Google lets Gemini read the mainframe. Dual Run, not the model, proves the replacement works

Legacy modernization is where generative AI looks most useful and most dangerous at once. A model that reads two million lines of COBOL and explains the business rules saves months. A model that also tells you the rewritten system behaves the same is grading its own homework. Google’s modernization launch on October 5 contains a clean answer to that problem, and it is not an AI feature.

What Google announced

In a post by Souvik Choudhury, Senior Director of Product Management, and Tom Nikl of the Cloud Modernization and Migration team, Google introduced Google Cloud Modernize, a portfolio that brings migration, VMware, mainframe and application modernization tools together. The parts that matter here:

• Modernization Hub, a new in-console experience where developers and architects analyze source code, map dependencies and plan modernization for Java, .NET and mainframe applications.
• App Modernization CLI (CodMod), which uses Gemini to analyze large repositories, understand legacy architectures, map hidden dependencies and generate modernization recommendations.
• Mainframe Assessment Tool, which parses mainframe code, extracts business rules and maps application and data dependencies.
• Dual Run, which replays live production transaction streams against both the mainframe and the new cloud application at the same time, to verify functional equivalence before cutover.

The post does not attach a GA or preview label to Modernization Hub or CodMod, so check the status of each component in the console and documentation before you plan around it. Other parts of the launch carry explicit labels (the Agentic Quick Estimator is GA; the EKS-to-GKE agentic migration is in Public Preview).

The customer quote is the telling one. Claudio Balbo, Head of IT Architecture at Intesa Sanpaolo, says the bank must reassure its leadership, its internal control units and its regulators before moving forward with mainframe modernization, and that Dual Run “is gradually providing the evidences to build such confidence to all three groups.”

The pattern: AI proposes, a deterministic check disposes

1. Gemini reads code, maps dependencies
2. Business rules extracted as hypotheses
3. New services built and reviewed
4. Dual Run replays live transactions on both
5. People triage every difference
6. Risk and control approve cutover

• Understand: Gemini reads code, maps dependencies · Business rules extracted as hypotheses
• Rebuild: New services built and reviewed
• Prove: Dual Run replays live transactions on both · People triage every difference · Risk and control approve cutover

AI workHuman decision

The design separates two jobs that are easy to blur. Understanding is probabilistic: Gemini’s reading of a legacy codebase is a set of hypotheses about what the system does, and some will be wrong. Proof is deterministic: the same production transaction goes into both systems, and the outputs either match or they do not. The model does not get a vote on the second job.

This is the same principle we saw in GPT-Synopsys, where verification tools have the last word, and in Snowflake’s accounts payable system. Modernization is simply the highest-stakes place to apply it.

What “equivalent” has to mean before you start

Dual Run gives you a comparison engine. It does not decide what counts as a match. That definition is your work, and it is where most of the effort goes:

• Field-level rules. Timestamps, generated IDs, rounding and ordering will differ legitimately. Decide which differences are expected and write them down before the first replay, so nobody tunes the rules to make the dashboard green.
• Coverage over time. A week of replay misses month-end close, year-end, interest capitalization and rare product types. Plan the run to cover the calendar of the business, not just its volume.
• Triage ownership. Every difference is either a bug in the new system, a bug in the old one that people relied on, or an accepted change. Each needs a named decision, and the third kind needs business sign-off.
• Evidence packaging. Intesa’s point is that three audiences need confidence. Keep a record of runs, differences, root causes and decisions that an internal auditor can follow without the engineering team in the room.

Custody does not pause during replay

Replaying live production transactions means copies of production data, often customer financial data, flowing into a new environment before it is officially in production. That environment needs production-grade access control, logging and retention from the first replay, not from cutover day. For U.S. banks, it also means the replay environment is part of the third-party and change risk story examiners will ask about.

What to do now

1. Separate understanding from proof in your modernization plan: AI for the first, deterministic comparison for the second.
2. Treat AI-extracted business rules as hypotheses and test each one against replay results.
3. Write equivalence rules before the first replay and change them only through review.
4. Plan replay coverage around the business calendar, including closes and rare products.
5. Secure the replay environment as production from day one.
6. Build the evidence pack for three audiences: leadership, internal control and the regulator.

The bottom line

Google’s modernization launch is full of AI, but its most important component is a comparison engine. That is the right architecture for any regulated system: let the model accelerate understanding and rewriting, and let real transactions, compared deterministically, decide whether the new system is ready.