Why AI Rollouts Stall: The Transformation Debt Nobody Budgets For

An enterprise can run a full cloud ERP rollout and still watch its planners reconcile data in spreadsheets. That gap — between technology that is technically live and technology that actually changes how the business runs — is the argument at the centre of a new analysis by Srinivasan Narayanan, an Oracle Cloud Solution Delivery Lead and IEEE senior member who has guided large manufacturers through ERP-driven supply chain transformations across multi-plant and multi-distribution-centre environments.

His core claim is blunt. If every new technology wave — digital transformation, cloud, advanced analytics, now generative AI and autonomous agents — is being asked to fix the same problems, the constraint isn't technology. It's the organization's ability to absorb transformation.

What does 'operational absorption' actually mean?

Narayanan draws a hard line between two things executives often blur. Adoption is the ability to acquire and implement technology. Operational absorption is the ability to integrate that technology into processes, decisions, data, governance and daily ways of working, so it becomes part of how the business operates.

The distinction matters because technology cycles are compressing. Companies can no longer finish one transformation, run in a stable environment for years, then prep for the next. New capabilities arrive while earlier rollouts are still stabilizing. The winners, he argues, won't be the best implementers — they'll be the organizations best at continuously absorbing change.

A concrete test: check whether the last transformation changed daily work. If planners still reconcile by hand, or key decisions still happen outside the system after a cloud ERP go-live, ask which problems the rollout was supposed to solve before buying another tool.

Why the same problems keep returning

Many transformation barriers aren't technology problems at all. They sit in poorly designed processes, inconsistent data, unclear ownership, unnecessary customization, fragmented decision-making and a culture that lets temporary workarounds harden into permanent practice.

Narayanan calls the accumulated mess "operating model debt." It builds the way technical debt does. A business exception triggers a customization. A process gap spawns a spreadsheet. An urgent requirement produces a point-to-point integration. A temporary approval becomes permanent. Each call looks reasonable alone. Together they make every future change harder, because each new capability must thread through existing complexity, custom integrations and entrenched workarounds.

This bites hardest with AI. An agent working across planning, procurement, manufacturing, inventory, maintenance or logistics depends entirely on the processes, data, decision rules and governance underneath it. Weak foundations don't get fixed by adding intelligence on top. Worse, automation can amplify the weaknesses, because decisions and actions now happen faster and with less human intervention.

Fix the sequence: business problem first

The predictable failure mode, he writes, is reversing the sequence — seeing a new capability, then hunting for places to use it. Every investment should open with two questions:

• What business problem are we trying to solve?
• What measurable value will it deliver?

And a third, once you're on your second or third wave: why are we still solving this problem? If the cloud program was supposed to kill fragmented processes and manual reconciliation, AI should build on that foundation — flagging supply risks earlier, improving planning decisions, autonomously managing well-defined exceptions — not papering over fragmentation the last project failed to eliminate.

Three shifts that build absorption capacity

Narayanan prescribes three moves:

• Modernize the digital core without recreating legacy complexity. Every customization should be judged not just by the problem it solves today but by the future complexity it creates. ERP, planning, manufacturing, logistics, maintenance and data platforms should run as one connected core, not independent projects.
• Move from better information to better decisions. Define decision rights before granting AI autonomy: which calls stay human, where AI recommends, what it may execute, when it must escalate, and who owns the outcome.
• Build capabilities that outlive the technology. Firms need people who understand processes end to end, can interpret data, challenge recommendations and redesign workflows — not experts in every emerging model. Governance, likewise, needs repeatable mechanisms for value assessment, risk management and retiring what no longer earns its keep, rather than a new model per wave.

Measure speed to value, not speed to deployment

Go-live is the beginning of value realization, not the end, he writes. The metrics that count:

• Time from deployment to stable business operation
• Percentage of transactions on standardized processes
• Reduction in manual workarounds and human-handled exceptions
• Time to integrate each new capability into core processes
• Measurable value realized after deployment

A project can land on time and on budget and still have changed nothing about how the business runs.

The executives best prepared for the next era won't be the ones who adopt every technology first. They'll be the ones whose processes, data, governance and workforce get stronger with each wave — because every transformation, Narayanan closes, should make the next one easier.