'Is AI the Answer to Legacy Modernization? A Former IBM Strategist Says “The Economics Don’t Add Up”'

On Monday morning, changes to a core business system brought some billing processes to a halt. One person searches for old design documents while another team checks the impact on connected systems. Only a handful of people know the system well. An urgent fix is needed, but no one can determine how much can safely be changed.

In situations like this, it would help if AI could read the code, identify dependencies, and even suggest fixes. But can AI finally solve the legacy system problems that have been put off for years?

A Kyndryl survey covered by ZDNET Japan reflects both optimism and caution. AI, particularly agentic AI, could accelerate modernization. Yet many companies have found that past modernization efforts failed to deliver results that justified the time and expense. A former IBM strategist also argues that the economics of addressing legacy systems do not add up. It is too soon to assume that adopting AI will automatically make modernization cost-effective.

Why Modernization Keeps Getting Put Off#

Modernizing legacy systems, including mainframes, has long been a challenge for CIOs. Each wave of technology—web services, microservices, cloud computing, virtualization, and x86-based Windows and Linux servers—has brought renewed efforts to make systems more flexible.

Even so, modernization often slips back into the category of “something to tackle in five years.” Existing systems are complex, and years of modifications have introduced behavior that design documents alone cannot explain. They also have numerous connections to surrounding systems, making it difficult to predict where a change to one function will have an impact. Because business operations must continue during migration, old and new systems need to run in parallel.

As a result, modernization is rarely a simple matter of replacing something old with something new. It becomes a lengthy project encompassing requirements validation, code analysis, data migration, testing, and adoption by the people who use the system. The full scope may be unclear when estimates are made, leading to delays and additional costs along the way.

In Kyndryl’s survey of 2,000 senior IT decision-makers, only about half of companies undertaking modernization said they had achieved improvements in IT operations. Fewer than half reported benefits in innovation. Eighteen percent said the value delivered was limited or unclear. Nearly half of respondents had also faced schedule delays or cost overruns.

With burdens this substantial, securing investment approval can be difficult even when the need for modernization is clear. This gives weight to the former IBM strategist’s observation that “the economics don’t add up.” The point can be read as a warning that a uniform approach to replacing every company’s complex systems may fail to deliver results that justify the cost.

AI Can Start by Reducing What We Don’t Know#

The hope for AI here is not that it will rebuild an entire system in one go. Its initial role is to support investigation and documentation that have traditionally relied on manual work, making modernization easier to assess and plan.

In Kyndryl’s survey, only 10% of technology leaders said they were using agentic AI as a tool for IT modernization. Fewer than 10% of organizations were confident that they fully understood the dependencies between old and new systems. Being able to read the code being modernized does not necessarily mean understanding its data flows, related business processes, or calls to other systems.

Agentic AI can assist with several tasks in these situations. It can, for example, uncover hidden dependencies, generate candidate code, and document existing specifications and work performed. Shortening tasks that people previously investigated one by one could free up time for assessing the impact of changes and planning tests.

However, AI-generated explanations and code cannot simply be assumed to be correct. Older systems may contain undocumented business rules and exception handling. AI analysis needs to be checked against the knowledge of operational staff, existing logs, and test results. Generated code must also be verified to ensure that it runs in the target environment and meets business requirements.

Kyndryl warns that organizations that do not use agentic AI could fall significantly further behind schedule than early adopters. This does not mean that companies adopting AI are guaranteed to succeed. What matters is not just whether AI is used, but whether the scope is focused, people verify the results, and project outcomes can be measured.

Test the Economics on a Small Scale Before Replacing Systems#

When adding AI to a modernization project, start with a limited trial rather than an organization-wide plan. Instead of reducing concerns about financial viability to a debate over whether to adopt AI, measure which tasks deliver value for the money spent.

One possible approach is:

• Choose one area. Focus on a clearly defined problem, such as a process that takes too long to troubleshoot or a program that requires an impact assessment every time it changes.
• Record the current workload. Document the person-days needed to investigate dependencies, prepare design documents, and get ready for testing, along with current turnaround times and rework.
• Start by using AI for analysis and documentation. Defer changes to the runtime environment or production code, and use AI in ways that allow staff to verify its results.
• Run a trial lasting several weeks to a few months, comparing time spent, accuracy of findings, rework, and additional costs. Include not only usage fees but also the staff time required to check AI output.

The key is to measure improvements in business operations, rather than the volume of code generated. Even if investigation takes less time, overall costs will not fall if verification work increases or testing uncovers defects. Conversely, if AI reduces investigation time and rework for a particular task, and staff can reproduce those results, that provides a basis for expanding to the next area.

Set the evaluation criteria in advance as well. For example, record how much time is saved in checking dependencies, how many critical processes AI misses, and whether the trial is cost-effective after accounting for labor costs. Numerical targets will vary by system and organization, so use your own current performance as the baseline rather than applying someone else’s success story directly.

Tomorrow’s Legacy Systems Will Need Reviewing Too#

Systems built today will become legacy systems as they remain in operation. Even if AI helps write code faster, future staff will face the same investigative burden unless the reasons for each process and the business operations that depend on it are documented. When using AI for modernization, maintain the rationale behind decisions and the change history alongside the generated code.

Modernization is also about more than eliminating old technology. According to Kyndryl’s report, companies’ motivations for modernization now extend beyond cost reduction and moving away from legacy systems. As AI drives greater demand for IT, making environments that mix old and new systems work effectively is becoming another challenge. Sovereignty concerns also influence technology choices.

That is why people must still decide what to retain, what to change, and which business processes to tackle first. AI may reduce the burden of examining code and dependencies, but it does not automatically determine the value of an investment or the level of acceptable risk.

Rather than assuming that adopting AI will solve legacy problems, begin by measuring the current effort and cost of a single business process. Define what AI will handle and what people will verify, then use the trial results to decide whether to move on to the next modernization target.

Source: https://japan.zdnet.com/article/35253404/