The IT modernization debate has been loud recently. In late February, Anthropic claimed its Claude Code tool will help IT teams “modernize their COBOL codebase in quarters instead of years”, triggering a 13.2% drop in IBM’s share price. The announcement has since sparked widespread discussion and divided the tech industry. On one side, AI’s creators and evangelists argue that it will accelerate modernization. On the other, tech buyers at banks and other enterprises are asking questions about AI’s potential use cases, and more importantly, its limitations in modernization programs. Those are exactly the right questions. But the current debate is focussed on the wrong area. IT modernization is a nuanced practice, particularly within highly regulated sectors like banking and finance where the compliance implications of mistakes are severe. Language is just the tip of the iceberg AI-assisted COBOL modernization in banks is real and increasingly feasible. COBOL is a language. Although translation at the scale of banking applications is complex, it is still fundamentally a language problem that LLMs can handle effectively with the right governance architecture in place.But the legacy modernization market is not just COBOL. Across the globe, banks are running mission-critical applications built on low-code and no-code platforms from the 1980s and 1990s, such as Gen, CA Telon, SUPRA, MANTIS and NATURAL/ADABAS. These systems sit at the core of day-to-day operations. And while they are expensive to support and increasingly difficult to staff, they remain absolutely vital for banks. The overarching modernization challenge for these platforms is not language translation, but runtime substitution. The code these platforms generated depends on proprietary infrastructure that controls much of how the app behaves, including how data is handled, how transactions are managed and how calculations are performed. In many cases, these critical functions are embedded within the system rather than being fully visible in the code. Take Gen as an example. A banking application built on the platform is generated as code, such as COBOL, but how it functions is tightly bound to the Gen runtime. Important business logic, including how data processing and ensuring financial calculations meet regulatory requirements, sits within the platform rather than being explicitly defined in the code. When it comes to modernizing this app, what looks like a simple translation task is really rebuilding how the entire system operates. That process requires clearly defined, engineered rules rather than AI-driven interpretation. The same fundamental constraint applies to all modernizing all legacy applications, not just Gen. But on top of technical barriers, there are legal constraints to consider. The terms and conditions of Gen prohibit de-compilation or reverse engineering of code written by the platform. Any modernization project that processes this code directly, including adding it to a LLM, carries licencing risks that procurement and legal teams at banks are now actively scrutinizing. This wider IP issue is a material constraint that eliminates the possibility of AI-first transformation. Effective, governed legacy modernisation Modernizing legacy banking systems must be deterministic. The same inputs should produce reliable, consistent outputs on every run, and runtime substitution logic needs to be encoded as explicit, inspectable, versioned rules. However, an LLM’s responses can vary from one day to the next – or even hourly – as the model evolves, leading to inconsistent outputs. Determinism is important as auditors and regulators require financial institutions to demonstrate exactly what transformation was applied and why. A deterministic pipeline produces that evidence mechanically. Establishing the same verification in an LLM’s probabilistic pipeline requires a full-blown investigation. In complex banking apps comprising millions of lines of code, the scale and the level of scrutiny required completely negate any efficiencies AI offers during the build process. AI’s potential role in modernization However, AI does have a role to play in legacy modernization, but the current focus on COBOL is an oversimplification. In fact, many use cases for banks differ significantly from the code rewriting argument currently being discussed. One genuine and exciting opportunity is aiding discovery and documentation. Here, AI can help banks to understand what an IT estate contains, how components relate and the embedded business logic. Estates that have been running for 30 years and were maintained by developers who have since retired are often poorly documented. These estates are common in banks, but AI can accelerate the process of making that estate understandable. AI can also support banks in testing and assurance during the modernization process. Generating test cases from documented business logic, identifying coverage gaps and flagging behavioural divergence are all tasks where AI augments a deterministic pipeline, rather than replacing it. Banks can’t wait for AI to catch up Almost every bank across the globe will be running Gen, Telon, SUPRA and MANTIS estates. Licence pressure, talent scarcity and operational resilience requirements are forcing modernization decisions now. The question for banks is not whether to exit, but how to do so safely. Banks must recognize that AI cannot resolve their legacy challenges – at least not yet. For legacy modernization to be effective, financial institutions need specialist tooling built on deterministic pipelines. Runtime substitution logic must be grounded in deep platform knowledge and governance frameworks need to satisfy regulators and auditors. This approach already exists today, while the AI models needed to deliver this effective, compliant modernization do not. No#BankingTechnology #LegacyModernizationAndrew Mainhart head of legacy system modernizationTXP22 Jun, 2026