Krishnamurthy told Patel he could do the job with a two-person team. The budget line that changes is spending on AI tokens, not engineers.
That math matters for banks, where old code comes with big maintenance bills. Large financial institutions put 70% to 85% of their IT budgets toward maintaining older systems, Amazon Web Services said. Estimates from IT services firm DXC show 45 of the world’s 50 largest banks rely on mainframe applications for critical operations, PYMNTS reported.
The spending is high because the work is manual and complex. Before any code can be rewritten, someone has to find out what it does, and the people who wrote it have largely retired. Itaú Unibanco, Brazil’s largest bank, said in an AWS case study that its engineers spent 50% of their time analyzing what existing code did and another 30% manually creating tests. Douglas Santos, Itaú’s mainframe platform executive, said the bank had billions of lines of COBOL, a programming language from 1959, with scarce documentation.
Itaú is not unusual. Systems written in COBOL still run an estimated 95% of U.S. ATM transactions, and hundreds of billions of lines remain in production across finance, airlines and government, PYMNTS reported. More than 40% of banks still run COBOL-based cores, according to DXC estimates PYMNTS cited, and 71% of mainframe teams report understaffing, AWS said. Bankdata, the IT provider for a group of Danish banks, said in 2025 that it still had more than 70 million lines of code on the mainframe.
Itaú Cuts Mainframe Migration Time by Up to 70%
Itaú has put AI agents on the part of its systems that has resisted migration longest. The bank and GFT Technologies said that a tool they built together, Batch2Cloud, cut migration time for the bank’s batch systems by up to 70% and sped project delivery by up to 3.3 times in the initial implementation phase. Batch systems are the scheduled routines that run closings, settlements and reconciliations, and they have sat on mainframes for decades. GFT said the agents handle repetitive tasks, analysis, code generation and validation, while specialist teams keep responsibility for technical decisions, architecture and governance.
Other parts of Itaú’s migration show similar gains. Santos said application discovery that once took three months was done in two weeks using AWS Transform and Kiro, Amazon’s AI development tools. Business-rule extraction that required weeks of expert interviews now takes hours. In applications where the tools were used end to end, migration time per application fell about 75%. More than 70% of the bank’s business services already run on AWS, and it says it is on track to retire its mainframes by the end of 2028.
Few banks are that far along. Danske Bank, Denmark’s largest bank, used AI agents to modernize one mainframe application of about 140,000 lines across 583 files, AWS said. AWS separately credited Danske with 60% overall acceleration and four times faster forward engineering.
Banks Must Prove the New Code Moves Money Correctly
Faster modernization does not remove the cost of getting it wrong. U.K. regulators fined TSB 48.65 million pounds (approximately $65.28 million) in 2022 after a 2018 platform move locked customers out of their accounts, PYMNTS reported. Legacy banking code holds decades of transaction rules, interest calculations and regulatory requirements. A generated replacement that drops one exception can misprice a loan or fail a payment.
That is why the AI output goes through the same gates as any other code change. Itaú built a validation framework to confirm that every business-critical rule survived extraction before it was rebuilt, and it said the generated code passed its security and quality pipeline on the first pass.
AI is already inside bank technology teams even where the cores have not moved. A PYMNTS Intelligence report, “Financial Services Pulls Ahead in the Enterprise AI Race,” found a 25% adoption rate for AI in API orchestration and integration among financial services firms, and 30% named data quality and fragmentation as their leading obstacle to further deployment.
For all PYMNTS AI and digital transformation coverage, subscribe to the daily AI and Digital Transformation Newsletters.
