## Overview This case study presents two parallel mainframe modernization initiatives at Western Union and Unum Insurance, both leveraging AWS Transform—an agentic AI service announced in 2024 and enhanced in May 2025—in combination with partner solutions from Accenture and Pega. The core innovation involves a "composable" agentic AI framework where multiple specialized AI agents work together through orchestration layers to automate the end-to-end transformation of decades-old COBOL mainframe systems into modern cloud-native applications. Western Union, a 170-year-old financial services company, needed to modernize its money order platform (serving 110+ million transactions annually in North America) to support an ambitious growth strategy: increasing consumer services revenue from 15% to 25% of total revenue by 2028, representing $1 billion in growth. The existing mainframe-based system had become a bottleneck, with dwindling engineering expertise, inflexible architecture, and poor documentation accumulated over 35+ years of layered development. Unum Insurance, a 175-year-old company providing employer-sponsored benefits to 3.7 million lives, faced similar challenges with their Colonial Life claims processing systems. Their claims examiners worked across 7 different windows with fragmented workflows, while claims managers spent 3 hours daily (7,000 hours annually across 9 managers) manually assigning work—all supported by undocumented mainframe code maintained by tribal knowledge. ## The Composable Agentic AI Architecture The technical foundation of both solutions centers on AWS Transform's composable agent architecture, which addresses a critical challenge in the modernization market: tool confusion and integration complexity. Rather than forcing customers to choose between competing vendor tools, the composable approach allows multiple specialized agents to work together through standardized protocols. The architecture consists of several key layers. At the base sits AWS's broader LLM and agentic AI portfolio, including Amazon Bedrock (providing access to models like Claude, Llama, and Nova), agent orchestration capabilities, and the AWS Transform application layer itself. The Transform service provides base agents for assessment, VMware, mainframe, and .NET modernization at no additional charge to AWS customers. Partner agents are registered with AWS Transform through a standardized process. Partners develop agents in their local environment using tools like Strand Agent (an open-source framework) and AWS Transform primitives. Once tested locally, agents are deployed to Amazon ECS (Elastic Container Service) and registered with AWS Agent Corps, making them available for orchestration with Transform's native capabilities. The orchestration layer supports two primary patterns. In the "supervisor" pattern, multiple agents work side-by-side on parallel tasks—for example, Accenture's FSI (Financial Services Industry) knowledge base agent can work simultaneously with Transform's business rules extraction agent. In the "linear" pattern, agents execute sequentially (transformation → testing → data migration). The orchestration is enabled by the Model Context Protocol (MCP), described as a "smart API" that reduces hardwired coding requirements by allowing agents on MCP servers to communicate flexibly. Security architecture includes VPC-enabled runtime for data isolation, fine-grained IAM policies controlling agent access, and comprehensive logging and observability for agent actions—critical considerations given that mainframe workloads often contain sensitive financial and compliance data. ## Western Union Implementation Western Union's pilot focused on their money order platform, targeting three key problem areas. First, they needed to break away from terminal-based interfaces that prevented scaling customer support (they couldn't train tier-one representatives on terminal windows). Second, they required support for three distinct personas: end customers needing self-service tools, retail agents needing to assist customers, and 15+ back-office teams (accounting, fraud prevention, presentment operations, refunds, customer support) requiring efficient workflows. Third, they needed to accelerate clearing and settlement speeds to match modern banking expectations for mobile deposit scenarios. The technical execution involved the Accenture composable solution working alongside AWS Transform. The project processed 2.5 million lines of code, with 53,000 lines of COBOL converted to Java in approximately 1.5 hours. Technical documentation covering 31 elements was generated in 19 hours using Accenture's specialized agents, while business requirements documentation was completed in 5 hours. Data migration included 21 VSAM files and 25 DB2 tables transferred without accuracy issues. The timeline represents a significant acceleration: infrastructure setup, code transformation, and testing initiation occurred within 1.5 months—roughly half the time traditional approaches would require. The project aimed to move from discovery through testing in 6 weeks, compared to the previous 3+ month timeline. From an LLMOps perspective, Western Union emphasized several production considerations. The chatbot interface for simplified engineering access maintains "current context and nomenclature," allowing the system to preserve 35 years of institutional terminology rather than forcing wholesale naming changes. This contextual preservation reduces onboarding friction for both new and existing resources. Will Holway, who runs consumer services