Ready to Bring AI Automation to Your Business?
Let's DiscussWhen a company develops a powerful technological solution, there is often a temptation to apply it to every problem. For Anthropic, that solution is Claude, and the latest challenge is modernizing COBOL programs.
In February 2026, Anthropic announced that Claude Code could significantly accelerate COBOL modernization, potentially reducing projects that traditionally take years to just a few quarters.
The announcement unsettled investors. IBM shares declined approximately 13% on February 23, their sharpest single-day percentage fall since October 2000.
But does faster code conversion necessarily translate into successful modernization?
Unfortunately, the difficulties associated with legacy COBOL systems extend well beyond programming languages. The greater challenge involves understanding decades of accumulated business logic, institutional practices, and undocumented decisions.
First introduced in 1959, COBOL continues to support critical applications across banking, insurance, government, and other industries.
Many of these systems have undergone decades of modifications, making their underlying architecture increasingly difficult to understand.
(a) Interconnected and Complex Code: Legacy programs frequently contain tightly coupled subprograms and dependencies. Altering one component can unexpectedly affect several others.
(b) Hidden Business Knowledge: Important business rules are embedded within lengthy programs, often reflecting subject matter expertise (SME) and institutional knowledge accumulated over decades.
(c) Insufficient Documentation: Many applications lack reliable technical documentation. Existing records may not reflect subsequent modifications.
(d) Shortage of Experienced Programmers: A large proportion of programmers familiar with these systems have retired, creating significant knowledge gaps.
(e) Multiple COBOL Implementations: Different dialects, compiler extensions, and customized implementations complicate standardized conversion. Some estimates cite approximately 105,000 variants, although this number is difficult to verify.
These challenges explain why replacing legacy applications is considerably more complicated than rewriting their instructions.
During modernization, engineers may spend approximately 60% to 70% of their effort analyzing existing applications and reconstructing their underlying business logic.
Consider a banking application containing exceptions introduced decades ago to accommodate specific regulatory requirements or unusual transactions. Such rules may exist exclusively within the code, with little explanation of their original purpose. An LLM might accurately translate these instructions into Java or Python. However, without understanding their business significance, it could inadvertently change how exceptional transactions are handled.
The converted application might function correctly under normal conditions while producing incorrect results in less common scenarios.
Some industry estimates suggest that failure rates for legacy modernization projects approach 80%, although reported rates vary according to project scope and the definition of failure.
The underlying concern remains important: critical institutional knowledge can disappear during technology transitions.
Contemporary LLMs are capable of analyzing code and identifying patterns. Their ability to independently recover undocumented business reasoning, however, remains uncertain.
Despite these difficulties, enterprises cannot depend indefinitely on aging applications.
Estimates suggest that more than 200 billion lines of COBOL remain operational worldwide, supporting activities across substantial portions of the approximately $30 trillion U.S. economy. Although these applications have demonstrated considerable reliability, they are often expensive to modify, difficult to integrate, and challenging to scale.
The declining availability of experienced COBOL programmers compounds these problems. Consequently, modernization or replacement will become increasingly important over the next decade.
The economics are equally concerning since at an estimated conversion expense of $25 per line, modernizing an application containing one million lines could cost approximately $25 million. For large enterprises maintaining multiple interconnected systems, the financial implications are substantial.
Clearly, relying exclusively on traditional software engineering may prove prohibitively expensive.
A combination of human expertise, LLMs, specialized AI techniques, and rigorous testing may offer a more viable solution.
Experienced engineers and subject matter specialists, supported by AI tools, should first identify hidden business rules, system dependencies, and undocumented exceptions.
Once business requirements are understood, LLMs could potentially convert approximately 60% of suitable code, depending on application complexity. This should be treated as a working hypothesis rather than an established benchmark.
Other algorithms, including Random Forests, can assist with classification, dependency analysis, and identifying conversion patterns. Deterministic transformation methods and specialized engineering tools would remain essential for translating the remaining components.
Engineers must rigorously compare legacy and modern application behavior through regression testing, transaction reconciliation, and exception testing.
Even minor discrepancies in financial or mission-critical applications can create substantial operational risks.
Claude could considerably improve productivity in analyzing and translating COBOL programs. However, faster code generation does not necessarily guarantee accurate preservation of business functionality.
Successful modernization will depend on combining AI capabilities with institutional expertise, appropriate engineering techniques, and comprehensive validation.
The fundamental challenge lies in ensuring that AI-driven code conversion retains the business logic, institutional knowledge, and operational reliability embedded in legacy COBOL systems over decades.
And that is a considerably more difficult problem than code conversion alone.