Enterprises are building the plane before they have a fleet
A new wave of VentureBeat Pulse Research, based on a June 2026 survey of 101 enterprise organizations (100+ employees), reveals a stark tension: companies are racing to build sophisticated orchestration layers for AI agents, but the vast majority of those so-called agents are still glorified chatbots.
The headline finding is blunt. When asked to honestly assess their own portfolios, 71% of respondents said a quarter or fewer of their deployed AI agents are true multi-step orchestrated workflows. Only 10% have crossed the halfway mark. The rest are single-prompt chatbot wrappers — dressed up in agent clothing, but doing nothing an orchestration layer is actually for.
This gap between ambition and reality is the central story of the research. Enterprises are standardizing on model-provider platforms, pouring money into workflow tooling, and designing hybrid control planes — all before most of their agents can execute a multi-step task reliably.
Anthropic’s Claude dominates the platform race
When it comes to which platform enterprises are betting on, one name stands out. Anthropic’s Claude is the primary orchestration platform for 40% of respondents — more than double the next contender. Microsoft sits at 18%, OpenAI at 13%, and Google and Amazon trail in single digits. Open-source frameworks like LangChain and LangGraph, which dominate technical discussions, barely register in enterprise deployment.
The logic behind the choice is what the researchers call “model gravity.” The single biggest factor driving platform selection — cited by 21% of respondents — is native alignment with a state-of-the-art base model. Enterprises are picking the orchestration environment that comes with the frontier model they already want to build on.
But satisfaction is lukewarm. Respondents rated their platforms at 3.94 out of 5 overall, with “ease of implementation” the weakest score at 3.85. And 96% plan to change their orchestration approach within the year. These are tools enterprises tolerate, not love.
Reliability rules — but most agents can’t deliver it
What do enterprises actually want from orchestration? The answer is boring but brutal: reliability. Task completion reliability (32%) and multi-step workflow management (28%) together account for 59% of primary success metrics. Developer productivity and end-user experience lag far behind.
This makes the chatbot trap even more pointed. Enterprises define success as dependable multi-step execution, yet most of their deployed agents can’t do multi-step work at all. The ambition is real; the portfolio is not.
The trap is unevenly distributed. Among smaller enterprises (under 2,500 employees), 77% say a quarter or fewer of their agents do true multi-step work. For larger organizations, that figure drops to 62% — still high, but meaningfully better. The chatbot trap is, directionally, a mid-market condition.
Hybrid control planes: The hedge against lock-in
Enterprises are designing their control architecture with one fear in mind: vendor lock-in. By the end of 2026, 51% expect a hybrid control plane — part provider-native, part external. Only 6% plan to hand control entirely to a provider-managed service.
The reason is clear. When asked what worries them most about letting control live inside a model provider, 35% said vendor lock-in, up from 24% in an earlier April-May wave. Security and permissioning limitations (28%) and inflexibility across models (21%) round out the concerns.
This is a notable shift. In the earlier survey, security was the top concern. By June, lock-in had taken the lead. The worry about provider platforms appears to be maturing from whether they can be secured to whether they can be replaced.
The hybrid control plane is the architectural hedge. Enterprises will build on a provider’s platform, but they will not be governed entirely by it.
Investment flows to tooling, but cost control lags
Where is the money going? Agent workflow tooling leads spending plans at 34%, followed by security and permissions enforcement at 25%, and scaling infrastructure at 20%. Monitoring and debugging draws a smaller 11%.
The weight on tooling and permissions over pure observability signals that enterprises are spending to build and harden orchestration, not merely to watch it run.
But fiscal control over token consumption remains reactive. More than a quarter of enterprises (27%) admit they have no real-time, programmatic way to stop a runaway agent before the bill arrives — they learn of it from the logs afterward. Another 32% rely entirely on native caps built into their platform, a control only as good as the provider’s tooling.
Only the enterprises building custom gateways (23%) or exploiting cross-model routing to arbitrage cost (19%) are treating token burn as an engineering problem to be controlled deterministically.
Again, size matters. About one in three smaller enterprises (34%) exercises only reactive control of agent spend, against 20% of larger ones. The mid-market is running the least mature agents on the least instrumented budgets.
The bottom line: The layer is real; most of the agents aren’t yet
This wave of research paints a clear directional picture. Enterprises have decided how they want to orchestrate agents — on model-provider platforms, with hybrid control planes, judged by reliable multi-step execution. The platforms, budgets, and strategies are being put in place.
But the deployed reality is thin. Seventy-one percent of enterprises admit a quarter or fewer of their agents are genuinely orchestrated. Only 10% are past the halfway mark. And more than a quarter cannot stop a runaway agent in real time.
The orchestration layer is being built ahead of the orchestrated portfolio it is meant to run. That is not necessarily a contradiction — it may be a roadmap. The question for subsequent waves is whether the deployed reality closes the gap on the ambition, or whether the chatbot trap proves stickier than the roadmap assumes.
For organizations serious about AI agent deployment, the takeaway is sobering: invest in the orchestration layer by all means, but be honest about what your agents can actually do. And if you can’t stop a runaway agent in real time, fix that before you let it run unsupervised.
