Mainframe modernization: approaches, risks, and how to decide per workload

Share in
17-11-2026

Mainframe modernization means changing how your mainframe applications, data, and operations are built, connected, and run, one workload at a time. It's rarely a simple stay-or-leave choice.

In Kyndryl's 2025 State of Mainframe Modernization Survey of 500 senior leaders at mainframe-using enterprises, only one respondent planned to move entirely off the mainframe. Yet 98% were moving some applications to other platforms. That gap is the decision you actually face.

Hyperscalers (large cloud providers such as AWS and Google Cloud) tend to pitch migration, while the IBM ecosystem tends to pitch modernizing in place. Each is right for some workloads and wrong for others. This guide compares the approaches, then gives you a scoring worksheet and a decision tree to route each workload on its own merits.

Back to top

What is mainframe modernization (and what it isn't)?

Mainframe modernization is the process of updating mainframe applications, data, and operations so they cost less to run, change faster, and connect with cloud and AI services. The methods include API (application programming interface) enablement and DevOps (automated build, test, and release practices). They also include replatforming (moving code to a new runtime), refactoring (converting code to a modern language), and migration to another platform.

Modernizing doesn't have to mean leaving. Many programs keep the IBM Z hardware and z/OS, IBM's mainframe operating system. What changes is everything around them, from how code is built and tested to how data is shared and how other systems call the core.

Mainframe modernization without leaving the mainframe: COBOL applications, CICS and IMS, Db2, and IBM Z stay in the core while REST APIs, data sharing, and how code is built and tested change around it

The term gets used loosely, so separate two scopes:

  • Mainframe modernization covers the whole estate, from hardware to applications. Middleware is in scope, such as CICS (Customer Information Control System, IBM's transaction server) and IMS (Information Management System, a transaction manager and database). So are Db2 (IBM's relational database) and the batch jobs run by JCL (Job Control Language) scripts.
  • Mainframe application modernization covers only the application code and its interfaces, typically programs written in COBOL or PL/I, two business programming languages dating from the 1950s and 1960s.

Scope sets your budget and your risk. IBM notes that mere code translation often fails to address the real problem; in IBM's view, successful projects also handle data architecture, runtime replacement, and transaction processes. Convert only the code, and you can still inherit a slow batch window (the time slot reserved for bulk batch processing) and a brittle data model.

Back to top

Are mainframes becoming obsolete?

No. IBM is still building new mainframes, and its hardware roadmap stayed active through 2025 and 2026:

  • IBM z17. IBM announced the z17 on April 8, 2025, with general availability on June 18, 2025. Its Telum II processor has an on-chip AI accelerator, and IBM, extrapolating from internal tests, rates z17 at up to 450 billion inference operations a day. Inference means running trained AI models on live data.
  • Smaller z17 models. In July 2026, IBM added single-frame and rack-mount z17 configurations, generally available on August 12, 2026.
  • Next-generation chip. In August 2026, IBM previewed a dual-architecture processor for future IBM Z and LinuxONE (IBM's Linux-only mainframe line) systems. Each core is designed to run both IBM Z and Arm instructions (the chip architecture behind most smartphones and many cloud servers).

Customers aren't walking away either. In the 2026 BMC Mainframe Survey of more than 1,300 mainframe practitioners and decision makers, 94% expressed long-term confidence and continued investment in the platform.

What's aging is the operating model around the machine. Respondents in Kyndryl's 2026 State of Modernization Report say nearly half of their mainframe environments are more than ten years old.

The real constraints are scarce skills, slow release cycles, and cost measured in MIPS (millions of instructions per second), the capacity unit many teams budget against. Mainframes won't disappear soon. The way you staff, build, and connect them, though, has to change.

