The $800 Billion COBOL Problem: How Anthropic’s AI Play Could Upend IBM’s Legacy Modernization Empire

="">

For decades, the world’s largest banks, insurers, and government agencies have been held hostage by a programming language older than the moon landing. COBOL — the Common Business-Oriented Language first developed in 1959 — still processes an estimated 95% of ATM transactions, 80% of in-person transactions, and runs systems that touch $3 trillion in daily commerce. Now, Anthropic, the San Francisco-based AI company behind Claude, is making a bold bet that artificial intelligence can do what armies of consultants and billions of dollars in modernization budgets have repeatedly failed to accomplish: make COBOL systems not just tolerable, but genuinely manageable for the long haul.

The implications for the technology industry — and particularly for IBM, which has built a sprawling consulting and software business around COBOL modernization — are enormous. If AI tools can dramatically reduce the cost and complexity of maintaining, understanding, and even translating legacy COBOL code, the calculus that has driven multibillion-dollar modernization contracts for years could shift overnight.

A Language That Refuses to Die

The scale of COBOL’s entrenchment in global infrastructure is staggering and often underappreciated outside of enterprise IT circles. According to reporting by TechRadar, there are an estimated 800 billion lines of COBOL code still running in production environments worldwide. The U.S. federal government alone relies on COBOL for critical systems including Social Security payments, IRS tax processing, and Veterans Affairs benefits administration. Major financial institutions run their core banking platforms on COBOL, and many of the world’s largest insurance companies process claims through COBOL-based mainframe applications that were written decades ago.

The problem is not that COBOL doesn’t work. In many respects, it works extraordinarily well — these systems are fast, reliable, and battle-tested over half a century of continuous operation. The problem is that the workforce capable of maintaining these systems is aging out. The average COBOL programmer is well into their 60s, and universities stopped teaching the language years ago. When the COVID-19 pandemic hit in 2020 and unemployment claims surged, several U.S. states discovered — to their considerable embarrassment — that their COBOL-based systems couldn’t handle the load, and they couldn’t find enough programmers to fix them.

Anthropic’s Pitch: AI as COBOL’s New Best Friend

Anthropic’s approach, as detailed by TechRadar, is not to kill COBOL but to extend its life — potentially indefinitely. The company argues that large language models like Claude can read, interpret, and document COBOL code at a speed and scale that would be impossible for human programmers. Where modernizing a COBOL system once required armies of consultants spending years mapping workflows, interviewing retiring programmers, and painstakingly documenting business logic buried in millions of lines of code, AI can compress that process dramatically.

The AI can analyze COBOL codebases to identify what each module does, map the dependencies between programs, generate documentation that doesn’t currently exist, and even suggest or execute translations into modern languages like Java or Python. Perhaps most critically, AI can serve as an institutional memory replacement — capturing the deep domain knowledge embedded in COBOL code that would otherwise walk out the door when the last generation of COBOL programmers retires.

Why IBM Has the Most to Lose

IBM’s relationship with COBOL is deeply intertwined with its corporate identity and its bottom line. The company’s mainframe business — built on the z-series hardware that runs the lion’s share of the world’s COBOL workloads — generated $3.4 billion in revenue in 2023. But the real money isn’t just in hardware. IBM’s consulting arm, along with its software tools for application modernization, represents a massive revenue stream built on the premise that COBOL modernization is extraordinarily difficult, time-consuming, and expensive.

IBM has invested heavily in its own AI-powered modernization tools, including watsonx Code Assistant for Z, which is designed to help translate COBOL to Java. But IBM’s commercial incentive is complex: the company simultaneously wants to sell modernization services and maintain the relevance of its mainframe platform. If a third-party AI tool — particularly one from a well-funded competitor like Anthropic — can make COBOL maintenance cheap and accessible, it threatens both sides of IBM’s legacy business. Organizations might decide they don’t need to modernize at all if AI can make their existing COBOL systems comprehensible and maintainable at a fraction of the current cost. Alternatively, if they do modernize, they might not need IBM’s expensive consulting engagements to do it.

