Enterprise AI organizations are moving quickly to build agentic systems—but the most important bottleneck isn’t the lack of orchestration platforms. It’s deployment maturity. A new VentureBeat Pulse Research survey of 101 enterprises (100+ employees) finds that companies are consolidating their agent orchestration bets onto major model-provider ecosystems, investing in workflow tooling and permissions, and planning hybrid control planes to reduce lock-in risk. Yet when respondents are asked to describe what they’ve actually deployed, a striking gap appears: most “agents” still behave like chatbot wrappers rather than true multi-step orchestrated workflows.
That mismatch—between orchestration ambition and orchestration reality—may be the defining story of enterprise agent adoption right now. And it has practical consequences for reliability, governance, and cost control. In other words: enterprises aren’t just struggling to “get agents working.” They’re struggling to get from conversational demos to production-grade execution loops that can safely run multi-step tasks end-to-end.
A consolidation wave is underway, but it’s not happening where engineers expect
One of the clearest signals in the survey is where orchestration is actually running today. Across the 101 enterprises, orchestration deployments concentrate heavily on model-provider platforms. Anthropic’s Claude leads by a wide margin at 40%—more than double the next platform. Microsoft follows at 18%, with OpenAI at 13%. Google and Amazon also appear in the mix, and together these major providers account for roughly 80% of deployments (81 of 101).
Meanwhile, the open-framework ecosystem that dominates many technical conversations—LangChain/LangGraph—and custom in-house builds remain marginal, each landing in single digits. A small 3% of respondents report that they are not orchestrating at all.
This pattern matters because it challenges a common assumption: that enterprises would gravitate toward orchestration frameworks first, then plug in models later. Instead, the survey suggests the opposite. The orchestration layer is being selected largely as an extension of the model strategy, not as an independent engineering choice.
Respondents rate their current orchestration platforms at 3.94 out of 5 overall (109 answered). “Value for money” scores essentially the same (3.94), while “ease of implementation” is the weakest at 3.85. That’s not a disaster score, but it’s also not enthusiasm. The more telling detail is that 96% of respondents plan to change their orchestration approach within the year. In other words, enterprises are tolerating today’s setup because it works well enough to move forward—but they’re actively searching for something better.
The driving force is “model gravity,” not orchestration features
When asked what influenced their orchestration platform choice, enterprises point to “model gravity”—the pull of the underlying base model and its native alignment with frontier capabilities. In the survey, model gravity is the single largest factor at 21%. Flexibility across models and tools and ease of development are close behind, each at 17%. Security and permissions enforcement account for 14%, and total cost of ownership is 11%. Performance (latency/memory) sits last at 4%.
This ordering is revealing. It implies that, for most enterprises, the orchestration platform is not primarily evaluated as a standalone product. It’s evaluated as the environment that best supports the model they want to standardize on, while still leaving enough room to avoid being trapped.
That lock-in anxiety shows up later in the report, but it’s already present in the selection logic. Enterprises want optionality, even if they start by consolidating around a provider.
What enterprises say they optimize for: reliable multi-step execution
If the platform story is about consolidation, the success story is about reliability. The survey asks what enterprises optimize for as their primary success metric for orchestration. Two categories dominate:
Task completion reliability at 32%
Multi-step workflow management at 28%
Together, those two metrics account for 59% of responses (60 of 101). Developer productivity is meaningful but secondary at 17%, and end-user experience is relatively minor at 9%.
This is a crucial point: enterprises define “agent success” less as a flashy conversation and more as dependable execution. They want systems that can carry a task through multiple steps and finish the job without falling apart midstream.
And that definition is exactly what makes the next finding so pointed.
The chatbot trap: most deployed “agents” aren’t doing real multi-step work
When respondents are asked to assess their own portfolios honestly—what share of their deployed “agents” are true multi-step orchestrated workflows versus single-prompt chatbot wrappers—the results expose the central gap of the report.
Combining the bottom two bands, 71% of enterprises (72 of 101) say a quarter or fewer of their deployed “agents” are genuinely orchestrated. Only 10% (10 of 101) report that more than half of their deployed agents meet the “true multi-step orchestrated workflow” bar.
In plain terms: most enterprises are calling their systems “agents,” but most of those systems are not yet behaving like agents in the operational sense. They are closer to assistants that answer prompts than orchestrators that execute multi-step plans.
