Enterprise organizations deploying AI agents face a critical gap between governance capabilities and cost visibility, according to new research spanning 107 companies.
The typical enterprise runs three orchestration platforms simultaneously, selecting them for model flexibility rather than loyalty to any single vendor. Microsoft leads in primary usage while Anthropic dominates forward-looking plans by a significant margin. This multi-platform approach reflects enterprise demand for vendor independence alongside access to leading AI providers.
Cost control emerges as the thorniest problem. One in five enterprises lacks real-time mechanisms to halt runaway agents before expenses spiral. This blind spot persists despite organizations reporting stronger governance frameworks for agent behavior and security protocols.
Enterprises prefer deliberately hybrid control planes that combine leading AI providers with provider-independent technologies. Counterintuitively, companies fear provider-resident controls less for lock-in concerns and more for inadequate security and permissioning features built into those platforms. Organizations want flexibility to mix vendors without sacrificing oversight.
The research reveals a maturity mismatch. Governance tools have evolved to catch problematic agent actions. Cost metering has not. Enterprises can monitor what agents do but struggle to track what those actions cost in real time. This creates operational risk where well-behaved agents still generate unexpected bills.
The multi-platform orchestration pattern reflects rational hedging. Single-vendor orchestration would simplify cost tracking and governance, but enterprises prioritize flexibility across models and providers. They accept operational complexity as the price of avoiding vendor lock-in.
Organizations deploying production agents at scale need immediate attention to cost instrumentation. Robust governance without cost visibility leaves enterprises exposed to budget surprises. The research suggests most enterprises have solved the "what is the agent doing" problem but remain stuck on "what is it costing."