AI-augmented mainframe modernization

The readiness assessment is $15,000 and takes three weeks, and it produces the number for everything after it: effort per workload, the wave sequence, and a fixed-price proposal. A pilot on one business-critical workload starts at $60,000. Waves are priced individually because the honest drivers are estate size, language mix, data complexity, integration count and how much test data exists. Any tooling licence is itemised in the proposal rather than hidden in the day rate.

Historically, years, and many never finish: only 22% of started mainframe modernizations were called a success in a 2023 survey of 400 executives. AI compresses the two phases that consumed that calendar, discovery and test writing. It does not shorten the parallel run, and should not. Our assessment takes three weeks and a first pilot workload runs weeks four to ten. After that, delivery is wave by wave rather than one date, so modernized workloads reach production within months while the rest of the estate is still being worked. How many waves your estate needs is what the assessment tells you before you commit.

COBOL, PL/I, Natural and Assembler on z/OS, with the JCL, CICS, IMS, Db2, VSAM and IDMS around them, and RPG and CL on IBM i. The transformation engine differs per language, which is why the assessment names the tooling per workload instead of assuming one translator covers the estate. Where a language has no mature translator, the path is usually Reimagine: extract the rules, rebuild from them, and prove parity the same way.

Partly, and the distinction matters. AI tooling produces code that compiles and a first draft of the tests, and it does that in days rather than months. What it cannot do on its own is guarantee that the new program produces the same numbers as the old one on your data. A 2024 ICSE study of 1,700 translated samples found the language models it tested produced correct translations between 2.1% and 47.3% of the time, and Gartner predicts that more than 70% of mainframe exit projects started in 2026 will miss their intended benefits because generative AI capabilities are overestimated. That is why generated code is reviewed by engineers and gated on parity tests before it goes anywhere near production.

Three ways, in order. Every extracted business rule is traced to the source file and line it came from. Characterization tests capture the current behaviour before anything changes, and the modernized workload has to match it on production data. Then the two run in parallel on real volume, including the nightly batch, until a named engineer signs the wave off. Parity is a gate, not a report.

Functional parity is not enough on a mainframe, so performance is part of the gate criteria. During the parallel run we measure batch completion against your actual window and transaction response times against your current baseline. If either misses, the wave does not advance until the architecture or the tuning fixes it.

Analysis and transformation run in a cloud account or environment you control, with an audit trail of what each tool produced and which source lines it came from. Model-training, retention and data-residency terms for every tool in the chain are reviewed with your security team and agreed in writing before any code is uploaded.

AWS, Azure, Google Cloud or a private cloud. The target is chosen in the assessment on your constraints, not on a partnership: data residency, the platforms your teams already run, the licensing you already hold, and where the data the workload depends on lives. We deliver on all of them, and the landing zone, Infrastructure as Code and observability work is the same discipline whichever you pick.

No, and a partner who says otherwise is selling. Retain is one of the six paths, and for some workloads it is the right one: integrated with the cloud estate rather than migrated. Retire is another. Most estates carry jobs nobody has run in years, and the analysis finds them before you pay to move them.

Then you take the deliverables with you. The assessment is fixed-price and zero lock-in: the inventory, the dependency map, the path per workload with its reasoning, the wave plan and the priced proposal are yours to use with any partner. The pilot works the same way, with acceptance criteria published before it starts. We would rather win the waves on the pilot’s results than on a contract clause.