The survey also suggests this trap is not evenly distributed. Smaller enterprises show a stronger version of the problem: 77% of smaller organizations say a quarter or fewer of their agents do true multi-step work, compared with 62% among larger enterprises. Larger organizations appear to be further along in genuine multi-step deployment, which fits the broader pattern of enterprise maturity: more resources, more integration work, more operational discipline.
But the headline remains: the orchestration layer is being built ahead of the orchestrated portfolio it’s meant to run.
This is not necessarily a contradiction—it may be a roadmap in disguise
It’s tempting to interpret the gap as failure: enterprises invested in orchestration infrastructure but didn’t deploy agents that use it. However, the survey’s framing suggests a different interpretation.
Enterprises are building orchestration platforms, budgets, and control architecture precisely because the orchestrated portfolio is still thin. In other words, the infrastructure is being prepared for the next phase of deployment, not retrofitted after the fact.
That said, the gap still has consequences. If most deployed “agents” are single-prompt wrappers, then the reliability metrics enterprises claim to care about are not yet being tested at scale. Multi-step reliability is hard to achieve, and it requires more than a model and a prompt. It requires workflow tooling, state management, guardrails, and operational controls that can handle failures gracefully.
Those controls are where investment is going next.
Investment priorities: workflow tooling and permissions lead, monitoring trails
The survey asks what orchestration-related investments will grow most over the next year. The top spending category is agent workflow tooling at 34%. Security and permissions enforcement follows at 25%. Scaling infrastructure is next at 20%. Monitoring and debugging draws 11%, and another 11% report flat budgets.
This allocation reinforces the reliability-first narrative. Enterprises are funding the machinery that strings steps together dependably, and they’re funding the access control and permissioning needed to make those steps safe. Monitoring is important, but it’s not the main budget driver—suggesting that many organizations still see orchestration as something they must harden before they can fully observe and optimize it.
There’s also a subtle implication here: if your deployed agents are mostly chatbot wrappers, you may not yet have the operational complexity that demands heavy observability spend. But as soon as you move into multi-step execution, monitoring becomes non-negotiable. The budget pattern suggests enterprises are trying to get the workflow engine working first, then instrument it more deeply as usage expands.
Control plane expectations: hybrid is the default, not the exception
By the end of 2026, enterprises expect the primary control plane for agents to be hybrid. In the survey, 51% anticipate a hybrid model—provider-native plus external orchestration. Only 6% expect to hand control to a provider-managed service outright.
When respondents explain why, vendor lock-in emerges as the dominant concern. If control sits inside a model-provider platform, vendor lock-in is the top worry at 35%. Security and permissioning limitations follow at 28%, and inflexibility across models and tools is at 21%.
This is a meaningful shift in emphasis. The survey notes that in an earlier April–May wave, security and permissioning limitations were the leading concern (32%), with lock-in second (24%). By June, lock-in had moved to the top. The worry appears to be maturing: enterprises are increasingly focused not only on whether provider platforms can be secured, but on whether they can be replaced.
Hybrid control planes are therefore not just a technical preference. They’re a governance strategy. Enterprises want to benefit from provider-native capabilities while keeping the ability to route, constrain, and potentially migrate execution logic outside the provider’s control.
Three strategic moves cluster for the year ahead
The survey also asks what major changes enterprises anticipate in their orchestration strategy over the next 12 months. Three moves cluster near the top, almost evenly split:
Building in-house control (25%)
Standardizing on one framework (24%)
Moving agents from sandbox to production (23%)
Only 4% expect no change.
Taken together, these choices describe a transition from experimentation to operational consolidation. Enterprises want fewer frameworks, more production exposure, and more ownership of the control layer. Notably, this appetite for custom in-house control planes aligns with the hybrid posture: standardize where it helps, but keep the governance and execution control where you can manage it.
Fiscal control is lagging—and that’s where “agent risk” becomes bill risk
Even if orchestration maturity is behind, cost governance is often the first place enterprises feel pain. The survey’s final finding addresses fiscal control over token consumption—specifically the risk that an autonomous loop could exhaust a budget before anyone intervenes.
Here, the picture is less mature than the orchestration roadmap.
More than a quarter of enterprises (27%) admit they