Legacy systems were long regarded as a drag: hard to understand, expensive to operate, barely connectable to modern web and cloud architectures. Two current reports paint a more nuanced picture. Google Cloud is using artificial intelligence to analyse mainframe applications and gradually move them to the cloud. In parallel, a study by Rocket Software describes how mainframes in many companies are no longer perceived as an obstacle to AI, but as a building block of AI strategies.

Iterative instead of big bang

The approach described by Google aims at iterative modernisation: applications are not replaced in a single big cut, but moved to the cloud in stages. AI supports two tasks that typically cause the greatest effort in modernisation projects – the analysis of the existing estate and the transformation of the code.

The analysis is the underestimated part. Grown systems contain business logic that has emerged over decades in COBOL or comparable languages – COBOL is a programming language developed for commercial mass processing that still runs productively in banks, insurance companies and administrations today. Documentation is often missing, and the people who made the decisions are no longer with the company. Anyone who starts here without tool support spends months on archaeology.

This is exactly where AI support takes effect: it can make dependencies visible, describe program structures and prioritise candidates for migration. The actual transformation – for example, the conversion of COBOL logic into a modern target language – is only one step of several. The description of the undertaking as more than a pure language translation indicates that data storage, interfaces and the operating model must also be considered.

From obstacle to data source

The second perspective turns the usual narrative around. According to the study by Rocket Software, mainframes even strengthen AI initiatives – through their data basis and their reliability. That is plausible: core systems contain consistent transaction data maintained over long periods. For AI applications that are to build on reliable enterprise data, this is an advantage over scattered data silos in newer systems.

At the same time, the study names a problem that many operators know: the shortage of skilled workers in mainframe operations. Here too, AI is supposed to help by making knowledge accessible that was previously tied to individual experts. Both reports thus converge on the same point, albeit from different sides: AI lowers the cost of understanding legacy systems – whether one subsequently wants to migrate them or continue operating them.

What the sources do not say

The available information remains at the conceptual level. There are no reliable figures on time savings, error rates or costs, nor any statements about how much rework machine-generated code requires in practice. This gap is relevant, because experience shows that it is not the initial translation that decides project success, but the question of whether the result is professionally verifiable and maintainable in the long term.

The two perspectives are also not congruent. One source argues for the path from the mainframe to the cloud, the other for the mainframe as part of future AI architectures. This is not a contradiction, but a strategic fork in the road: not every legacy system has to be replaced in order to enable modern applications. Sometimes it is enough to connect core systems via clean interfaces and to handle further development in modern web frontends.

Transferable to smaller legacies

Very few mid-sized companies operate a mainframe. However, the mechanics are the same, only on a different scale. A ten-year-old PHP application, a self-built ERP module or a TYPO3 installation with heavily customised extensions is also a legacy system: valuable in business terms, technically hard to move, tied to a few key people.

The lessons from the mainframe discussion can be transferred:

  • Analysis first. Before discussing target architectures, the existing estate should be described – what business logic exists, what dependencies there are, which parts are actually used.
  • Migrate in stages. A step-by-step rebuild allows corrections and keeps the system running at all times. Complete redevelopments more often fail because of scope and duration.
  • AI as a tool, not as a decision-maker. Machine-generated analyses and code suggestions speed up the work but do not replace professional sign-off.
  • Treat data as an asset. If legacy systems contain the most reliable data basis in the company, structured access to them is more important than their rapid replacement.
  • Plan for knowledge transfer. What only a few people understand today should be documented in the course of modernisation – this is a concrete benefit of AI-supported code analysis.

Realistic expectations

The announcements show a direction, not a shortcut. Modernisation remains a project with business decisions, testing effort and organisational questions. What is changing is the entry threshold: undertakings that were previously not touched at all because of the analysis costs are becoming more calculable.

This applies to the COBOL application in the data centre just as much as to the grown individual software in medium-sized businesses.

For web and software projects, this above all means a changed sequence: anyone planning a modernisation should begin with a tool-supported inventory instead of jumping straight to the target architecture.And anyone who has so far assumed a replacement should check whether a clean connection of the legacy system via interfaces does not bring benefits faster.In both cases, AI is an accelerator for understanding – responsibility for architecture and acceptance remains with the project team.

Sources