Agentic AI’s modernization promise faces a production reality check

Vendor claims are moving faster than production adoption

Modernization vendors are selling AI as a coordination layer for work that has historically been fragmented across application teams, infrastructure groups, security reviewers, finance functions and outside systems integrators. Accenture Edge’s AWS offerings, for example, include an agent-led data discovery accelerator intended to provide visibility into data sources and dependencies during cloud migration and modernization projects, as well as AI-powered migration offerings for virtual machines.³

Those offerings reflect a plausible technical use case. Dependency discovery is one of the hardest parts of legacy modernization because old systems often contain undocumented interfaces, brittle batch jobs, embedded business logic and unowned data flows. Agents that inspect estates, generate documentation, recommend migration paths and compare outputs could make programs more coherent.

But Kyndryl’s 10% production figure suggests that most enterprises are not yet operating at that level of confidence.¹ Many are likely still experimenting, limiting agents to advisory tasks, or using AI within developer workflows rather than allowing it to orchestrate modernization programs.

That distinction matters. A proof of concept that generates a dependency map is not the same as a production operating model that changes delivery economics across a multi-year transformation.

The economics case is still unresolved

The modernization business case typically rests on several expected returns: lower maintenance costs, faster product delivery, reduced operational risk, better resilience, improved data access and fewer dependencies on scarce legacy skills. Agentic AI could contribute to those outcomes if it shortens discovery, reduces manual code analysis, improves test coverage or helps teams prioritize what to modernize first.

Yet the Kyndryl data points to a market where value realization remains uneven. If nearly half of programs are still delayed and over budget, and nearly one in five organizations reports limited or unclear value, AI adoption should be evaluated against those baselines rather than against vendor demos.¹ CIOs should ask whether agents improve the full program economics or merely add another tooling cost to an already complex transformation portfolio.

The same issue is visible in adjacent technology finance discussions. Yarken’s analysis of Forrester’s FinOps observations notes that autonomous agents may be able to identify cloud cost actions, but a recommendation that looks financially sound can be wrong without business context, contractual awareness, ownership data and approval rules.²

Modernization has the same problem at a larger scale. An agent may identify a service as a candidate for retirement, refactoring or migration, but the decision may depend on compliance commitments, downstream reporting, batch windows, customer commitments or a product roadmap that the model cannot infer from code alone.

In that sense, the economics case for agentic modernization depends on context integration as much as model capability. The agent must understand not only what exists in the estate, but also what the business can safely change.

Controls are becoming part of the adoption bottleneck

Enterprise deployment of agents is also constrained by governance and trust. RTInsights’ October 11 roundup highlighted new work on compliance standards for autonomous agents, including a proposed framework for showing that agents follow established rules without exposing sensitive data.⁴ The same roundup pointed to production AI infrastructure and governance capabilities designed to run agents with enterprise controls.⁴

That governance push is not incidental. Modernization agents may need access to source code, infrastructure metadata, configuration files, vulnerability data, application logs, data catalogs and business process documentation. They may also recommend changes that affect system availability or security posture. The more consequential the action, the stronger the need for evidence, traceability and human approval.

The Information Technology and Innovation Foundation made a similar point in comments to the National Institute of Standards and Technology on modernizing the National Vulnerability Database for the AI era. It argued that AI-driven prioritization should be explainable, source-traceable and auditable, with records that allow analysts to reconstruct and review automated decisions.⁵

While the filing concerns vulnerability management rather than application modernization, the principle carries over: automation can accelerate decision-making only if teams can validate what the system did, why it did it and whether it was appropriate for the environment.

For modernization leaders, this turns agent governance into a delivery prerequisite. Before agents can move from copilots to execution systems, organizations need clear boundaries: what agents may observe, what they may recommend, what they may change, when approval is required and how every decision is logged.

Industry narratives are strongest where legacy pain is highest

The most aggressive claims for agentic modernization are appearing in sectors with heavy legacy burdens, including insurance. In-surely’s October 11 analysis argues that agentic systems could help brokers and carriers decompose legacy code, propose migration paths and validate outputs continuously, while reducing reliance on scarce mainframe talent.⁶ It also frames delay as a competitive cost, citing pressures such as digital-native competitors, rising maintenance costs and talent drain.⁶

That narrative will resonate with CIOs in industries where core systems have become business constraints. In banking, insurance, government, logistics and healthcare, legacy applications often limit product speed and data availability while consuming disproportionate maintenance budgets. If agentic AI can reduce the cost of understanding and changing those systems, the payoff could be significant.

But the same industry examples also show why adoption is likely to be incremental. In-surely describes a realistic path as modernizing workflows around the core first, then migrating the core itself, measured by outcomes such as faster submissions, cleaner data and lower operating cost.⁶ That is more credible than a wholesale autonomous rewrite of critical platforms. It also aligns with the Kyndryl data: production use exists, but it remains narrow.

What CIOs should require before scaling

For CIOs and engineering leaders, the practical response is to treat agentic modernization as a measured operating change rather than a procurement shortcut. The first requirement is a baseline. Leaders should know current modernization cycle times, defect rates, rework levels, documentation coverage, dependency discovery costs and run-rate maintenance spending before attributing improvements to AI.

Second, teams should separate advisory use from execution use. Agents that summarize code, draft documentation or identify dependencies carry different risks from agents that alter infrastructure, generate migration scripts or refactor production code. Each step up in autonomy should require stronger controls, test evidence and accountable human ownership.

Third, modernization programs should connect agent recommendations to business and financial context. As the FinOps example shows, technical optimization without business awareness can create new risk.² A modernization agent needs access to application ownership, service criticality, compliance requirements, product dependencies and cost models if its recommendations are to be economically meaningful.

Finally, leaders should demand value evidence from vendors that maps to modernization outcomes, not generic AI productivity. The relevant measures are fewer missed dependencies, faster conversion of specific code classes, reduced manual documentation effort, lower test remediation effort, shorter migration windows, fewer incidents and clearer post-program savings.

Agentic AI may still become an important modernization tool. The Kyndryl findings do not invalidate the technology’s potential. They do, however, show that the market is early, the production base is limited and the economics case remains unresolved.

For now, the burden of proof sits with vendors and adopters alike: show that agents can improve modernization outcomes, not just participate in modernization activity.