Some of the world’s most important financial and government systems still run on COBOL code written decades ago. The language itself remains dependable; the growing problem is that fewer people understand the sprawling applications built with it.
Anthropic’s claim that Claude Code can accelerate COBOL modernisation exposed how economically significant that problem has become. IBM shares fell 13% in their worst one-day decline in more than 25 years, while Accenture and Cognizant also dropped. Investors quickly imagined AI compressing years of lucrative consulting work.
AI attacks the archaeology of old systems
Modernising a large COBOL estate traditionally begins with painstaking discovery. Teams map dependencies, document workflows and determine what could break if one component changes. Institutional knowledge is often incomplete because many original developers have retired.
Anthropic argues that Claude Code can automate much of this exploration and analysis, giving teams detailed maps and risk information in quarters rather than years. The opportunity is enormous: hundreds of billions of COBOL lines remain in daily production, and the language reportedly handles about 95% of US ATM transactions.
AI is well suited to reading large codebases consistently, tracing connections and producing documentation. Reducing that preliminary burden could make projects economically viable that organisations have postponed for years. It could also loosen the dependence on a shrinking pool of specialists.
But understanding code is only the beginning. A generated explanation can guide engineers; it cannot by itself establish that a rewritten financial system will preserve every business rule, operational guarantee and regulatory control.
Translation is not platform modernisation
IBM’s response drew the essential distinction. Converting COBOL into another language is different from replacing the platform around it. IBM’s mainframe proposition includes z/OS, transaction-processing architecture, encryption and decades of hardware-software optimisation. None of those capabilities disappears because AI can translate source code.
The market reaction also overlooked that IBM has pursued the same opportunity. Its watsonx Code Assistant for Z is designed to help customers understand and modernise COBOL applications. Royal Bank of Canada has used it to map dependencies and develop modernisation blueprints, while the National Organisation for Social Insurance reported cutting an eight-hour analysis task to roughly 30 minutes.
Nor is COBOL synonymous with IBM mainframes. IBM noted that around 40% runs on Windows, Linux and other distributed platforms. The story therefore reaches beyond one vendor’s hardware business into a much broader market for legacy application work.
Customers have long had alternatives to mainframes, yet many stay because migration risk and platform qualities matter. AI may change the cost of understanding a system without automatically changing the case for moving it.
Consulting value moves toward judgment
The sharper threat is to work priced around manual discovery. If AI maps dependencies and drafts documentation quickly, clients may buy fewer hours for those tasks. Consultancies will need to concentrate value in architecture, validation, process redesign and accountable execution.
That change resembles what is happening elsewhere in software. A new AI capability triggers immediate questions about which existing revenue stream it can compress. The initial market verdict can run ahead of operational reality, but it still identifies where margins are vulnerable.
For banks and public institutions, the development is less about choosing sides between Anthropic and IBM. It creates a chance to inspect systems that have been too expensive or risky to touch. The responsible path remains incremental: use AI to expose structure and accelerate analysis, then keep experienced people in charge of validation and migration.
COBOL has not suddenly become obsolete, and mainframes have not lost their operational qualities. What has changed is the price and speed of understanding old code. That could finally unlock modernisation—while shifting the most valuable human work from code archaeology to deciding what should change and proving that it changed safely.
