Shares of IBM closed the day lower by nearly 13.2%, at $223.35 per share, after Anthropic on Monday said Claude Code could be used to automate the exploration and analysis work that drives most of the complexity in COBOL modernization, a key IBM business. IBM has long sold mainframe systems that are optimized for large-scale transaction processing, where COBOL has often been used.
Short for Common Business-Oriented Language, COBOL is a dominant code system developed in the late 1950s often used in business data processing, such as payment processing and retail transaction systems. An estimated 95% of ATM transactions in the U.S. use COBOL, according to Anthropic, making it a prime target for cost-efficient AI disruption.
Pia Singh
Hundreds of billions of lines of COBOL run in production every day, powering critical systems in finance, airlines, and government. Despite that, the number of people who understand it shrinks every year, Anthropic wrote in a Monday blog post.AI excels at streamlining the tasks that once made COBOL modernization cost-prohibitive.
Anthropic and its Claude chatbot have been making quite a lot of headlines lately — and I haven’t even gotten to the most consequential yet. In an ideal world, what Claude promises here would be a genuine improvement, enabling fast modernization and optimization of antiquated systems.
But alas we live in the real world, where, like it or not, we must address negative consequences and explore trade-offs and various risk scenarios. I’m not convinced that the current stage of LLM agents can even reliably achieve what the blog post envisions. Did Claude’s training data included enough codebases written in COBOL to support their claims? Was it trained in the specific business logic it would need to map and redraw complex workflows? And if not, one might suspect that this blog post functions more as a marketing ploy, aiming to convince at least a couple of these legacy businesses to open up their confidential systems to Claude to train upon — which Anthropic would then be able to use to attract more customers.
Sure, it’s cool to imagine letting an AI loose on the existing COBOL code base and waiting for it to deliver fresh code that supposedly replicates everything the original code did, while being faster and easier to maintain. But… how can you be sure that the new software actually runs the same workflows and delivers the same results as before? Do you… put the same AI in charge of regression testing and quality control!?
What if, in the quest for ‘streamlining’, the bot removes what it labels redundant sections of code that were designed to handle edge cases or complex regulatory requirements? I have been through enough system migrations to know that even if an updated system manages to solve legacy issues, it usually introduces fresh problems that may take weeks or months to discover and understand. Handing over critical system such as banking transactions and airline control to poorly understood bots that might suddenly start deleting everything in sight while ignoring stop commands is inviting a disaster of unparalleled proportions.
The market reaction thus seems rather irrational. If Claude, or another similar agent, were able to successfully replace numerous consultants and COBOL experts, then indeed consultancy companies such as IBM would lose large swaths of their clients and recurring business. But is this a realistic outcome on a short to medium timeframe? The blog post that triggered this selloff doesn’t even have concrete examples of Claude achieving this in a small, controlled test; this is nothing more than a rough concept at this point. As someone, I forgot who, said in a recent podcast, most investors don’t understand the technology around AI, what it’s realistically capable of and what its limitations are, leading to these wild swings on the market with little connection to fundamentals.
Post a Comment