Data center server racks behind the point that mainframes are not becoming obsolete, with 94% of more than 1,300 practitioners confident in the platform and the real constraints of scarce skills, slow release cycles, and cost measured in MIPSBack to top

Why enterprises modernize mainframes now

Enterprises are modernizing now because five pressures have converged, and each one lands on a different workload. Kyndryl's 2025 survey puts numbers on most of them:

  • Skills gap. In the Kyndryl 2025 survey results, 70% of organizations struggle to find the right multi-skilled talent.
  • Cost. MIPS consumption and software licensing tie budget up in running systems rather than changing them. Change itself is getting cheaper, though. Kyndryl's average cost of modernizing on the mainframe fell from $9.1 million in its 2024 survey to $7.2 million in 2025, mainly through lower licensing, hardware, and labor costs.
  • Speed of change. Digital channels expect frequent releases. Mainframe changes often wait for scheduled release windows and manual regression testing (re-checking that existing functions still work after each change).
  • Data and AI access. Your most valuable transaction data sits on the mainframe, yet it's hard to feed into analytics and AI. Kyndryl found 88% of respondents have implemented or plan to implement generative AI in their mainframe environments.
  • Regulation and security. In the same survey, 94% say regulatory compliance strongly influences their modernization decisions, and 32% cite security as a reason they kept an application on the mainframe.

Three of these drivers pull in opposite directions:

  • The skills gap pushes code toward modern languages and tools.
  • Regulation and security pull critical workloads to stay where they are.
  • Data access argues for opening the core through APIs rather than moving it.

With drivers in conflict, any one answer for the whole estate will be wrong for part of it. Deciding per workload avoids that trap.

One mainframe workload pulled three ways: regulation and security pull it to stay, the skills gap pushes code toward modern languages, and data access argues for opening the core through APIsBack to top

Mainframe modernization approaches compared

Six approaches cover most programs, and each can pay back when it fits the workload. Kyndryl's 2025 respondents reported return on investment (ROI) of 288% for modernizing on the mainframe, 297% for integrating with cloud, and 362% for moving workloads off. They're self-reported, so read them as evidence that no path is a dead end, not as a ranking.

The table rates the six approaches against each other, not in absolute terms:

Approach

Speed to first value

Delivery risk

Cost reduction potential

Skills needed

Keeps mainframe?

Modernize in place

Fast

Low

Low to medium

Mainframe plus DevOps

Yes

API-enable and integrate

Fast

Low

Low

API and integration

Yes

Rehost or replatform

Medium

Medium

Medium

Cloud operations plus COBOL

No

Refactor

Slow

High

High

COBOL, Java, test automation

No, or partly

Replace

Medium to slow

High

Medium to high

Business analysis, package setup

No

Retire

Fast

Low

High for that application

Data archiving

No

Modernize in place (DevOps, APIs, z/OS upgrades)

Modernizing in place keeps the code and platform but changes how you work on them, through Git-based source control, automated testing, and newer hardware. IBM says COBOL Elevate for z/OS, scheduled for availability from September 18, 2026, optimizes COBOL programs without source code changes or recompilation.

Best when: the workload is stable, critical, and dependent on mainframe service levels and security.

Trade-off: you still depend on scarce mainframe skills and mainframe costs.

API-enable and integrate (hybrid)

API enablement wraps existing CICS or IMS transactions in REST APIs (standard web interfaces), so digital channels can call the core without anyone changing its code. You end up with a hybrid setup, where the core stays on the mainframe and newer systems call it.

Best when: digital channels need mainframe data or transactions quickly.

Trade-off: the application keeps its technical debt (the cost of past shortcuts in its code), and the API layer becomes one more system to run.

Rehost / replatform

Rehosting or replatforming moves COBOL applications onto a compatible runtime on cloud or commodity servers (standard, non-mainframe hardware), with little code change.

Best when: you want to exit the hardware but keep proven code.

Trade-off: the language and architecture stay the same, so the skills problem moves with the code.

Refactor (incl. automated code conversion)

Refactoring converts COBOL or PL/I into Java or another modern language, usually in stages. AI tools now shorten both the conversion and the testing.

AWS Transform for mainframe claims it compresses multi-year timelines into months. On the testing side, IBM watsonx Code Assistant for Z generates unit tests that compare the new Java service with the original COBOL.

Best when: business logic changes often and you need developers who can work on it.

Trade-off: converted code can drift from the original logic, so you need functional-equivalence testing (proof that new code returns the same outputs for the same inputs).

Replace or retire

Replacing swaps a custom application for a software-as-a-service (SaaS) product; retiring switches it off. Retirement is often the cheapest first step. Bridgestone decommissioned unused applications before migrating its four truly critical ones.

