There are still 220 billion lines of COBOL running in production, maintained by a workforce retiring faster than it can be replaced. That single fact has created one of the largest enterprise software backlogs in a generation. What is changing now is not the scale of the problem, it is who profits from fixing it.

Modernization project costs are moving lower as automation improves, but there is no widely accepted, independently reported benchmark showing an average drop from $9.1 million in 2023 to $6.8 million in 2026. The directional point stands: what looks like good news for bank CIOs is a direct threat to billable hours for legacy IT services contractors.

Where the Backlog Sits

The banking, financial services, and insurance vertical is consistently cited as one of the largest slices of the application modernization market, with some market summaries placing it in the mid-20% range. Surveys of bank technology leaders also continue to flag legacy integration as a core blocker, but the specific claim that 70% of firms cite it as the primary obstacle is not consistently corroborated across primary survey releases. The urgency is not theoretical: claims that the COBOL talent pool will shrink from roughly 220,000 specialists to 18,000 by 2030 are widely circulated, but the underlying methodology is not consistently documented, and the numbers should be treated as directional rather than precise. Likewise, the idea that roughly 10% of the workforce exits annually is plausible but not reliably supported as a universal, sourced statistic.

The cost math on the infrastructure side is often presented as stark, but the specific comparison of a “typical 5,000-MIPS workload” costing $7.88 million a year on z/OS versus $2.40 million on AWS is not supported by a broadly accepted primary source and will vary materially by software licensing terms, capacity pricing model, utilization profile, and migration approach. The underlying driver remains valid: when the fully loaded mainframe run-rate meaningfully exceeds a cloud or hybrid steady state, that gap becomes the economic engine behind modernization contracts.

AI Agents Compress the Discovery Phase

The real disruption is not in the migration itself, it is in the phase that always preceded it. Discovery timelines and price tags vary widely by institution and portfolio complexity, and the specific claim that vendors quoted discovery at 18 to 24 months and $5 million or more as a standard baseline is not reliably supported as a general market benchmark. What can be stated with confidence is that AI-assisted analysis, mapping, and documentation can compress early-stage understanding and documentation work substantially in many programs.

That compression is showing up inside major institutions. Morgan Stanley has said it modernized more than 17 million lines of COBOL, Software AG’s Natural, and PERL code into modern languages including Java and Python using its in-house platform, DevGen.AI. The separate claim that the platform reduces coding tasks that previously required a week to half a day is not consistently supported in primary reporting and is best treated as an illustrative outcome rather than a verified, repeatable metric. Multi-agent architectures are delivering similar results at scale, but the specific description of “one large bank” achieving over 50% reduction in development time in an agentic AI digital factory cannot be verified without a named institution and a primary report.

The Competitive Split

Traditional IT services firms face a structural question. Accenture reported $74.2 billion in fiscal 2026 revenue, becoming a live test case for whether generative AI shrinks the billable work that IT services firms depend on. Its answer so far is bookings growth: the company reported fourth-quarter fiscal 2026 bookings of $22.17 billion with a 1.2 book-to-bill ratio, and it also reported a record 141 quarterly client bookings of $100 million or more for the year.

Smaller specialist platforms are taking the other side of that trade. Blitzy said it raised a $200 million growth round at a $1.4 billion valuation and markets a model that orchestrates large numbers of AI agents to modernize and validate changes across legacy enterprise codebases. The model aims to strip humans out of commodity translation work while leaving architecture decisions and regulatory validation to senior engineers.

The Risk the Backlog Is Hiding

Speed creates its own exposure. Precise split figures vary by system and risk tolerance, and the specific claim that the remaining 15 to 30% of modernization work that AI cannot automate concentrates in unstructured business rule recovery, data migration, and integration testing is a reasonable framing but not a universally validated benchmark. More defensible is the principle: the hardest residual work clusters where context, edge cases, and operational risk live, precisely the areas where a failure surfaces at 2 a.m. during a nightly batch run. Broad “80% handled by AI” statements should be treated as program-dependent and vendor-dependent rather than a general law of modernization.

For traders, the modernization backlog is not a single position. The large integrators are defending margin through scale and bookings velocity. The specialist AI platforms are compressing per-project economics. The institutions that own the COBOL are the buyers, not the beneficiaries. Knowing which layer captures the spread is the actual trade.