The Economy of Legacy has changed
How AI is changing the way we tackle technical debt
On Monday 23rd of February 2026, the IBM share price fell 13.2%. The worst single-day decline since October 2000. The trigger wasn’t an earnings miss or a leadership change. It was a blog post.
That morning, Anthropic published a piece explaining how Claude Code could automate the exploration, dependency mapping, and analysis phases of COBOL modernisation. The thesis, in their own words: “Legacy code modernisation stalled for years because understanding legacy code cost more than rewriting it. AI flips that equation.”
The Someday List Just Got an Honest Price
Every CIO I speak to has the same list. The names change, but the contents don’t; replatform a legacy estate; generate tests for the twenty-year-old codebase nobody dares touch; document the critical system whose architect left in 2019; migrate workloads off some ancient Java version; produce a complete audit trail for the regulator who started asking over a year ago...
All valuable, all on the someday list, but none of it happening.
The hardest part of the business case for any modernisation programme has always been “is it valuable enough to justify the people-hours?”
The cost of a human team doing the work was almost always higher than the value extracted. So the work sat. And sat. Until the next regulator, the next breach, or the next retirement made it urgent enough to fund a one-off project.
Where the Economics Have Shifted
Three categories of work have crossed the business case threshold.
Language migration. Take a COBOL estate you’ve been quietly trying to retire for fifteen years. The first task in any migration is understanding what the system actually does; historically a six-figure consulting engagement before a single line of code gets translated. McKinsey published a case study of one FinTech client with 20,000 lines of COBOL that was estimated at 700 to 800 human hours to migrate properly. After deploying genAI agents, they cut that estimate by 40%. A separate engagement with a top-15 global insurer reported greater than 50% acceleration in modernisation efficiency and testing. Those numbers reframe the conversation entirely.
Major version upgrades. Java 8 to 21. .NET Framework to .NET Core. Spring Boot 2 to 3. These come with release notes the size of a phonebook listing every breaking change. The work has always been mostly mechanical: identify which deprecated APIs are in use, produce the required changes, validate against existing tests. But the mechanical part still took a team of senior engineers six to twelve months. With the help of AI, the bill of materials for a framework upgrade now looks more like a sprint than a programme.
Regulatory uplift. The cadence of new regulation isn’t slowing (DORA, NIS2, EU AI Act). Traditionally, a regulation lands, you get twelve to eighteen months, and someone has to read three hundred pages of legal text, work out which controls apply, then go on a six-month evidence-hunting expedition through your CI logs, ticket system, and seventeen spreadsheets from 2019. AI can now handle much of the heavy lifting of reading the regulation, producing a structured control list, mapping each control to artifacts your platform already produces, and surfacing evidence gaps as engineering tickets. Compliance stops being a quarterly fire drill and becomes a continuous property of the system.
What AI Won’t Do
Agentic AI won’t magically solve all of this. It compresses the analysis, dependency mapping, and translation work that historically consumed the majority of any modernisation budget. Everything else still requires people. The data layer redesign. The runtime translation. The regulatory sign-off cycles that still take twelve to eighteen months in financial services. The organisational change management that determines whether the new system actually gets used.
What’s shifted is the balance of economics. The expensive, low-value toil is now cheap. The expensive, high-value judgement is now where your teams’ time actually goes.
Pick One. Start Small.
Most CIOs are framing AI as a question about new product development. Where can we ship faster? Where can we build something we couldn’t build before?
The highest-ROI AI investments for most enterprises won’t be in greenfield at all. They’ll be in the work you were never going to fund. The modernisations sitting on the someday list. The compliance backlog quietly accruing every quarter. The undocumented systems that have become organisational single points of failure.
Pull the list out. Reprice it. Start with one item; something with clear boundaries, moderate complexity, and a high enough profile that success will be visible but low enough risk that failure won’t take down operations. Use AI to do the analysis and documentation work that always made the business case impossible. Keep your engineers on the architectural, regulatory, and organisational calls because those are still where the value is created and the risks live.
The boring backlog you’ve been postponing for a decade just became the most strategic thing on your desk. Don’t waste it by overestimating what’s been solved, or by underestimating what’s now possible.