Why is AI making enterprise modernisation more complex, not less?

Who answered and what was measured

Kyndryl commissioned the research firm Coleman Parkes to survey 2,000 senior IT decision-makers and line-of-business leaders in mid-market and enterprise organisations. Respondents' employers averaged $3.8 billion in global revenue. The regional composition was 35% Europe, 30% Asia-Pacific, 25% North America and 10% Latin America. Twelve sectors were represented, including technology, banking, manufacturing, healthcare, retail, telecommunications and media, energy, transport and government. Roles included CIOs, CTOs, heads of IT and leaders in operations and management.

The report supplies a denominator beside many charts and distinguishes the full sample from two 1,000-person modules focused on mainframe and cloud or on applications and networks. It also identifies narrower bases, such as 391 respondents whose organisations had moved workloads back to mainframes and 192 actively using agentic AI across several modernisation tasks. Those details help prevent percentages from being mistaken for findings about all 2,000 participants.

Important information is still missing from the published methodology. It does not state the survey's field dates, how respondents were recruited, country-level quotas, response rate, weighting, margin of sampling error or the exact screening questions. The five-continent description in the announcement is broader than the four regional groupings specified in the report. Results should therefore be read as the reported experience of this selected senior-leader sample, not a census of global businesses or a measure of actual system performance.[1]

The biggest constraint may be knowing what is already connected

Respondents in the application-and-network module reported an average of 383 business-critical applications. Only 9% said they had a fully mapped, current graph of application dependencies, while 54% described their understanding before modernisation as partial or limited. One quarter said they had many undocumented or 'dark' applications. This is consequential because a model, data pipeline or security change can fail through an old interface that a project team did not know existed.

The same visibility problem complicates cost claims. Nearly half, 46%, described predicting and controlling cloud AI workload costs as very challenging. Across the full sample, 18% said past modernisation had delivered limited or unclear value; 53% selected reliability and resilience as a major source of realised value, compared with 35% selecting better user or customer experience. These are multiple-choice self-reports, not audited savings or customer-outcome measures, but they challenge the idea that infrastructure investment automatically becomes visible business transformation.[1]

Agentic AI shows an association, not a proven shortcut

Ten per cent of organisations had put agentic AI into production across several modernisation use cases. Among active users, common tasks included automated testing, cited by 43%, legacy-estate discovery and assessment at 39%, documentation and code refactoring at 31% each, and code translation at 27%. The appeal is plausible: software agents can search configuration data, code and operational records more quickly than teams manually assembling a dependency map.

Early adopters were less likely to say they were behind schedule: 33%, compared with 49% among non-users. Among the 192 active multi-use adopters, 33% reported clear return on investment and plans to expand, while another 39% reported early benefits that were not yet fully measurable. The report appropriately calls the results promising but unproven at scale. Organisations able to deploy agents may already have cleaner data, stronger engineering teams, more money or better governance. The cross-sectional survey does not randomly assign the technology or observe results over time, so it cannot isolate an agentic-AI effect.

Reported barriers reinforce that caution. Security posture was selected by 33% of respondents considering agentic deployment, integration with existing toolchains by 32%, lack of trust in AI decisions by 30%, and governance or audit requirements by 30%. An agent that can modify code or navigate operational systems expands the importance of permissions, test coverage, rollback and human accountability. Faster activity is not the same as safer modernisation.[1][2]

What this changes for organisations and workers

For employees and customers, the finding is less about abstract architecture than service reliability. Banks, hospitals, retailers and public bodies can run critical applications across several generations of technology. If AI adds new network, data and security dependencies without revealing old ones, outages and compliance failures can reach people through missed payments, delayed care, inaccessible services or exposed records. Teams need inventories that connect technical components to the business services and groups who rely on them.

The report also finds that decision authority is dispersed: respondents selected an average of three groups with budgetary or decision-making control, while nine stakeholder groups each received substantial influence rankings. That can slow approval, but it also reflects legitimate interests in cost, safety, law, workforce design and customer outcomes. A useful modernisation measure should therefore combine operational indicators such as incidents and deployment time with energy use, staff workload, accessibility, security and user outcomes. A single migration count or promised AI saving is not enough.

Geography matters too. Sixty-three per cent expected sovereign-cloud use to increase, yet 65% said sovereignty decisions were addressed selectively or mainly through compliance and only 6% anchored broader transformation in sovereignty principles from the outset. Governments differ on data location, access, procurement and cross-border transfers. An architecture that is efficient in one jurisdiction may create legal or operational lock-in in another, particularly for multinational organisations and public services.[1]

Commercial context and what would change the assessment

Kyndryl is an enterprise technology services provider that sells consulting, implementation and managed modernisation work. It funded the study and frames orchestration, continuous modernisation and agentic tools as the answer to complexity. Coleman Parkes conducted the survey, but the public report does not disclose an independent academic protocol or preregistration. That commercial interest does not make the responses false; it makes independent validation and transparent methods especially important before treating the report as proof that a specific service or architecture works.

Confidence would rise if Kyndryl published the questionnaire, field dates, country and company-size quotas, recruitment and weighting methods, anonymised response tables and uncertainty intervals. Repeated panels could test whether the same organisations that adopt agentic tools later improve delivery speed, costs, incident rates and user outcomes. Independent studies should compare like-for-like projects, including failed or cancelled deployments, and verify claimed returns against operational and financial records rather than leader perceptions alone.

The present evidence supports a bounded conclusion. Large organisations in this survey expect AI to increase the breadth and coordination demands of their technology estates, while many report weak visibility into application dependencies. It does not show that every company needs more platforms, that postponement alone caused the reported harms, or that agentic AI will solve the complexity it also helps create. The practical priority is to map systems, owners, risks and human outcomes before accelerating change—and then measure whether modernisation makes services demonstrably better.[1][2]