What happened
AWS’s Artificial Intelligence blog post, “Modernizing mainframe Assembler to Java using Amazon Bedrock and AWS Transform,” presents a methodology for moving legacy Assembler applications to Java on AWS. It frames the problem as a skills and continuity risk: Assembler programs can underpin recurring business operations, while the people who understand their low-level logic may be leaving. The post says rewriting manually can take years and risks losing undocumented rules. It does not report a named customer deployment or measured result; it offers a process teams can apply.
The workflow has five phases. First, Amazon Bedrock—used through Kiro IDE or the Bedrock console—helps extract business rules and data mappings from Assembler code into a Business Rules Extraction (BRE) document. Second, the BRE is turned into a functional specification describing the program, COBOL data structures, process flow and coverage of business rules. Third, that specification guides generation of intermediate COBOL. Fourth, the COBOL is validated against the original Assembler code. Finally, AWS Transform converts COBOL into Java Spring Boot applications. AWS says the intermediate representation gives subject-matter experts a more human-readable stage to review before the final transformation.
The post gives operational guidance for the first phases. It recommends choosing a foundation model with strong code reasoning and a large context window, setting temperature to zero for consistency, and adding documentation, data dictionaries and business context alongside source code. For processing multiple programs, it says teams can call the prompts through the Amazon Bedrock Converse API and save each output in Amazon S3 for use in the next phase. The post recommends least-privilege IAM access and an S3 bucket with encryption and Block Public Access enabled. It says Bedrock does not use prompts or model outputs to train AWS models or share them with third parties, and advises using sanitized test data rather than production data or personally identifiable information.
Human review is built into the method, not treated as a final sign-off. AWS says SMEs should validate AI-generated BRE documents before specification work begins, record discrepancies in a validation log, and update the BRE. It recommends supporting review with flowcharts or data-flow diagrams and checking rule accuracy, data mappings, edge cases and error conditions. The generated COBOL should then undergo functional equivalence testing before conversion to Java. For existing COBOL source, AWS says teams can skip the Assembler-specific stages and provide COBOL directly to AWS Transform.
What the archive adds
New Runtime’s earlier coverage on agentic AI for cloud migrations on Amazon Bedrock AgentCore described dividing migration work into specialized stages. This is a new, more specific modernization workflow: it applies staged orchestration to Assembler-to-Java conversion and makes business-rule review and equivalence testing explicit checkpoints.
Trend signal
This supports AI-native organizations move toward review and orchestration: AWS’s method allocates work between AI generation and human validation, rather than treating code production as the only objective. The post is a vendor methodology, not evidence that teams have achieved a particular productivity or quality gain, so it illustrates the operating pattern without establishing its outcomes.
Why now
AWS ties the modernization problem to the shrinking pool of people able to read legacy Assembler and the risk of losing undocumented business logic as experienced maintainers retire. Its proposed intermediate COBOL stage is intended to create a reviewable bridge before migration to Java.
Why it matters
For teams modernizing systems that encode decades of business rules, the main risk is not producing syntactically plausible code; it is changing behavior that no longer has clear documentation. AWS’s workflow makes the intermediate artifacts—rule extraction, specifications and COBOL—objects that experts can inspect, challenge and correct. That gives teams several opportunities to catch errors before they become embedded in a new service.
The method also shows where a model-assisted process needs governance: constrain access, avoid sensitive production data in prompts, retain outputs between steps and test functional equivalence. Those controls do not prove that a conversion is correct, but they make the process more auditable and help direct scarce domain expertise to explicit decisions. Teams should treat the prompts as scaffolding for analysis and testing, not as a substitute for understanding the system or validating its behavior.
What happens next
- AWS will publish more guidance for applying staged AI workflows to legacy modernization. (6 months, 55%). Watch for: additional AWS modernization guides with explicit validation stages.
Limits and uncertainty
AWS presents a methodology and example prompts, not independently verified conversion results. It provides no benchmark for time saved, conversion accuracy, defect rates or the cost of expert review in this post. The suggested model choice is general guidance rather than a named model recommendation, and the article’s approach depends on SMEs who can judge business rules and test data with expected outputs. Teams would need to assess their own codebase, dependencies, test coverage and operational constraints before treating the process as suitable for production migration.
Skill to build now
Practice turning a legacy program into a reviewable rule inventory before asking a model to rewrite it. For one non-production routine, use the article’s BRE structure to record each rule’s inputs, outputs, conditions and purpose; have a domain expert mark discrepancies, then map approved rules into a functional specification. Keep test cases with expected outputs so later code changes can be checked against observed behavior.
Reader takeaway
The valuable part of AWS’s proposal is the sequence of inspectable artifacts and review gates, not simply using a model to translate old code. Put business-rule validation before code generation and functional equivalence tests before the final language conversion; without those checks, an intermediate representation can just carry mistakes forward.