Best when: a commodity product does the same job, or the application has little remaining value.

Trade-off: replacement forces business process change, which is often harder than the technical work.

Back to top

How to decide workload by workload

Workloads on the same mainframe differ in risk, change rate, and value, so decide one workload at a time. Legal & General shows what that looks like.

According to an OpenLegacy case study, the insurer separated "softcore" ancillary systems such as document management from "hardcore" systems such as claims processing. It then moved the softcore applications to modern platforms and kept the hardcore systems on the mainframe behind APIs.

This matrix maps common workload profiles to the approach that usually fits:

Workload profile

Likely approach

High-volume core transactions, strict service levels

Keep on mainframe and modernize in place

Core logic needed by digital channels

API-enable (hybrid)

Batch reporting, low change frequency

Rehost or move to cloud

Frequently changing business logic

Refactor incrementally

Duplicates a commodity SaaS function

Replace

Rarely used, low value

Retire

Real profiles blur, so use the two tools below for the in-between cases.

Score each workload

Score each workload from 1 (low) to 5 (high) on these five criteria:

Criterion

Question to ask

Score (1 to 5)

Business criticality

What breaks if this is down for an hour?

 

Change frequency

How often does the business ask for changes?

 

Data gravity

How many other systems read or write this data?

 

Regulatory constraint

Are there data residency or audit rules?

 

Skills available

Do you still have people who understand it? (5 = yes)

 

Then combine the scores into two numbers and read them with these rules of thumb:

  • Stay pressure = criticality + data gravity + regulatory constraint (range 3 to 15). A score of 12 or more means you should keep the system of record (the authoritative source for that data) on the mainframe.
  • Change pressure = change frequency + (6 − skills available) (range 2 to 10). A score of 7 or more means you should act on the code early through APIs or incremental refactoring.
  • Both high points to hybrid. Both low sends the workload to the decision tree below for rehost, replace, or retire.

Here is an illustrative example for three workloads at a fictional insurer. Treat the thresholds as starting points to tune, not industry standards:

Workload

Criticality

Change

Data gravity

Regulation

Skills

Stay pressure

Change pressure

Reading

Claims processing

5

4

5

5

2

15

8

Hybrid: keep core, API-enable, refactor what changes

Document management

2

2

2

3

2

7

6

Tree: SaaS exists, so replace

Internal reporting batch

2

1

3

2

3

7

4

Tree: logic rarely changes, so rehost

Run each workload through a decision tree

Use the decision tree for any workload the scores leave unclear:

Is the workload still needed? ↓ No → Retire ↓ Yes Does a commodity SaaS do the same job? ↓ Yes → Replace ↓ No Does it need near-zero downtime and mainframe-grade throughput? ↓ Yes → Keep and modernize in place, or API-enable ↓ No Does the business logic change often? ↓ Yes → Refactor incrementally ↓ No → Rehost / replatform

Most portfolios end up hybrid. Kyndryl's 2025 data reflects this: 99% of respondents operate hybrid environments, and organizations expect to move 28% of their applications off the mainframe on average.

Back to top

Can AI replace the mainframe or COBOL developers?

No. AI can't replace the mainframe or the people who understand it, but it is changing how fast you can modernize. AI tools now read legacy code, write documentation, convert languages, and generate tests.

AWS reports that Itaú Unibanco cut reverse-engineering time, the work of recovering business rules from existing code, by more than 90%. The bank used AWS Transform for that step and the Kiro coding agent for forward engineering, which builds the new application.

Even AWS, which sells these tools, says humans stay in the loop. Asa Kalavade, VP of AWS Transform, told ITPro: "You do need those COBOL experts, because how do you validate this code?"

Practitioners are cautious too. In BMC's 2026 survey, 40% would let AI recommend code-management actions, but only 23% would let it complete them.

