Enterprise AI orchestration builds infrastructure for agents that aren't there yet
A survey of 101 enterprises with 100 or more employees paints a consistent picture: the orchestration layer is being built before the orchestrated portfolio exists. Anthropic's Claude holds 40% of primary platform usage, more than double any rival. Enterprises are investing in workflow tooling, planning hybrid control planes to avoid vendor lock-in, and 68% expect to switch orchestration platforms within twelve months. Yet 71% admit a quarter or fewer of their deployed agents are genuine multi-step workflows. Most of what runs under the agent label is still a chatbot wrapper. The infrastructure race is running well ahead of the reality it was meant to govern.
The platform concentration has a straightforward explanation. Model gravity leads the selection factors at 21%, meaning enterprises pick the orchestration environment attached to the frontier model they have already standardized on. Claude's 40% share is the result of that gravitational pull, not a platform victory on its own terms. The next two factors, flexibility across models and tools (17%) and ease of development (17%), show that buyers are also making the choice with an exit in mind. Vendor lock-in is the risk enterprises name most often when asked what worries them about provider-resident control (35%), ahead of security and permissioning limitations (28%). The concentration is real but clearly read as provisional by the same buyers creating it.
The chatbot trap surfaces directly in the self-assessment. Asked what share of their deployed agents are true multi-step orchestrated workflows versus single-prompt wrappers, 71% of respondents put themselves in the bottom two bands: a quarter or fewer. Only 10% have crossed the halfway mark. This gap is not a contradiction of the earlier findings. It is their consequence. Enterprises are building the platforms, budgets, and control architectures precisely because the orchestrated portfolio is still thin. The orchestration layer is scaffolding for work that has not arrived yet.
The fiscal control gap compounds the maturity problem. More than a quarter of enterprises (27%) have no real-time, programmatic way to stop a runaway agent before the bill arrives. Another 32% rely entirely on the native caps built into their primary platform, which ties cost control to the same provider whose lock-in they are trying to avoid. The enterprises building custom gateways (23%) or routing across models to arbitrage cost (19%) are treating token consumption as an engineering problem. The majority are not. That 27% figure matters because it means more than a quarter of this cohort cannot exercise the most basic production control over a system they are actively moving toward deployment.
The switching intent is the other signal that separates this layer from the rest of the stack. Sixty-eight percent plan to adopt a new, additional, or replacement orchestration platform within twelve months, the highest intent-to-change of any layer VentureBeat tracks. Among those in motion, the largest single group (29%) has no shortlist. OpenAI leads named candidates at 16%, followed by LangChain/LangGraph at 12% and Anthropic at 7%. The independent frameworks draw roughly double their current usage footprint in forward consideration, which mirrors a pattern VentureBeat's security tracker found for specialist vendors. Enterprises are concentrated on model-provider platforms, broadly dissatisfied with the orchestration layer, planning to move, and mostly undecided on where to go. The most consolidated layer of the stack is also the least settled.
The sample is self-selected (n=101, single June 2026 wave), not a probability sample, and the methodology notes it should be read directionally rather than as a confirmed trend. That limitation is important. The findings are consistent with patterns visible in practitioner discussions and adjacent survey data, but the exact figures reflect a specific cohort of AI-active technical decision-makers, not enterprise AI deployment broadly. The directional signal is clear enough to act on: enterprises are building for orchestration maturity they have not yet reached, and the infrastructure gap between investment and deployment is worth tracking in subsequent waves.
The hybrid control plane expectation (51% by end of 2026) is the architectural conclusion from all of this. Enterprises want the model's gravity without the provider's governance. They want the platform without the lock-in. They are investing in tooling and permissions, consolidating frameworks, and moving agents from sandbox to production. But they are doing it while running mostly chatbots, with reactive cost controls, and without a clear successor platform. The orchestration infrastructure is a bet on a portfolio that has not materialized yet. Whether that bet pays off depends on whether the deployed reality closes the gap on the ambition before the next wave of platform selection arrives.