Anthropic published a new deep-dive on how AI can break the long-standing cost barrier to modernizing COBOL systems, arguing that the “understand the legacy first” phase used to be more expensive than rewriting—until AI flipped that equation. (Source)

Why COBOL modernization is uniquely hard

COBOL still runs critical infrastructure across finance, airlines, and government, and Anthropic cites estimates that it processes ~95% of ATM transactions in the U.S. The bigger issue: the experts are retiring, documentation is outdated, and decades of patches have embedded business logic directly inside production code.

What AI changes (the expensive part)

The post highlights that AI can automate large chunks of exploration and analysis—mapping program entry points, tracing execution paths, surfacing hidden dependencies (shared data structures, file ops, initialization sequences), and generating workflow documentation by following data from input to output. This is the work that previously required “armies of consultants” over long timelines.

A practical modernization playbook

Anthropic’s recommended approach is incremental: (1) AI-assisted discovery and dependency mapping, (2) risk/opportunity analysis to identify safe early targets, (3) human-led planning for architecture + compliance + priorities, then (4) component-by-component implementation with continuous testing to ensure modernized outputs match legacy behavior. The promise is fewer “big bang” migrations and more controlled, testable transitions.

Why it matters

If AI can reliably compress discovery + documentation + risk analysis, modernization becomes a repeatable engineering workflow instead of a multi-year consulting project—unlocking upgrades to systems that still run huge parts of the real economy.