Enterprises are being told that artificial intelligence can modernize the aging software systems that run their core operations. A former IBM strategist now argues the economics of that proposition simply do not hold up, and the claim deserves scrutiny.

What You Need to Know

Legacy IT systems remain the operational backbone for banks, insurers, airlines and government agencies, and replacing them is among the most expensive projects an organization can undertake. AI vendors have positioned large language models as a shortcut for translating, documenting and migrating decades-old code. The counterargument is that the cost of verifying AI-generated output at enterprise scale often exceeds the cost of the original migration. That tension now sits at the center of how CIOs evaluate AI spending.

The Economics Behind the Skepticism

The core objection is not that AI cannot read old code. It is that the verification burden scales faster than the productivity gain. A model can propose a translation of COBOL into Java in seconds, but confirming that translation preserves decades of undocumented business logic can take human engineers weeks.

That asymmetry matters because legacy systems rarely fail in obvious ways. They fail in edge cases buried in code written by engineers who retired years ago. AI can generate plausible replacements, yet plausibility is not correctness in a payroll or settlement system.

Where the Cost Curve Bends

Vendors typically price AI modernization tools on seats or tokens. The hidden expense sits in review, testing and compliance sign-off. Those costs do not shrink when code generation accelerates, which is why the former IBM strategist's critique resonates with practitioners who have lived through failed migrations.

Organizations that have attempted full rewrites of core banking or insurance platforms know the pattern. Projects run long, budgets expand and the original system often stays in production longer than planned.

Why This Matters

The stakes extend well beyond one consulting opinion. Enterprise software budgets for the next several years will be shaped by whether AI-assisted modernization is treated as a genuine cost reducer or as an expensive experiment.

CIOs who accept vendor claims without independent verification risk committing capital to projects that stall. Those who dismiss AI entirely may miss legitimate gains in narrow tasks such as code documentation, test generation and dependency mapping.

Regulated industries face the sharpest consequences. A bank that migrates a core ledger using AI-generated code without rigorous validation exposes itself to audit findings, operational outages and regulatory penalties that dwarf any licensing savings.

  • Verification cost: Human review of AI-generated code often exceeds the savings from faster generation.
  • Hidden business logic: Decades of undocumented rules in legacy systems resist automated translation.
  • Regulatory exposure: Auditors in banking and insurance require traceable, validated change control.
  • Vendor incentives: Tool pricing models reward usage, not successful migration outcomes.

What You Need to Know About the Vendor Pitch

AI vendors have spent the past two years repositioning their products as modernization platforms. The pitch is compelling on a slide: feed legacy code into a model, receive modern code out, reduce reliance on scarce COBOL and mainframe specialists.

The reality inside enterprises is messier. Legacy estates are heterogeneous, poorly documented and tightly coupled to downstream systems. A single misread field length can cascade into reconciliation failures that take days to trace.

None of this means AI has no role. Targeted use in documentation, test scaffolding and impact analysis can deliver measurable value. The dispute is about scope, and about whether the technology can carry the weight vendors place on it.

The Broader Pattern in Enterprise AI

This debate mirrors a wider correction underway in enterprise AI adoption. Early enthusiasm has given way to harder questions about return on investment, data readiness and change management.

Legacy modernization is simply the most expensive test case. If AI cannot demonstrate economic advantage there, the burden of proof shifts to every other enterprise use case making similar claims.