Executive Summary
Legacy systems are hard to replace because they work. The modernization question is how to prove the new code works the same way.
Legacy modernization changes old applications so they are cheaper to run, easier to change and staffed by people who can be hired. AI has changed the work: agents analyze code, extract business rules, write documentation, convert languages and generate tests. The cloud migration guide covers moving workloads between infrastructures; this guide covers changing the code itself.
The approaches differ by how much code changes and how equivalence is proven. Hyperscaler agents run modernization as a service on their cloud. Mainframe-native assistants help teams understand and change COBOL where it runs. Automated conversion translates COBOL to Java. Verified rewrite services compare old and new outputs. Architecture tools refactor large Java and .NET systems. Rehosting runs old workloads in the cloud without rewriting.
Why Legacy Modernization Is a Separate Purchase
Legacy systems carry core business rules that nobody has fully documented, written in languages fewer people know. GAO says it identified 11 of the most critical federal legacy systems, and three agencies had documented modernization plans that included all key practices for success. GAO lists a publication date of Jul 17, 2025. The risk sits in the plan and the proof of equivalence as much as in the tools.
Traceability is the new standard to ask for. Amazon Web Services says every line of generated code traces back through requirements to the original source code it replaces, creating a complete audit trail. Amazon Web Services says customers can modernize Windows, mainframe and VMware applications up to 5x faster with agentic AI-powered automation of analysis, planning, documentation and transformation tasks.
Keep, Convert, Rewrite, or Rehost
There are four broad paths. Keep the code and use AI to understand, document and change it. Convert it automatically to a modern language. Rewrite it with AI and verify the results. Or rehost it in the cloud without changing the code. Programs can combine paths across applications.
Keeping the code is a real option. IBM says watsonx Code Assistant for Z accelerates mainframe application development and modernization with AI and automation. IBM says developers can generate new COBOL code with natural language prompts, using a fine-tuned LLM trained on IBM Z. Rehosting avoids code changes entirely: Kyndryl says Kyndryl Cloud Uplift, formerly Skytap, runs IBM Power workloads such as AIX, IBM i and x86 in Azure without rewriting or refactoring.
| Path | What you get | What you take on |
|---|---|---|
| Keep and assist | Documentation, explanation and faster changes | The platform and its costs stay |
| Automated conversion | Code translated to Java or another language | Readability and performance of output |
| Verified rewrite | New code checked against old outputs | Building the comparison harness |
| Architecture modernization | Monoliths split into services | Long programs on large codebases |
| Rehost | Old workloads running in the cloud | Technical debt carried forward |
How do you evaluate AI Legacy Code Modernization?
Evaluate these products on a real application, not a sample program. The capabilities that decide results are discovery and dependency analysis, business rule extraction, documentation, code transformation, test generation, equivalence verification, and data migration.
Equivalence is where approaches differ. CloudFrame says it combines deterministic models with generative AI to build knowledge graphs and deliver functionally equivalent Java. CloudFrame describes transforming COBOL assets to maintainable, performant and cloud-ready Java. Mechanical Orchard says Imogen rewrites legacy systems into modern, cloud-ready code. Mechanical Orchard describes a software dark factory as one that writes and verifies software without a human in the loop.
| Capability | What it does | Buyer translation |
|---|---|---|
| Discovery and dependencies | Maps programs, data and jobs | The basis of every plan. |
| Business rule extraction | Turns code into readable rules | Have owners confirm them. |
| Documentation | Explains code in plain language | Useful even if you never convert. |
| Code transformation | Converts or rewrites code | Read the output yourself. |
| Test generation | Builds tests from behavior | Needed to prove equivalence. |
| Equivalence verification | Compares old and new results | Ask for evidence on your data. |
| Data migration | Moves files and databases | Plan it from the start. |
| Traceability | Links new code to old | Supports audits and debugging. |
Vendors in AI Legacy Code Modernization
Representative vendors include Amazon Web Services, CloudFrame, IBM, Kyndryl, Mechanical Orchard and vFunction. The list is alphabetical, is not a ranking, and is not for sale. How we choose vendors
The approaches below differ by how much code changes and who does the work. Programs can use several, so each approach describes where a product is strongest.
Conversion specialists have long track records. mLogica says that since 2004 it has delivered distributed and mainframe database and application modernizations in more than 600 large scale engagements. TSRI says it converts code from legacy to modern languages, UI, databases and architectures. Google Cloud describes assessment and reverse engineering of legacy mainframe applications using AI at scale, and AI code transformation and application reimagination with agentic workflows. Ownership is changing too: Astadia says it is now part of Amdocs.
| Vendor | Approach | Where it fits |
|---|---|---|
| vFunction | Architecture modernization | Context for modernizing Java and .NET applications |
| CloudFrame | Automated conversion | COBOL to functionally equivalent Java |
| Amazon Web Services | Hyperscaler modernization agents | AWS Transform for mainframe and other applications |
| IBM | Mainframe-native assistants | watsonx Code Assistant for Z |
| Kyndryl | Rehost and services | Rehosting and modernization services |
| Mechanical Orchard | Verified AI rewrite | Imogen rewrites with verification |
The six approaches were set first, then a representative was selected for each; the list is not for sale. About twenty vendors and service firms sell legacy modernization. Any of them, named here or not, can present its own case in the Spotlight below.
How much should you budget for AI Legacy Code Modernization?
Modernization is priced by lines of code or programs, by project, by subscription to tools, or by cloud service usage, and most programs include services. Ask each vendor for the pricing unit, what is included in a fixed price, and who pays for testing and fixes after conversion.
The value shows up in lower mainframe and license costs, easier hiring, faster changes and reduced operational risk. Model it per application, and include the cost of running old and new systems in parallel.
| Meter | You are charged for | Grows with | Where it goes wrong |
|---|---|---|---|
| Lines of code or programs | Code converted or analyzed | Portfolio size | Dead code counted and converted |
| Fixed-price project | An agreed scope | Scope | Change requests after discovery |
| Tool subscription | Assistants and analysis tools | Developers | Paying before the plan is ready |
| Cloud service usage | Agent and compute time | Workload | Uncapped usage during trials |
| Services | Planning, testing and cutover | Complexity | Testing underestimated |
How long does implementation take for AI Legacy Code Modernization?
Modernization programs succeed application by application. Start with discovery, prove the approach on one application, then scale.
Map the portfolio, dependencies and data. Extract business rules and confirm them with owners.
Modernize one representative application. Build the test harness and compare outputs on production data.
Choose a path for each application. Sequence by risk and value.
Modernize in waves. Run old and new in parallel, then retire the old.
What should you ask vendors about AI Legacy Code Modernization?
Ask each vendor to answer on one of your real applications.
-
Are you keeping the mainframe for now?Yes Shortlist mainframe-native assistants for understanding and change.No Continue.
-
Must you prove the new system matches the old exactly?Yes Shortlist automated conversion and verified rewrite.No Continue.
-
Is the legacy code Java or .NET rather than COBOL?Yes Shortlist architecture modernization tools.No Hyperscaler agents or rehosting fit.