The Hype Gave Way, and the Complexity Showed Up
Aiman Ezzat, CEO of Capgemini, offered a pointed assessment of where enterprise AI actually stands. Speaking publicly this week, Ezzat said clients are realizing AI is far more complex than they were initially led to believe.
His summary: "Quick savings, some agents, a small platform" was the early pitch. That framing is giving way to something harder.
For ERP owners and finance and operations leaders, that shift is directly relevant. Your AI roadmap almost certainly started with a use case and a data problem. Ezzat says the actual stack of decisions runs deeper than that.
Four Layers That Compound
Ezzat described a sequence of realizations clients are working through. First, it was data readiness. Then it became data context and schema design. Then the control plane — how AI decisions are governed and sequenced across a process. Then model selection: which decisions call for a large language model (LLM) and which are better handled by a smaller, purpose-built model (SLM).
And underneath all of it, data sovereignty — where the model runs, which jurisdiction governs it, and who controls the weights and logs.
None of these decisions is independent. A control plane choice affects which models are practical. A sovereignty constraint shapes which cloud infrastructure is even available. The complexity of optimizing multiple variables simultaneously is what Ezzat says is driving clients back toward large, multi-year transformation programs rather than incremental agent deployments.
His argument is counterintuitive but worth examining: the variable count for deploying a single agent into one process is approaching the variable count for an end-to-end redesign. If that is true, scoping narrow to reduce risk may not actually reduce it.
What This Means for Your ERP Roadmap
Capgemini's active engagements give concrete shape to Ezzat's argument. For HMRC in the UK, Capgemini is migrating a critical tax platform to SAP S/4HANA on a sovereign cloud, explicitly framed as a foundation for future AI capabilities.
For a major North American automotive manufacturer, Capgemini is redesigning enterprise shared services across finance, HR, procurement, supply chain, and IT through an AI-enabled global business services model, with projected productivity improvement of 50% to 54% over seven years.
Those are not AI pilots layered onto existing ERP workflows. They are process redesigns where the ERP modernization and the AI architecture are planned together.
For mid-market operations leaders, the practical implication is a sequencing question. An AI initiative that assumes the ERP data layer is ready — without confirming data context, schema quality, and control logic — is building on the same unstable foundation Ezzat describes.
The ROI Discipline Is Not Going Away
Ezzat warned that without a disciplined approach to value capture, AI spend will not yield visible returns. Clients are still investing, but capital is moving toward initiatives with measurable outcomes. Isolated pilots are being deprioritized.
That shift affects vendor relationships too. If your ERP vendor's AI roadmap is still framed around individual use cases rather than process-level redesign, that gap is worth surfacing in your next roadmap review.
The Audit Question for Your Next Planning Cycle
Before your organization commits another budget cycle to AI, work through Ezzat's four layers as a checklist. Is the data context and schema defined — not just the data source? Is the control plane designed for the process, or assumed to be a standard platform default? Has model selection been evaluated against the actual decision types in each workflow? And has sovereignty been scoped to the jurisdictions and infrastructure constraints your organization actually operates under?
If any of those questions does not have a named owner and a current answer, the complexity Ezzat describes is already in your program. It just has not surfaced yet.