Amazon's Q2 earnings reveal that enterprise AI deployment has shifted decisively from model capability to infrastructure reliability. AWS reported a $25 billion AI revenue run rate growing at triple-digit percentages, but the substantive signal lies in how Amazon is positioning the problem: production agents require secure infrastructure, persistent memory, identity controls, data connections, and observability—not just polished interfaces. Andy Jassy framed this explicitly as a production challenge rather than a model race, which resets expectations for CX teams currently evaluating agentic platforms. The implication is stark. An agent that summarizes conversations or suggests replies in a sandbox differs fundamentally from one that identifies customers, retrieves policies, triggers back-office processes, records outcomes, and escalates safely. This distinction matters because it exposes a procurement gap: many CX leaders are still asking "does this tool have AI?" when they should be asking "can it safely orchestrate work across our existing systems?" Amazon's answer—Bedrock AgentCore and the broader infrastructure layer—suggests that fragmented point solutions will struggle to justify their place when enterprises need consistent controls across channels, particularly as agents gain access to sensitive customer data and payment processes.
The economics of agentic AI will prove equally consequential for CX roadmaps. Amazon emphasised that inference costs, tool use, and reinforcement learning run primarily on CPUs rather than accelerators, positioning its Graviton chips as offering 30-40% better price-performance. Yet this optimism requires scrutiny. Gartner's Patrick Quinlan warned that organisations underestimating operational costs will face disappointment—agentic AI introduces new expenses tied to usage, compute, governance, integration, and specialist skills. For CX teams, this means cost per resolution must become a primary metric alongside containment rate and handle time. A cheaper model delivering poor outcomes or pushing complex work back to human agents is not a cost saving; it is a cost transfer. Security compounds this discipline requirement. Amazon highlighted that security now dominates enterprise AI conversations, and for good reason: customer-facing agents accessing sensitive records, policies, and payment processes create trust challenges that basic chatbots do not. The strongest deployments will grant agents sufficient context to resolve real problems whilst restricting visibility and action authority. For CX professionals, this signals that platform selection should prioritise those with deep infrastructure, data governance, and security capabilities—and that the operational burden of running agents safely may outweigh the appeal of cheaper models or point solutions that lack integrated controls.
Amazon’s Q2 Earnings Expose the Hidden AI Cost Crisis Crippling Enterprise CX CX Today