# Anthropic's COBOL Tool Rattled IBM, But AI Migration Still Hides Silent Bugs > Source: > Published: 2026-08-03 05:14:57+00:00 *AI can make a forty-year-old COBOL migration look finished before the dangerous part has even started. The code may compile, the tests may pass, and the database can still be one missed copybook away from trouble.* Picture the failure Ashutosh Singhal, founder of Veriprajna, laid out in a Medium essay on June 2, 2026. An AI agent migrates a bank's wire transfer module from COBOL to Java. The code compiles cleanly. The unit tests pass. Then the first transaction in user-acceptance testing trips the database's consistency checks anyway. Singhal doesn't name the bank, and you shouldn't treat the story as reporting on one specific institution. Treat it as a warning about the failure mode. The culprit in his account is a variable called TRN-LIMIT, defined in a COPYBOOK thousands of lines away from the code the AI is translating. That copybook carries a REDEFINES clause, a COBOL construct that lets the same block of memory hold different data depending on conditions set somewhere else in the program. The model reads TRN-LIMIT as a simple number. It isn't. In the mainframe environment, it holds a packed decimal. In the Java version, the tool treats it like a standard integer, and bad binary data goes into the database column. That's not a small translation mistake. That's the work. Research published this year points in the same direction. AgentModernize, an arXiv preprint by Sheikh Nazib Ahmed and Marnim Galib, argues that line-by-line legacy translation can compile and run while still breaking edge-case handling, validation logic, or hidden constraints. SEDCoT, another 2026 arXiv paper by Phillip Entin, Wenchao Gu, Alexander Knapp and Chunyang Chen, starts from the same uncomfortable fact: general-purpose language models still struggle with COBOL because it is a low-resource language with its own logic patterns. ## IBM had a bad day That research is landing in the middle of a louder fight. On February 23, 2026, Anthropic said Claude Code could automate much of the analysis behind COBOL modernization: mapping dependencies, documenting workflows, identifying risks and giving teams enough structure to move in quarters instead of years. Investors heard the threat to IBM immediately. Reuters reported that IBM shares sank 13.2% that day, their biggest one-day drop since October 18, 2000. IBM pushed back fast. Rob Thomas, IBM's senior vice president for software and chief commercial officer, wrote in a February 23 blog post that translating COBOL and modernizing a platform are not the same job. His line was plain: translation captures almost none of the actual complexity. He pointed to the stack around IBM Z, including z/OS, CICS, IMS, Db2, RACF and MQ, and argued that decades of hardware-software integration cannot be replicated by moving code. He has a point. You don't have to love IBM's moat to see the hole in the easy version of the Anthropic pitch. A mainframe application is rarely just a file of COBOL waiting to become Java. It is batch schedules, database contracts, job control language, encodings, access rules and business habits that may have survived three generations of management without ever being written down in one place. The installed base is still enormous. A 2022 global survey commissioned by Micro Focus and conducted by Vanson Bourne found between 775 billion and 850 billion lines of COBOL in daily use, with 92% of respondents describing COBOL as strategic to their business. In the public sector, the IRS's Individual Master File, its core individual taxpayer account database, dates to the late 1960s. A 2025 Government Accountability Office report said the IRS planned, as of September 2024, to complete IMF modernization in fiscal 2028 after spending $2 billion on the program through September 2024. You can see why the sales pitch works. The people who understand these systems are retiring, documentation is thin, and the code still moves money, taxes and claims. If an AI tool can read faster than a human analyst, every bank and government agency has a reason to try it. Frankly, they would be foolish not to test it. But testing it is not the same as trusting it alone. ## The missing line still matters Researchers have a specific name for part of the risk. Transformer models can lose accuracy on information buried in the middle of a long prompt, a pattern widely called lost in the middle. A single COBOL program can pull in copybooks thousands of lines long, and the dependency that matters may not sit at the top or the bottom where the model is most likely to keep its grip. Left alone, a model doesn't know what it missed. It writes code that looks right. Phase Change Software, a rival vendor selling legacy analysis tools, put the concern sharply after Anthropic's announcement: reading COBOL isn't the same as understanding it. That is a vendor argument, but it is not an empty one. A language model can narrate what a program appears to do without proving the narration is correct, and in a modernization project, a confidently wrong answer is worse than an honest failure. Gartner made the market version of the same argument on June 18, 2026. The firm predicted that more than 70% of mainframe exit projects initiated in 2026 will fail to produce their intended benefits because organizations overestimate generative AI's capabilities. Analyst Alessandro Galimberti said the gap between GenAI's marketing promise and its real ability to migrate complex legacy code is widening. The old causes of failure haven't gone away: unclear ownership, undocumented logic, scope creep and the risk of moving critical systems too quickly. None of this has stopped the rush. AWS says Danske Bank migrated 850 applications to the cloud in 15 months and has used AWS Transform and related tools on mainframe modernization work, according to a re:Invent 2025 session featuring Danske Bank's Ashish Shekhar. Google Cloud announced in April 2025 that its Mainframe Assessment Tool, powered by Gemini models, was generally available, and Kyndryl expanded its Google Cloud partnership in March 2025 to push GenAI into mainframe modernization. The market is not choosing between AI and no AI. That choice is already gone. The real choice is whether you use the agent as a fast analyst inside a verified process, or pretend it can replace the map, the tests and the people who know why the old system behaves the way it does. TRN-LIMIT is the reminder sitting under the whole pitch. The missed line is still the line that breaks the system. **Also read:** [Retail Traders Are Building AI Trading Bots Once Reserved for Hedge Funds](https://startupfortune.com/retail-traders-are-building-ai-trading-bots-once-reserved-for-hedge-funds/) • [Alibaba Says Its New Qwen3.8-Max Model Trails Only Anthropic's Claude](https://startupfortune.com/alibaba-says-its-new-qwen38-max-model-trails-only-anthropics-claude/) • [Microsoft's Project Perception AI Now Handles 90% of Security Work](https://startupfortune.com/microsofts-project-perception-ai-now-handles-90-of-security-work/)