Bar chart from BMC's 2026 survey showing 40% of practitioners would let AI recommend code-management actions but only 23% would let it complete them, beside the AWS Transform quote that COBOL experts are still needed to validate code

Four trends are shaping mainframe modernization in 2026:

  • Agentic AI for modernization. Agentic AI means AI agents that plan and carry out multi-step tasks. In Kyndryl's 2026 research, 10% of organizations run it in production for dependency mapping and code conversion.
  • AI inference next to the data. Every Kyndryl respondent with a mainframe plans to deploy AI on it, and 23% already have AI in production there.
  • Sovereignty-driven placement. Sovereignty means control over where data sits and which laws apply to it. Kyndryl names sovereignty or governance as the leading reason companies moved workloads back onto the mainframe in the past 12 months.
  • A broader software ecosystem. IBM's dual-architecture processor aims to run Arm-native Linux alongside z/OS on future systems.
Back to top

Risks of mainframe modernization and how to control them

Five risks recur in troubled mainframe programs, and each one has a specific control.

Hidden business logic and missing documentation

Business rules hide in decades of patches, and the people who wrote them have often retired. In Kyndryl's 2026 survey, just 9% of 1,000 application and network modernization respondents, not only mainframe shops, had fully mapped their application dependencies.

Control: run code analysis before you commit to an approach. Tools such as Google's Mainframe Assessment Tool extract business rules and group assets by call and data dependencies. Then interview your subject-matter experts (SMEs) and write the rules down before any conversion starts.

Data migration and integrity

Mainframe data rarely maps cleanly to a relational database. EBCDIC encoding (the mainframe character set), packed decimal fields, and hierarchical IMS structures all need conversion. Some errors can surface only in month-end or year-end batch runs, long after cutover (the switch from the old system to the new one).

Control: move workloads in phases and replicate data continuously with change data capture (CDC), which passes each database change to the target in near real time. Before every cutover, reconcile record counts and balances automatically.

Functional regression after refactoring

Converted code can compile and pass unit tests yet still differ in rounding, date handling, or rare edge cases.

Control: require functional-equivalence testing and a parallel run for every workload you convert. Google Cloud's Dual Run runs workloads on the mainframe and the cloud at the same time and compares outputs such as daily reports or database snapshots.

Parallel run during mainframe modernization: the same workload runs on the mainframe and the converted cloud system, CDC replicates database changes, and outputs are compared so a match leads to phased cutover and a mismatch is caught before cutover

Downtime on critical systems

A cutover can fail even when the data moves correctly. In April 2018, UK bank TSB migrated its customer data to a new IT platform.

According to the Bank of England, the data migrated successfully, but the platform immediately experienced technical failures. TSB took until December 2018 to return to business as usual, and regulators later fined it £48.65 million.

Control: cut over in phases by workload or customer segment, keep a tested rollback plan, and run old and new systems in parallel until outputs match.

Skills drain during the program

The SMEs you depend on are the same people retiring or being recruited away mid-program. Kyndryl's 2025 survey found 70% of organizations already struggle to find multi-skilled talent, and a multi-year program only raises that exposure.

Control: keep your SMEs on retention terms through decommissioning, and pair each of your mainframe engineers with a cloud engineer so knowledge transfers during the work.

Back to top

A phased mainframe modernization roadmap

A phased roadmap lets you prove value early and change course without starting over. You'll likely need that room, because 80% of organizations in Kyndryl's 2025 survey shifted their modernization strategy within a single year.

Five phases cover most programs, each with a clear exit point:

  1. Assess. Inventory your applications, interfaces, batch jobs, and MIPS consumption, then map their dependencies. Capture your baseline cost, release frequency, and incident rates now, because phase 4 measures against them. You exit when every workload has an owner and a dependency map.
  2. Decide. Apply the scoring worksheet and decision tree to each of your workloads. You exit when each one has an approach and a business case.
  3. Pilot. Pick a low-risk workload with visible value, typically an API enablement or a batch job. Staying clear of customer-facing transactions lets you test tools and team with less regulatory exposure. You exit when the pilot runs in production with functional equivalence proven.
  4. Scale in waves. Move your workloads in groups small enough to roll back within one maintenance window. Track cost, release frequency, and incident rates against your baseline, and exit each wave when those metrics hold or improve.
  5. Decommission. Switch off the old components only after data reconciliation is complete and your workload owners have formally signed it off.
Phased mainframe modernization roadmap with five phases, assess, decide, pilot, scale in waves, and decommission, each with its exit criterion and the assess baseline used to measure every wave

Budget for surprises in the assess and pilot phases. A CTO quoted in Kyndryl's 2026 report said around 25% of the original effort sometimes shifts into data cleansing, dependency mapping, and automated regression testing.

