are deployed directly on Z and invoked from CICS applications during transaction processing.
This makes in-transaction inference straightforward to implement. Your COBOL programs pass structured data to the inference service, receive a score or decision, and continue processing—all at transaction speed.
[00:23:58] – Steve Wallin, IBM
From a CICS perspective, this is intentionally simple. You define your input and output structures, populate containers, and use an EXEC CICS LINK to invoke the inference service. The model runs, and the response is returned to your application flow.
[00:25:36] – Steve Wallin, IBM
Another critical area is understanding and modernizing existing applications. Over the last few years, our tooling has improved significantly around code discovery, analysis, refactoring, testing, and validation.
We’ve focused heavily on fine-tuning models specifically for mainframe workloads, because generic models often struggle with COBOL, CICS, and z/OS constructs due to limited representation in their training data.
[00:27:02] – Steve Wallin, IBM
To illustrate this, consider tokenization. Languages like Java and Python often map cleanly to a single token in many models. By contrast, COBOL, CICS, and z/OS terms frequently break into multiple tokens, which means they require significantly more training data to achieve comparable accuracy.
We’ve invested heavily in pre-training and fine-tuning to address this bias, ensuring higher-quality results for mainframe-specific use cases.
[00:28:23] – Steve Wallin, IBM
This brings us to application workflows. With z/OS Connect, CICS applications can expose and consume RESTful APIs using OpenAPI standards. Copybooks can be mapped to APIs, and APIs can generate COBOL structures, enabling seamless integration across the enterprise.
More recently, we introduced the ability to expose those APIs as MCP interfaces, making them directly consumable by AI models and agents.
[00:29:47] – Steve Wallin, IBM
Once connected, models can discover available tools, retrieve data, and structure responses dynamically—without writing additional code. This opens up entirely new ways to interact with CICS-backed data and services.
For example, in an insurance scenario, a generative AI model can analyze claim text, retrieve policy data from CICS, perform textual analysis, and present a recommendation before a human agent even opens the case.
[00:30:56] – Steve Wallin, IBM
This ability to interact directly with VSAM-backed data and CICS services is already available today and enables a new generation of AI-driven workflows layered on top of existing business applications.
[00:31:33] – Amanda Hendley, Planet Mainframe
We’re good to keep going.
[00:31:35] – Steve Wallin, IBM
Great. Let’s move into operations and automation, starting with Assistant for Z.
In September, we released version 3, which marked a shift from simple chatbot-style interactions to fully agentic workflows. Assistant for Z now integrates MCP servers across CICS, IMS, Db2, SMPE, ServiceNow, and more—enabling automated diagnosis and action across the stack.
[00:32:50] – Steve Wallin, IBM
We’ve gone through three major phases:
Direct model calls, which rely entirely on pre-trained knowledge and are often unreliable.
RAG-based assistants, which ground responses in curated documentation.
Agentic workflows, where systems plan actions, gather live data, and generate responses based on real system state.
[00:34:05] – Steve Wallin, IBM
To demonstrate this, I showed how a generic model can produce a confident but incorrect response to a CICS error. Using grounded documentation improves accuracy—but still leaves work for the operator.
With an agentic workflow, however, the system can query live CICS regions, identify the root cause, and return an actionable resolution—dramatically reducing time to diagnosis.
[00:40:18] – Steve Wallin, IBM
This is where we see the most immediate client demand today: problem determination. Agentic workflows can gather evidence, summarize findings, open tickets, and eventually even propose corrective actions—while maintaining auditability and security controls.
[00:41:23] – Steve Wallin, IBM
Assistant for Z is designed as a platform. IBM provides core agents, ISVs can build their own, and clients can create agents tailored to their environments. These agents can run on z17 hardware using SPire or Spyre accelerators, or externally on x86 systems that call back into Z.
[00:42:43] – Steve Wallin, IBM
I often describe AI today not as intelligence, but as automation that was previously impossible. The real value comes from augmenting models with your data—using them as tools, not sources of truth.
[00:43:58] – Steve Wallin, IBM
Across the stack, this includes:
In-transaction inference for fraud and risk
Developer productivity through Code Assist for Z
Operations automation via Assistant for Z
Business workflow acceleration through MCP-enabled APIs
[00:46:24] – Steve Wallin, IBM
If there’s one takeaway, it’s this: the simplest way to start is often the documentation. From there, organizations can progressively adopt AI across applications, development, and operations—at their own pace.