operations, stressed that "tech for tech's sake is not a good thing"—the focus remained on delivering tangible business outcomes rather than technology adoption for its own sake. The solution's human-centric design focus represents a key LLMOps consideration: moving from terminal screens to modern UIs enables different organizational roles to interact with the system appropriately. Customer support representatives gain access to capabilities previously locked behind specialized mainframe knowledge, effectively democratizing system access while maintaining security and appropriate permissions. ## Unum Implementation Unum's implementation followed a similar architectural approach but integrated Pega's low-code/no-code platform alongside AWS Transform. Their transformation process began with a revealing challenge: simply locating the code. Jay Moody described encountering "Bob in his basement next to his mainframe that he's worked on for 30 years" who was understandably protective of code he'd built and maintained. This highlighted the importance of empathy in transformation projects—recognizing that modernization may threaten the relevance of longstanding expertise. The technical workflow started with uploading code to Amazon S3 buckets, connecting AWS Transform to analyze approximately 1.5 million lines of code, and extracting business rules from the mainframe systems. Transform's agents generated comprehensive business rules extract documents showing execution paths, personas, integrations, and data models—providing documentation for "80 years of code" that previously existed only in one person's knowledge. The Pega Blueprint component represents a distinctive aspect of Unum's approach. Blueprint is a free online tool that accepts various input formats including Transform's business rules extracts or even video recordings of users navigating green screens. Blueprint analyzes these inputs to extract workflows, organize them into stages and steps, and identify decision points through a visual drag-and-drop interface. Critically, Blueprint serves as both a technical tool and a collaboration platform. Business subject matter experts can sit with IT teams to "define, refine, and reimagine" processes without requiring deep technical knowledge. The resulting blueprints function as contracts between IT and business stakeholders, providing clear sign-off on what will be built and how. The completed Blueprint downloads import directly into Pega's cloud-native solution running on AWS infrastructure, inheriting AWS's 99.95% availability SLA, disaster recovery, encryption, and backup capabilities. Pega's "composable architecture" (sometimes called the "situational layer cake") implements rule layers where the appropriate rule fires based on context and timing. The Unum solution delivers several out-of-box AI capabilities including skill-based routing, AI-powered "fast pay" for straight-through processing, smart adjudication, and operational reporting. From an LLMOps perspective, these represent pre-trained, production-ready AI components integrated into the workflow automation platform rather than custom models requiring extensive training and fine-tuning. ## Results and Production Impact Western Union's pilot demonstrated significant speed improvements, though Holway acknowledged they still had "a little ways to go." The accelerated timeline—half the traditional duration—positions them to meet aggressive 2028 revenue targets. The conversion of 53,000 COBOL lines in 1.5 hours represents automation at scale that would be impractical with manual approaches. Unum's results were more dramatic, particularly from a cost and timeline perspective. Previous vendor quotes estimated 7 years and $25 million for similar work. Using AWS Transform and Pega Blueprint, they achieved a COBOL-to-cloud migration with a testable application in 3 months—representing a 28x timeline compression if compared to the 7-year estimate. The operational impact on Unum's claims management is quantifiable. The 7,000 annual hours previously spent on manual work assignment by claims managers is eliminated entirely. Claims examiners consolidated from 7 different windows to a single unified experience. Most importantly, end customers (claimants) experience dramatically reduced wait times as automation enables faster adjudication and feedback on missing information. ## LLMOps Considerations and Tradeoffs While the results are compelling, several LLMOps considerations and potential tradeoffs emerge from critical analysis of these implementations. **Agent Orchestration Complexity**: The composable approach addresses integration challenges but introduces new complexity in orchestration. Managing multiple agents from different vendors (AWS, Accenture, Pega) requires careful coordination of agent registration, MCP server communication, and workflow orchestration. The presentation describes two orchestration patterns (supervisor and linear), but production deployments likely encounter edge cases requiring custom orchestration logic. The observability and logging capabilities are essential here, but the case study doesn't detail how conflicts or failures in multi-agent workflows are detected and resolved. **Testing and Validation**: Both implementations emphasize speed—COBOL conversion in hours rather than weeks—but the case study provides limited detail on testing rigor. Western Union mentions "start testing" in their 1.5-month timeline, and Unum achieved a "testable application" in 3 months, but neither provides specifics on test coverage, regression testing