Back to top

Choosing a mainframe modernization partner

Choose partners by workload, not by brand, because each partner type leans toward one answer. Outside help is common, too. In Kyndryl's 2025 survey, 74% of respondents rely on third-party providers.

Here is where each partner type is strong and what to watch for:

Partner type

Strength

Watch out for

Hyperscaler programs (AWS, Google Cloud)

Tools and migration funding

Bias toward moving off the mainframe

IBM and the mainframe vendor ecosystem

Deep platform expertise

Bias toward staying

Global systems integrator (SI) or specialist modernization firm

End-to-end program delivery

Cost and senior staffing levels

Application and cloud engineering partner

Surrounding apps, APIs, data pipelines, testing

Should not claim mainframe expertise

Hyperscaler funding is real. The AWS Migration Acceleration Program offers investment incentives for mainframe modernization, and you should weigh that money against the bias that comes with it.

Before you sign, ask every partner three questions:

  • Which approach have you advised a client against? A partner who has never recommended staying, or never recommended leaving, is selling one answer.
  • How will you prove functional equivalence? You should hear about parallel runs and output comparison, not only unit tests.
  • Who owns rollback during cutover? The contract should name the team and the trigger.

Hybrid programs also create substantial work that never touches COBOL, such as the APIs, cloud services, and test automation around the core. If that surrounding layer is where your team lacks capacity, an application engineering partner such as GMO-Z.com RUNSYSTEM can take it on.

That partner can handle system migration, application development, and testing for those workloads, while you keep the mainframe-side work with a specialist.

Back to top

Modernize the mainframe portfolio, not the platform

Mainframe modernization works when you treat it as a portfolio of workload decisions, and the survey evidence points that way. In Kyndryl's 2025 data, only one of 500 respondents planned a full exit, 98% were moving something, and all three broad paths reported positive returns.

Platform decision versus portfolio decision: one stay-or-leave answer for every workload compared with a mainframe core that stays and modernizes in place while other workloads are API-enabled, rehosted, refactored, replaced, or retired

Three mistakes undo many programs:

  • Big-bang rewrites that put every workload's risk into one cutover.
  • Vendor-led approach selection that applies one answer to the whole estate.
  • Skipped equivalence testing that lets logic drift reach production.

This quarter, build a workload inventory, score each workload with the worksheet, and run the unclear ones through the decision tree. The resulting short list tells you which partners you need, and in what order.

Bring that scored list to every vendor meeting. When a hyperscaler or an IBM partner pitches one answer for the whole estate, the list turns a platform debate into a workload-by-workload conversation. That's where the real savings and risks sit.

Back to top

FAQs

What are the main mainframe modernization strategies?

The six main strategies are modernize in place, API-enable, rehost or replatform, refactor, replace, and retire. Most programs combine several. In Kyndryl's 2025 survey, 98% of respondents were moving some applications off the mainframe, but only 1 of 500 planned a full exit.

Is IBM still making mainframes?

Yes. IBM announced the z17 on April 8, 2025, and it became generally available on June 18, 2025. IBM announced single-frame and rack-mount z17 models in July 2026 and previewed a dual-architecture IBM Z and Arm processor for future systems.

How long does mainframe modernization take?

Expect months for a focused scope and years for a full estate. AWS reports that Bridgestone moved four critical applications to its cloud in about seven months. AWS's Asa Kalavade says such projects have typically taken two, three, or five years.

What's the difference between mainframe migration and modernization?

Mainframe migration moves workloads off the mainframe to another platform, such as the cloud. Modernization is broader. It can keep workloads on the mainframe while changing how they're developed, connected, and operated, for example through APIs or DevOps.

Back to top
추천 기사
20년의 경험을 보유한 GMO-Z.com RUNSYSTEM의 전문가에게 기술 컨설팅을 받아보세요
GMO-Z.com RUNSYSTEM