[00:47:00] – Amanda Hendley, Planet Mainframe
There’s a question in the chat.
[00:47:02] – Steve Wallin, IBM
Sure. The question is whether any of the CICS AI capabilities require components to be running in an OpenShift environment.
The short answer is yes—for Assistant for Z specifically. The agent framework and supporting services run on OpenShift, whether that’s on x86 or on Z. OpenShift is currently a dependency for those components.
If that presents challenges, I’d really like to hear what they are. We’re actively looking at ways to optimize and simplify that stack as we move forward.
[00:48:26] – Amanda Hendley, Planet Mainframe
Thanks. Any other questions?
Steve, there was an earlier question asking whether your PowerPoint can be shared as a PDF.
[00:48:32] – Steve Wallin, IBM
Yes, absolutely. I’ll get that to you, Amanda.
[00:48:39] – Steve Wallin, IBM
Let me turn the question back to the group. Is anyone currently on an AI journey within your organization, or thinking about where to start? What questions do you have from that perspective?
[00:49:27] – Amanda Hendley, Planet Mainframe
We can absolutely do a follow-up session if there’s interest—particularly around configuration as code.
[00:49:33] – Steve Wallin, IBM
That would be great. We can certainly have someone from the team walk through how we’re approaching configuration simplification in CICS 6.3 and what’s coming next.
[00:49:33] – Steve Wallin, IBM
Another question: can you upload your own internal documentation into the models?
Yes—absolutely. With Assistant for Z, there are two key pieces:
An agent builder, which allows you to create custom agents and workflows
The ability to ingest your own documentation, processes, standards, and naming conventions to enrich the knowledge base used by the models
You can both extend the documentation available to the system and build agents tailored to your own automation needs.
[00:51:24] – Steve Wallin, IBM
Someone asked whether any of these AI capabilities are included directly with z/OS.
In general, no. Most of the AI capabilities we’ve discussed—Machine Learning for z/OS, Assistant for Z, and related tooling—are separate components that integrate with z/OS and CICS. We’re working closely across the stack, but these capabilities are delivered as complementary products.
[00:51:24] – Steve Wallin, IBM
There was also a question around co-location challenges. If you’re finding co-location difficult, I’d be interested in understanding where the friction is—whether it’s uncertainty around timelines, complexity of transformation, or lack of automation.
In cases where transformation is involved, discovery and understanding tools can be especially helpful. If it’s more of a lift-and-shift scenario, the challenges can be different. Either way, I’m happy to continue that conversation offline.
[00:52:58] – Amanda Hendley, Planet Mainframe
I’m seeing some chatter, but no additional questions coming in at the moment.
[00:53:05] – Steve Wallin, IBM
All right—happy to connect offline if anyone wants to follow up or dive deeper on any of these topics.
[00:53:13] – Amanda Hendley, Planet Mainframe
There’s one more question before we wrap up.
[00:53:17] – Steve Wallin, IBM
The question is where IBM watsonx Orchestrate fits into the stack.
Assistant for Z is actually powered by watsonx Orchestrate. It’s a shared technology foundation across IBM, but Assistant for Z is the product name you’ll interact with in the Z context.
[00:53:50] – Amanda Hendley, Planet Mainframe
Great—thank you.
To wrap us up, I want to thank our partners Broadcom and DataKinetics for supporting today’s session. If you’re interested in Db2 or IMS, we host user groups for those as well, and they meet on a similar cadence.
These sessions are recorded and published on our website with the slides and transcript, and they’re also available on YouTube if you’d like to explore past sessions.
Our next CICS session is scheduled for January 13, focused on the IT Stairway to Heaven and the CICS Testing Pyramid. Registration should be opening soon.
Thank you again, Steve, for a great presentation—and thanks to everyone who joined us today.
[00:55:16] – Steve Wallin, IBM
Happy holidays.
[00:55:18] – Amanda Hendley, Planet Mainframe
Happy holidays. Thanks everyone—bye-bye.
Virtual CICS Sponsor
Steve Wallin
Executive Director
IBM
Upcoming CICS User Group
November 13, 2025