SoftwareMining •
COBOL to Java & C#
• Predictable & Repeatable
Generative AI can provide valuable assistance during a COBOL modernization project. It can help developers understand unfamiliar programs, generate documentation, identify dependencies, suggest tests, and explain complex sections of legacy code.
Recent reports about an OpenAI model evaluation incident, where autonomous AI agents compromised Hugging Face infrastructure during testing, underline a key concern for regulated modernization projects: banks and government agencies need code transformation processes that are controlled, repeatable and auditable. See the BBC report and the OpenAI statement.
However, explaining or rewriting individual programs is different from translating an entire production application. Enterprise modernization requires consistent handling of business logic, control flow, data definitions, files, databases, transaction-processing APIs, error paths, and interactions between programs.
This article examines where generative AI can add value, where caution is required, and how AI-assisted analysis can be combined with deterministic translation.
Generative AI is most useful where modernization requires interpretation, summarisation, or developer assistance. It can help with:
Generative AI is valuable for code explanation and developer assistance, but enterprise application modernization presents additional challenges. The generated output may vary between model versions or prompts, and an individual COBOL program often cannot be translated correctly without the surrounding application context, including copybooks, called programs, databases, files, screen definitions, JCL, and transaction-processing services.
Enterprise COBOL applications also rely on platform-specific technologies, such as CICS, IMS, DB2 embedded SQL, VSAM, JCL, Unisys DMS-II, and HPE NonStop services, whose behaviour must be reproduced consistently across the entire application.
Readable or compilable Java or C# code does not, by itself, demonstrate functional equivalence. Validation should confirm that file processing, database updates, numeric precision, transaction behaviour, return codes, exception handling, and program interactions remain consistent with the original COBOL application. AI-generated code therefore typically requires expert review and execution-based testing before production deployment.
Deterministic translation applies defined rules consistently across the complete application. Its principal advantages are:
Learn more about how deterministic translation compares with generative AI in Claude Code vs Deterministic COBOL Modernization .
Generative AI and deterministic translation address different parts of modernization. Deterministic rules can be used for executable application conversion, while AI supports documentation, code explanation, discovery, test preparation, and developer productivity.
This allows AI to be used where interpretation is valuable, while deterministic processing provides repeatability, consistency, and controlled regeneration across the complete application.
The best modernization approach depends on the size, complexity, and objectives of the project. AI-assisted development may suit smaller or isolated tasks, while large mission-critical applications generally require a controlled, repeatable process.
When evaluating an approach, consider:
A representative pilot should use real application programs, execute business scenarios, and compare the results with the original COBOL application.
SoftwareMining has evaluated Claude Code using COBOL applications of varying size and complexity, assessing generated code, application context, repeatability, and suitability for enterprise modernization.
See Claude Code vs Deterministic COBOL Modernization .
Successful modernization should be demonstrated through execution-based testing using representative business scenarios and comparison with the original COBOL application. See Testing and Functional Equivalence for Migrated COBOL Applications .
Enterprise COBOL modernization should be evaluated on more than the appearance of the generated Java or C# code. The key questions are whether the translated application is complete, repeatable, testable, and suitable for controlled deployment.
When comparing modernization technologies, consider the following:
Can the complete application be translated, including business logic, data definitions, file handling, databases, error paths, and platform-specific technologies? Compilable code alone does not demonstrate a complete translation.
Does the same COBOL source produce consistent results each time it is translated? Enterprise projects typically require multiple translation cycles as applications continue to evolve.
Can the translated application be compiled, executed, and compared with the original COBOL system using representative business scenarios? Successful modernization should be demonstrated through execution-based testing rather than code review alone.
Does the approach support the technologies used by the application, such as CICS, IMS, DB2 embedded SQL, VSAM, JCL, or other platform-specific services?
How much manual correction is required? Can the application be regenerated after future COBOL changes? Is the process suitable for controlled change management, audit, and large application portfolios?
Generative AI and deterministic translation are complementary rather than competing technologies. AI excels at analysis, documentation, and developer assistance, while deterministic translation provides the repeatability and consistency needed for enterprise application conversion. Using each where it is most effective allows organizations to combine AI-assisted productivity with a controlled and repeatable modernization process.
If you're evaluating approaches to COBOL modernization, the following resources may also be useful:
Download PDF: Using AI Safely in COBOL Modernization