The Economy of Artificial Complexity
On Monday, IBM lost $31 billion in market value. Its stock dropped 13% — the worst single-day decline in over 25 years.
The cause wasn't a product failure, a security breach, or a quarterly miss. It was a blog post. Anthropic, the company behind the AI model Claude, published a piece explaining how its Claude Code tool can help modernise COBOL — mapping dependencies, documenting workflows, identifying risks in legacy codebases that would take human teams months to surface.
Investors read the post. Within hours, IBM shares were in freefall. Accenture and Cognizant followed. The entire legacy technology services sector took a hit because an AI company said it could do what consulting firms charge millions to do.
This is worth examining — not for what it says about AI, but for what it reveals about how the enterprise technology industry actually works.
The "News" That Wasn't News
Let's start with a basic question: was anything Anthropic announced actually new?
IBM launched its own AI coding assistant for mainframes in 2023. By mid-2025, IBM's CEO was saying it had "very wide adoption" — mostly for understanding existing COBOL codebases and deciding what to modernise. AWS has mainframe migration tools. Microsoft has mainframe migration tools. Kyndryl, IBM's own infrastructure spinoff, has migration services. Last week — days before the Anthropic blog post — Infosys chairman Nandan Nilekani publicly stated that AI has made rewriting legacy applications affordable and imperative.
On a personal level, I used AI to improve COBOL code months ago. Not as a demonstration. As actual work. The results were tangible — cleaner logic, better documentation, faster comprehension of code that had accumulated decades of modifications. The capability has been real and available for some time.
What Anthropic published on Monday was not a technological breakthrough. It was a well-timed marketing announcement that repackaged existing capability with polished language and positioned it as a disruption narrative. The technology community has known for at least a year that AI can meaningfully assist with COBOL analysis and modernisation. The financial markets, apparently, had not been paying attention.
Which raises a question about the quality of analysis driving billions of euros in investment decisions — but that's a discussion for another day.
The Claim vs. The Reality
Anthropic's blog post states that "with AI, teams can modernise their COBOL codebase in quarters instead of years." This is a carefully constructed sentence that deserves careful reading.
What Claude Code demonstrably does well is the analysis phase: mapping dependencies across thousands of lines of code, documenting how programs interact, tracing execution paths, and identifying risks. This is genuinely valuable work. It's also the first phase of modernisation, not modernisation itself.
IBM's response was telling. Senior Vice President Rob Thomas pointed out that translating COBOL is the easy part. The real work is data architecture redesign, runtime replacement, transaction processing integrity, and replicating decades of performance that was optimised at the hardware level. None of that was addressed in Anthropic's blog post.
Consider what a COBOL modernisation project actually involves in critical infrastructure. A system that processes 95% of a country's ATM transactions isn't just code. It's decades of business rules embedded in program logic that was never properly documented. It's workarounds for edge cases discovered during crises twenty years ago. It's performance optimisation that happened at the intersection of software and specific hardware. It's institutional knowledge that exists in the heads of people who are retiring or have already retired.
AI can read the code. AI can map the dependencies. AI can produce documentation that would take a human team months to compile. All of this is real, valuable, and — critically — already possible. What AI cannot yet do is replicate the contextual judgment required to decide what to migrate, what to rewrite, what to retire, and what to leave exactly as it is because touching it would introduce more risk than it eliminates.
The analysis is the radiograph. Someone still needs to read it and decide whether to operate, medicate, or leave the patient alone. And that decision requires something that no AI model currently possesses: thirty years of understanding why a particular workaround exists and what breaks when it's removed.
The Elephant in the Room
Here's the question that the market reaction answered without anyone explicitly asking it: if AI can genuinely accelerate COBOL modernisation, why was the reaction fear instead of excitement?
Think about what this capability represents if it's real. Systems that have been trapped in a sixty-year-old language can finally be updated. Operational risk from dependence on retiring specialists decreases. Maintenance costs that have been climbing for decades could collapse. Organisations could redirect resources from keeping the lights on to building something better.
This should be good news. For the organisations running COBOL, for the customers depending on those systems, for the industry as a whole.
But $31 billion evaporated. And the reason is structural.
IBM, Accenture, Cognizant, and dozens of smaller consulting firms have built enormous revenue streams on top of COBOL's complexity. Every undocumented dependency is billable. Every incomprehensible subroutine requires a specialist. Every retired programmer who took knowledge with them creates demand for a consulting engagement. The difficulty isn't just a problem — it's a product. It generates contracts, sustains teams, and justifies multi-year engagements.
