A blizzard had just ended in the New York City metropolitan area when red alerts started flashing on Wall Street trading desks. IBM’s stock was in free-fall, ending the day down 13% – shedding billions of dollars from investors’ portfolios (including mine).
What triggered this rapid sell-off? A blog post from the AI company Anthropic (https://resources.anthropic.com/code-modernization-playbook) stated that Claude, their now-enhanced product, could rapidly transmute all existing COBOL code running on mainframes in banks, brokerage firms, insurance companies, and countless organizations in the Federal government into Java or Python code. Anthropic’s premise was that their AI could read all the COBOL programs, the copybooks (maps of file layouts), the JCL (instructions for running the programs, which files to use as input, and where to produce output), and divine the necessary intersections and cross-correlations to yield a newly formed, more functional and maintainable series of programs.
But before I discuss why that is significant, let’s take a momentary pause. Here’s where my “prior computing life” comes into the story. In 1997, an IBM consulting group asked me to work on a special project, one that would be incredibly difficult and time-consuming, challenging, and ultimately rewarding. It was, I’ll admit, an honor and incredibly tempting. To ensure the project’s deadline and goal were met, IBM created this awesome product that would read all the COBOL programs, the copybooks, and the JCL, then depict the locations and cross-correlations of very specific fields in the programs and files. The client was PaineWebber (now UBS), and the assignment I took was for the Year 2000 Project (Y2K).
IBM called their product the Asset Analyzer, which is really what it was. The product made it incredibly easy to identify the two-digit date fields. What took time – and skill – was updating the programs and files, and then thoroughly testing to ensure that everything worked as expected. Back then, testing a single program was straightforward. Testing an entire application was more work- (and stress-) intensive. Ensuring date fields were passed correctly across various applications required extensive effort, including highly scripted runbooks and dozens of programmers, systems administrators, and technical support staff.
I used the expertise I developed during that four-year engagement in several other expansion projects over the course of a decade.
In 2004, I went to San Jose, California, to write an IBM publication on the latest successor to the original product, WebSphere Studio Asset Analyzer. By then, it had grown from its initial foray of expanding fields into a first-generation utility for rewriting COBOL programs into Java. I was writing my second book for IBM and became even more familiar with the ins and outs of accomplishing COBOL migration.
Fast forward to 2024. One of my trade publication websites had a lead story that IBM had just introduced the watsonx Code Assistant for Z, an AI product that would (wanna guess?) rapidly consume your mainframe legacy code and artifacts and assist your programming staff in the creation of Java programs that would be more easily maintained by younger staff who understood the more recent coding patterns.
So now, two years later, Anthropic wants to challenge IBM on its home ground, proposing an incredibly fast turnaround time from identifying programs to producing remediated versions.
I am going to assume that some of the dread accompanying the free-fall of IBM’s stock price stems from financial analysts who know that thousands of companies still rely on millions of COBOL programs running throughout the night on IBM mainframes. Anything that eliminates this legacy code could hurt IBM. But what I believe has them running scared is that they also recognize that the number of COBOL programmers has been decreasing – by a lot – every year. And despite IBM’s best efforts to get young people interested in coding with a nearly 70-year-old language, those efforts have not produced the kind of results financial analysts are hoping for. Always looking at the short-term, they can envision groups of younger, mostly foreign-born students taking classes in Java and Python, and using the auspices of the leading Indian consulting firms (e.g., Tata Consultancy Services, Infosys, Wipro, and LTIMindtree) to obtain H1B visas and help perform the offshore work that companies used so frequently in the early part of this century.
Bottom line: Anthropic may make out like a bandit if Claude can easily transform all that legacy code to something more modern. But my first caution is based on experience. It takes a frighteningly long time to identify all the components of an application. Omitting even a single asset can have unintended consequences for the resulting code. And I will further caution that it will take considerably longer to validate those programs than most stock analysts (and business leaders) probably anticipate. Standalone tests, quality assurance tests, integration testing, and more are not trivial efforts (if done correctly), and the costs are high.
In the end, I don’t doubt that IBM, with nearly three decades of experience, can – and will – find a way to retain its title as the leading AI transformation engine in the field, given this latest nudge. But it is good to know they must do so, looking over their shoulder at the new up-and-coming kid on the block.
Thanks, and safe computing!