Enterprise organizations are rapidly consolidating their agentic orchestration strategies around major model-provider platforms, with Anthropic’s Claude currently leading as the primary choice for 40% of surveyed technical decision-makers. Despite this shift toward centralized model-provider environments, a significant gap remains between the high-level ambition for autonomous, multi-step workflows and the reality of current deployments, where the majority of “agents” remain simple, single-prompt chatbot wrappers.
This trend toward platform concentration is driven by “model gravity”—the tendency for enterprises to choose an orchestration layer that is natively aligned with the frontier model they have already standardized on. While this approach simplifies the initial stack, it creates a tension between the desire for efficiency and the deep-seated fear of vendor lock-in. As organizations move from experimental sandboxes to production environments, they are increasingly seeking a hybrid control plane that allows them to maintain autonomy over their orchestration logic while leveraging the power of external model providers.
The Rise of Model-Provider Platforms
In the current enterprise landscape, orchestration platforms are dominated by a handful of major providers. According to data from a June 2026 assessment of 101 enterprises with 100 or more employees, Anthropic’s Claude stands as the primary platform for 40% of organizations. Microsoft follows at 18%, with OpenAI capturing 13%. These figures illustrate a clear preference for platforms that offer tight integration with a preferred base model, a phenomenon categorized as “model gravity.”
For these organizations, the selection of an orchestration platform is rarely about raw speed or latency, which ranked last in influence at 4%. Instead, the choice is driven by the need for reliable, multi-step execution. When asked to define success, 32% of respondents pointed to task completion reliability, while 28% prioritized multi-step workflow management. These metrics highlight a shift in focus: enterprises are no longer interested in simple conversational interfaces; they require robust, dependable systems that can carry out complex tasks from start to finish.
Despite this clear objective, the current state of enterprise AI portfolios tells a different story. When asked to evaluate their actual deployments, 71% of surveyed enterprises reported that a quarter or fewer of their “agents” are true multi-step orchestrated workflows. Only 10% of the organizations surveyed have successfully moved more than half of their agents beyond the stage of single-prompt chatbot wrappers. This discrepancy suggests that the infrastructure for sophisticated agentic orchestration is being built well in advance of the actual, complex workflows it is designed to support.
Building a Hybrid Control Plane
A defining characteristic of the evolving enterprise AI architecture is a shift toward hybrid control. By the end of 2026, 51% of enterprises expect to utilize a hybrid control plane, combining provider-native tools with external orchestration layers. Only 6% of organizations anticipate fully handing over control to a provider-managed service. This strategic design is a direct response to the risk of vendor lock-in, which was identified by 35% of respondents as their primary concern regarding provider-resident control.
This preference for hybrid structures marks a notable shift in sentiment. In a survey conducted between April and May 2026, only 34% of respondents expected to move toward a hybrid control plane, while a larger segment of 12% were open to fully provider-managed solutions. The rapid pivot toward maintaining external control suggests that enterprises are hardening their architectures against the possibility of being tethered to a single vendor’s ecosystem. Security and permissioning limitations, cited by 28% of respondents, further reinforce the need for internal control logic that can operate independently of the underlying model.
Investment patterns reflect these strategic priorities. With 34% of budgets directed toward agent workflow tooling, organizations are clearly focused on the machinery required to link steps together reliably. Security and permissions enforcement follows at 25%, while scaling infrastructure accounts for 20% of planned spend. This allocation demonstrates that the priority is to build and harden the orchestration layer rather than simply increasing the number of deployed models.
The Challenge of Fiscal Control
As enterprises scale their agentic operations, managing the costs associated with token consumption has emerged as a significant operational hurdle. Currently, fiscal control remains largely reactive. More than a quarter of enterprises (27%) lack a real-time, programmatic method to halt a runaway agent, meaning they only discover budget-breaking costs after reviewing logs. While 32% of organizations rely on native caps and throttles provided by their model platforms, these measures are often limited by the vendor’s own tooling, which can inadvertently increase dependence on a single provider.
A subset of organizations—23%—is opting to build custom gateways, while 19% are utilizing cross-model routing to optimize costs through arbitrage. These approaches represent a move toward treating token consumption as a deterministic engineering problem. Notably, this lack of mature fiscal instrumentation is more prevalent in the mid-market; approximately 34% of enterprises with fewer than 2,500 employees exercise only reactive control, compared to 20% of larger organizations. As these companies continue to shift agents from sandbox environments into production, the need for robust, real-time cost-control planes will likely become an increasingly urgent requirement for CIOs and CTOs.
The path forward for these enterprises involves a delicate balancing act: consolidating frameworks to improve operational efficiency, while simultaneously investing in custom control layers that prevent over-reliance on any single provider. The success of these initiatives will depend on how quickly the deployed reality of multi-step agents can match the current, ambitious orchestration strategies being put in place.
Future research waves will continue to monitor whether this gap closes as organizations move further into production. For ongoing updates on enterprise AI deployment standards and technical benchmarks, readers are encouraged to monitor future industry reports and stay engaged with organizational updates regarding AI governance and infrastructure development.
Worth a look