Claude Code takes on ancient COBOL systems

Every time you pull cash from an ATM, chances are you are relying on code written when Richard Nixon was in the White House.

Anthropic just laid out how its agentic coding tool, Claude Code, can tackle one of enterprise tech's most expensive nightmares: modernizing COBOL. An estimated 95% of US ATM transactions and hundreds of billions of lines of production code across banking, aviation, and government still depend on it. The engineers who built these systems retired years ago, universities rarely teach the language anymore, and the documentation simply does not exist.

Legacy rewrites usually stall because understanding old code costs more than writing new software. Anthropic claims Claude Code flips that equation, compressing migration timelines from years of consultant army work down to mere quarters.

Mapping invisible landmines

There are four typical paths for legacy COBOL: rehosting, refactoring, replacing with off-the-shelf software, or rewriting in modern languages like Java or Python. Rewriting pays the biggest dividends, but it historically carried catastrophic risk because nobody actually knows what the software does under the hood.

This is where Claude Code steps in. The tool ingests the entire codebase, traces execution paths, and documents workflows that nobody remembers. Crucially, it uncovers implicit dependencies, such as data shared across obscure files or global state, which standard static analysis usually misses. It then highlights high-risk entangled modules and points out isolated components that are safe to migrate first.

Once the map is drawn, human engineers step in to make strategic and regulatory calls. AI drafts functional tests to ensure modern code produces byte-for-byte identical outputs, wraps legacy blocks in APIs, and runs old and new modules side-by-side. If a single piece fails, you roll back a tiny change instead of months of consulting work.

Why it matters

We think Anthropic is playing a brilliant enterprise game here.

While most coding assistants are busy showing off CSS tweaks or drafting boilerplate Python, Anthropic is aiming Claude Code straight at the deepest pockets in the global economy: legacy financial institutions and federal agencies. Banks are terrified of touching COBOL, but the dwindling pool of maintainers is forcing their hand. Positioning AI not as a magic replacement button, but as a tireless reverse-engineering analyst, makes the pitch palatable to risk-averse CISOs.

For enterprise teams, this is worth testing immediately on an isolated batch job or reporting module. The barrier to entry drops dramatically when you do not have to sign an eight-figure consulting retainer just to figure out what your mainframe actually does.

That said, there are obvious pitfalls. Financial systems are notorious for bizarre arithmetic edge cases and obscure rounding logic. If engineering leaders get overconfident and trust AI-generated unit tests without meticulous human oversight, subtle calculation errors could sit undetected for months. AI can map the swamp, but you still need an experienced pilot behind the wheel.