If AI eliminates the complexity — or even significantly reduces it — it eliminates the revenue model that depends on the complexity existing.
The market didn't panic because the technology is dangerous. It panicked because the technology threatens the business model of inertia. Investors weren't calculating the risk to systems. They were calculating the risk to revenue streams that depend on those systems remaining difficult to understand, expensive to maintain, and painful to change.
This is worth stating plainly: a significant portion of the enterprise technology services industry is economically incentivised to keep legacy systems complex, opaque, and expensive. Not through conspiracy, but through the natural mechanics of how services are priced and sold. Complexity generates scope. Scope generates contracts. Contracts generate revenue. Simplification is the enemy — not because anyone says so explicitly, but because the incentive structure makes it so.
The Question That Was Never Asked
The LinkedIn posts reacting to Monday's events followed a predictable pattern. Some celebrated AI's disruptive power. Others defended IBM's mainframe business. A few offered dramatic warnings about companies that don't "embrace transformation."
The fourth question was almost never asked — the one that actually explains what happened: who benefits from things staying as they are?
In every organisation running critical COBOL systems, there are stakeholders whose budgets, teams, headcounts, and influence depend on those systems remaining complex. The IT director whose department exists because mainframe maintenance requires a dedicated team. The consulting partner whose contract renews because only their people understand the code. The vendor whose support agreement is priced based on the scarcity of expertise.
None of these people are villains. They're rational actors within an incentive structure that rewards maintaining complexity and punishes simplification. The IT director who successfully modernises their mainframe has just eliminated the justification for half their team. The consulting partner who makes legacy code transparent has just made their own engagement unnecessary. The vendor who enables easy migration has just given the customer a reason to leave.
This is why legacy modernisation stalls — not because the technology isn't ready, but because the humans around it are incentivised to delay. And this is what $31 billion in market reaction actually measured: not the risk that AI will fail at modernising COBOL, but the risk that it will succeed.
What Should Actually Happen
If the AI capability is real — and the analysis portion demonstrably is — then the rational response isn't panic. It's preparation.
Start with honest diagnosis, not migration plans.
Before any modernisation project, understand what you actually have. Not the architecture diagram from 2015 — the reality. What does the code do? Where are the undocumented dependencies? What business rules are embedded in program logic whose origins have been forgotten? AI tools can accelerate this phase dramatically. Use them for what they're good at.
Separate understanding from action.
The ability to analyse and document a COBOL codebase is not the same as the ability to safely migrate it. These are different phases requiring different skills, different risk profiles, and different timelines. Compressing the analysis from months to weeks is a genuine improvement. Compressing the migration to match is a recipe for the kind of production incident that makes Monday's stock drop look trivial.
Ask who benefits from delay.
In any legacy modernisation discussion, map the incentives. Who gains from the current state continuing? Who loses if modernisation succeeds? Not to assign blame, but to understand why previous attempts stalled and to design governance that accounts for these dynamics. If the people evaluating whether to modernise are the same people whose roles depend on not modernising, the outcome is structurally predictable.
Invest in the judgment layer.
AI can read the code. AI can document the code. AI can even suggest how to rewrite the code. What AI cannot do is decide whether rewriting is the right choice for a specific system in a specific operational context with specific risk tolerances. That judgment requires human experience — deep, contextual, and increasingly scarce. Organisations should be investing in that capability now, before the last generation of people who understand these systems is gone entirely.
The Real Disruption
The real disruption isn't that AI can modernise COBOL. It's that AI is making visible the economics that have kept COBOL in place.
For decades, the cost of understanding legacy systems was so high that it functioned as a barrier — protecting the incumbent service providers, justifying ongoing consulting engagements, and making "do nothing" the rational financial choice for most organisations.
If AI collapses that cost — and it's beginning to — then the question every organisation has been deferring becomes unavoidable: why are we still running this, and who is it serving?
The answer, for many systems, will be "because it works and the risk of change exceeds the benefit." That's a legitimate answer. COBOL running critical payment infrastructure at scale is not a problem to be solved — it's infrastructure to be respected.
But for others, the honest answer will be more uncomfortable: "because changing it would disrupt the people and contracts that depend on it staying the same."
Monday's $31 billion wasn't the market discovering that AI can modernise COBOL. It was the market discovering that the economy of artificial complexity — the entire ecosystem of revenue, roles, and relationships built on keeping legacy systems opaque — might have an expiration date.
That's not a crisis. That's a correction. And corrections, while painful, tend to point in the direction of reality.