AI Does Not Modernize a Mainframe. It Can Shorten the Proof.
What measured evidence supports in code translation, documentation recovery, and test generation, and where the risk still concentrates.
It is safe to say yes to an AI-assisted modernization pilot if it is a bounded release with a named business owner, a behavior ledger, independent acceptance tests, reconciliation evidence, and a rollback path. Do not sign off on a portfolio promise built from generated lines of code, successful compilation, or a vendor productivity figure.
The evidence in five moves
GAO's legacy-system evidence points to planning, disposition, cyber risk, interfaces, and delivery control as the durable constraints.
Program-repair studies show plausible patches that later fail manual correctness review. Existing tests are evidence, not semantic proof.
Research supports AI help where outputs can be checked. It does not establish unattended estate conversion or safe cutover.
AI can produce candidate code, explanations, and tests faster than scarce domain experts can validate business behavior unless the evidence pipeline changes too.
Public vendor figures lack the data needed to assess quality, rework, parallel-run cost, or whether the old component actually retires.
Owner, briefing, proof
The accountable business owner decides which outcomes and exceptions must survive, and signs the acceptance record.
A release packet names the behavior boundary, baseline, risk, interfaces, evidence gaps, and decision gate.
Traceable source-to-test-to-reconciliation-to-approval evidence, plus measured post-release behavior and retirement outcome.
There is no public, independent, multi-release study showing that AI reduces total legacy-modernization cost after validation, data conversion, integration, security review, parallel run, rework, and confirmed retirement. IBM's 60% productivity figure is a vendor claim, not independent evidence.
Gates before you scale
Choose one business flow, source component, data set, interface set, exception class, performance constraint, and retirement condition.
Link requirements, source routines, data transformations, interfaces, recovered documentation, generated candidates, tests, reviewer, and sign-off evidence.
Run source-versus-target comparisons, data reconciliations, contract tests, negative cases, and production-like performance checks. Inherited tests are necessary but insufficient.
Track rework, review load, defect escape, reconciliation pass rate, accepted-release lead time, and parallel-run cost beside code output.
Make rollback, ownership, and the old-component retirement decision explicit. A translation that creates a permanent parallel estate has not proved economic value.
Claim ledger
What would change this conclusion
An independently evaluated, multi-release legacy-modernization study that measures AI contribution through validation, data conversion, interface work, security review, parallel run, rework, production outcome, and confirmed retirement. That would answer whether AI reduces total modernization cost or only moves work into the proof queue.
Where this can lead
A sponsor can begin with a release-level diagnostic: owner, behavior boundary, evidence gaps, and acceptance test. If the diagnosis finds a recurring proof gap, build the ledger and release controls across the modernization portfolio.