The Consultant Economy Built on Complexity

The COBOL modernization industry extends well beyond IBM. Accenture, Deloitte, Tata Consultancy Services, Infosys, and dozens of smaller specialized firms have built substantial practices around helping enterprises understand and migrate their legacy systems. These engagements are notoriously long and expensive. A large-scale COBOL modernization project at a major bank can run into the hundreds of millions of dollars and take five to ten years to complete — with no guarantee of success. High-profile failures, including abandoned modernization projects at major financial institutions, have become cautionary tales in the industry.

The consulting model depends on a fundamental asymmetry of knowledge. The client organization often doesn’t fully understand its own COBOL systems — the original developers are gone, documentation is sparse or nonexistent, and the business logic is encoded in ways that only become apparent through painstaking analysis. Consultants step into this knowledge vacuum and charge accordingly. If AI can fill that vacuum instead, the pricing power and value proposition of traditional modernization consulting could erode significantly.

Technical Realities and Limitations

Industry experts caution that AI-driven COBOL analysis is not a silver bullet. COBOL systems don’t exist in isolation — they are embedded in complex environments involving job control language (JCL), CICS transaction processing, IMS and DB2 databases, and intricate batch processing schedules that have been refined over decades. Understanding the code itself is only part of the challenge; understanding how it interacts with the broader infrastructure, how data flows between systems, and what happens when something goes wrong requires contextual knowledge that current AI models may struggle to fully replicate.

There are also significant concerns about accuracy. When AI translates COBOL to a modern language, even small errors can have catastrophic consequences in systems that process financial transactions or government benefits. A misinterpreted rounding rule or an incorrectly translated date format could result in billions of dollars in errors. Testing and validation of AI-generated translations will remain a labor-intensive and high-stakes process, regardless of how good the AI becomes at the initial translation.

The Strategic Calculus for Enterprise CIOs

For chief information officers at large enterprises, Anthropic’s entry into the COBOL space introduces a new variable into an already complicated decision matrix. The traditional options have been stark: spend enormous sums to rewrite systems in modern languages, or continue paying a dwindling pool of expensive COBOL specialists to keep the lights on. AI offers a potential third path — using intelligent tools to extend the viability of existing systems while gradually modernizing at a more manageable pace.

This “AI-assisted maintenance” model could be particularly attractive to organizations that have been burned by failed modernization attempts. Rather than ripping out systems that work and replacing them with new code that introduces new risks, they can use AI to document, optimize, and incrementally improve their existing COBOL infrastructure. The financial case is compelling: if AI can reduce the cost of COBOL maintenance by even 50%, the business case for a risky, multi-year modernization project becomes much harder to justify.

A New Front in the AI Enterprise Wars

Anthropic’s move into legacy code modernization also signals a broader strategic ambition. The company, which has raised more than $7 billion in funding from investors including Google and Salesforce, is clearly looking beyond chatbots and content generation to find enterprise use cases where AI can deliver measurable, high-value results. Legacy code modernization — a market that some analysts estimate at $30 billion annually — fits that profile perfectly.

The competitive dynamics are intensifying. Microsoft, through its partnership with OpenAI, has been positioning GitHub Copilot and Azure-based AI services for enterprise development workloads. Google has its own code generation and analysis tools built on Gemini. And IBM, despite its potentially conflicted position, is not standing still — the company has been aggressively marketing its watsonx platform as the enterprise AI solution of choice for mainframe environments. The question of who will own the AI-powered future of legacy code management is far from settled, but the stakes — measured in hundreds of billions of lines of code and trillions of dollars in daily transactions — could hardly be higher.

What is clear is that the decades-old status quo around COBOL — characterized by eye-watering consulting fees, perpetual modernization projects, and a steadily shrinking talent pool — is facing its most serious disruption yet. Whether that disruption comes from Anthropic, from IBM’s own AI tools, or from some combination of competitors remains to be seen. But for the first time in COBOL’s 66-year history, the technology industry may have found a way to address the language’s maintenance crisis without requiring a complete rewrite — and that alone represents a fundamental change in how enterprises think about their oldest and most critical systems.