approaches, or how they validate that complex business logic is preserved correctly across transformation. AWS Transform includes automated testing capabilities announced at the same event, but the case study doesn't elaborate on how these are applied. **Model Selection and Performance**: The architecture description mentions access to multiple models through Bedrock (Claude, Llama, Nova) but doesn't specify which models were used for different tasks or how model selection decisions were made. Different agents likely use different underlying models based on vendor preferences, but the case study doesn't address potential inconsistencies in output quality or formatting across agents, or how prompt engineering and model tuning were approached. **Cost Structure**: The presentation notes AWS Transform capabilities are available "without no charge to customer and partner," but this likely refers to the service access rather than the compute costs for running agents, storing code in S3, or operating the infrastructure. The comparison to Unum's previous $25 million quote suggests significant cost savings, but a detailed cost breakdown (AWS services, Pega licensing, Accenture consulting) isn't provided. **Knowledge Preservation vs. Re-engineering**: Western Union emphasized maintaining "current context and nomenclature" from 35 years of history, while Unum highlighted the opportunity to "reimagine" processes. This represents a fundamental tradeoff in modernization: preserving institutional knowledge and familiar terminology versus taking the opportunity to rethink outdated approaches. The case study doesn't deeply explore how teams decide which aspects to preserve versus redesign, or how AI agents handle these nuanced decisions. **Human-in-the-Loop Considerations**: Both implementations emphasize human-centric design and empathy (particularly Moody's repeated emphasis on people over technology), but the case study doesn't detail how human review and approval are integrated into the automated workflows. Given that these systems handle critical financial transactions (money orders, insurance claims), there are likely regulatory and risk management requirements for human oversight that aren't fully articulated. **Scalability Beyond Pilots**: Western Union's implementation is explicitly described as a "pilot" in Q4 2025, and while Unum's is presented as more complete, questions remain about scaling. How do these approaches handle mainframe systems with 10 million+ lines of code? How do they manage dependencies across multiple interconnected mainframe applications? The case study focuses on relatively isolated systems (money order processing, claims processing) rather than enterprise-wide transformation. **Vendor Lock-in**: The composable approach theoretically reduces lock-in by allowing multiple vendor agents to work together, but in practice, both implementations show deep integration with AWS services (S3, ECS, Agent Corps, Bedrock) and specific partner platforms (Accenture agents, Pega Blueprint). Migrating away from this stack after transformation would be non-trivial, though arguably less concerning than being locked into an aging mainframe. ## Partnership and Ecosystem Dynamics The case study reveals an interesting ecosystem strategy by AWS. Rather than building every modernization capability in-house, AWS created a platform enabling migration competency partners (Accenture, Pega, IBM, Infosys, Deloitte, TCS, and ISV partners like OpenLegacy) to integrate their specialized tools and industry knowledge. This approach allows AWS to "scale through partners" and expand their "technology portfolio" without building everything directly. The presentation explicitly addresses "market confusion" caused by overlapping tool proposals from different vendors, positioning the composable approach as resolving this through integration rather than competition. However, this requires significant coordination and standardization (via MCP and agent registration processes) that may not eliminate all integration challenges in practice. The mention of customers being able to "bring their own agent" suggests an open ecosystem beyond the named partners, though the practical requirements for developing, testing, and registering custom agents aren't detailed. ## Conclusion These case studies represent sophisticated production deployments of agentic AI for enterprise transformation rather than experimental pilots. The combination of AWS Transform's code analysis and transformation capabilities with Accenture's industry knowledge agents and Pega's low-code workflow platform demonstrates the potential of multi-agent orchestration for complex legacy modernization. The results—28x timeline compression, 7,000 hours of eliminated manual work, and conversion of millions of lines of undocumented COBOL in hours—are impressive, though readers should maintain healthy skepticism about vendor-presented success stories and recognize that pilot results don't always translate perfectly to full-scale production deployments. From an LLMOps perspective, the implementations demonstrate several mature practices: secure multi-agent orchestration, integration of specialized AI capabilities (code analysis, business rules extraction, workflow generation), emphasis on human-centric design and organizational change management, and production-ready deployment on scalable cloud infrastructure. However, questions remain about testing rigor, cost structures, and how these approaches handle the full complexity of enterprise-wide transformation beyond isolated system modernization.