Arthur D. Little’s Blue Shift Institute has released AI’s Hidden Dependencies, a report examining the infrastructure systems underpinning enterprise AI adoption. The research highlights growing reliance on energy availability, compute capacity, and supply chains—dependencies that many organizations have not fully priced into their AI strategies.
In a follow-up interview with ERP Today, Dr. Albert Meige, global director of the Blue Shift Institute, explained why those dependencies create risks for enterprises embedding AI into ERP systems, where AI is becoming embedded in always-on, business-critical operations.
When AI Dependency Becomes Systemic Risk
Meige said most enterprises still assess AI risk too narrowly. Attention remains focused on downstream concerns such as model accuracy or bias, while upstream dependencies receive far less scrutiny.
“The majority of businesses don’t currently measure the systemic risks around access to AI infrastructure – instead, they are solely looking at downstream risks such as inaccurate or biased results from AI models,” Meige said.
That gap matters as AI moves from isolated deployments into core operational systems.
According to Meige, AI dependency crosses into systemic risk when infrastructure constraints begin to affect core operations. At that point, organizations lose the ability to absorb volatility without operational impact.
“Early warning signs to monitor include exponentially growing costs, a lack of resource availability in particular areas, such as compute power, and supply chain consolidation that limits flexibility,” he said.
Those signals tend to surface first outside traditional risk frameworks.
Meige noted that enterprises often recognize infrastructure constraints only after AI is already embedded in business-critical systems, leaving limited room to adjust architectures, contracts, or sourcing strategies.
Inference Economics Reshape AI Costs Inside ERP Systems
The report identifies a structural shift reshaping AI economics. Inference, rather than training, is emerging as the dominant source of energy use and long-term cost. That transition has direct implications for ERP environments, where AI increasingly operates continuously rather than in experimental bursts.
“At a basic level it will push up day-to-day operational costs, particularly as inference-heavy agentic and multimodal AI gain traction, as these can exponentially shift required budgets,” Meige said.
ERP systems amplify that exposure. Forecasting, planning, compliance monitoring, and close processes run continuously, embedding AI into workflows that cannot be paused or throttled easily.
“The other factor for businesses that use ERP systems is that they don’t have control over the level of AI (and the costs) implemented within their chosen solution or where compute workloads run,” Meige said.
When ERP platforms act as systems of record, cost volatility or infrastructure constraints propagate quickly across finance, supply chain, and operations, raising the stakes of architectural decisions made upstream by vendors.
AI Supply-Chain Concentration and Infrastructure Bottlenecks
The report also highlights structural concentration across the AI value chain. While application ecosystems remain relatively diverse, upstream layers show consolidation.
Chips, advanced lithography, and hyperscale cloud platforms represent the most concentrated—and fragile—segments of the stack. According to Meige, those layers are also where disruption is most likely to occur as AI demand accelerates.
“Given the rapid growth of AI, many enterprises are exposed to top-level bottlenecks, which are the weakest links,” he said.
Enterprises are not standing still. Meige said organizations are already responding by diversifying sourcing strategies and rethinking deployment models.
“They are working to build resilience through more diversified sourcing and hybrid architectures that mix on-premise, sovereign, and public environments,” he said.
Those efforts, however, cannot fully offset upstream concentration. They instead buy flexibility and time when upstream bottlenecks materialize.
AI Infrastructure Risks Elevate the Importance of ERP Governance
AI can amplify operational value in ERP environments because it is embedded into systems of record, but that also concentrates cost, access, or vendor risk exposure.
Meige described this as a compounded dependency. Organizations rely on AI infrastructure providers, and ERP customers are further exposed because vendors control how AI is implemented.
The governance burden increases as a result. Meige said ERP users should prioritize what the report describes as “no-regret” actions—moves that strengthen resilience regardless of how infrastructure risks unfold.
Those actions include gaining visibility into the real footprint of AI usage, keeping costs predictable, and preserving the freedom to adapt across providers and jurisdictions. Because cost pressures and access constraints will not surface uniformly, forecasting alone is unlikely to be sufficient.
“The only viable approach is to consider different AI growth scenarios at a board level, monitor the market and look at the ‘no-regret’ actions that can be taken to build efficiency, adaptability, and resilience,” Meige said.
What This Means for ERP Insiders
AI infrastructure risk is a GRC issue. AI exposure increasingly sits outside traditional IT risk registers, yet it affects financial controls, operational continuity, and third-party risk simultaneously. GRC frameworks must expand beyond model risk to capture infrastructure dependency as a cross-domain control concern.
ERP concentrates AI risk into audit-critical systems. Embedding AI into systems of record ties infrastructure volatility directly to financial reporting, compliance workflows, and controls assurance. GRC leaders must reassess how vendor roadmaps and compute dependencies affect auditability, resilience, and accountability.
Scenario planning complements traditional controls. AI cost and access shocks will surface unevenly across regions and vendors, limiting the effectiveness of fixed compliance checklists alone. GRC programs benefit from adding scenario analysis, adaptive controls, and clear decision rights.
