Laurent Giraid stands at the forefront of the enterprise AI revolution, bringing years of technical experience in machine learning and natural language processing to the complex world of agentic orchestration. As organizations transition from simple experimental chatbots to sophisticated, multi-step autonomous systems, Laurent has become a vital voice in navigating the architectural and ethical challenges that come with scaling these technologies. His perspective is grounded in the harsh realities of production environments, where the initial excitement of model performance often clashes with the pragmatic needs of security, cost control, and governance. This conversation explores the shifting landscape of AI control planes, examining why the modern enterprise is moving away from single-provider lock-in and toward a more fragmented, hybrid approach to agent management.
The following discussion synthesizes the current state of AI adoption within large-scale organizations, highlighting the transition from a “one model” strategy to a pluralistic orchestration stack. We explore the driving forces behind platform selection—prioritizing flexibility and security over raw model performance—and the emerging gap between the desire for autonomous agents and the actual maturity of deployed portfolios. Laurent also sheds light on the looming “fiscal cliff” for AI spend, where a significant portion of enterprises still lacks the real-time tools to prevent runaway execution costs, emphasizing that while the technology is ready for governance, the financial guardrails are still lagging behind.
Many organizations now manage an average of three different agent orchestration platforms simultaneously. What specific operational challenges and strategic advantages arise when an enterprise refuses to settle on a single “control plane” for their AI agents?
The reality on the ground is that 85% of enterprises have moved past the idea of a monolithic provider, with 64% now juggling three or more platforms. This plurality is born out of a desperate need for flexibility; when 29% of decision-makers cite “flexibility across models and tools” as their primary driver, they are essentially saying they don’t want to be held hostage by a single vendor’s roadmap. Strategically, this allows a company to leverage Microsoft AI Foundry for 70% of their stack to satisfy existing enterprise agreements while simultaneously using OpenAI’s SDK for 68% of their experimental workflows. However, the operational friction is palpable, as managing a mean of 3.1 platforms requires a highly skilled technical team—often composed of the 22% of software and ML engineers we see leading these projects—who must bridge the gaps in security and visibility across disparate environments. It is a fragmented way of working, but for the modern CTO, the risk of being locked into a sub-optimal model outweighs the headache of managing a diverse, sometimes messy, orchestration portfolio.
When we look at how companies choose their orchestration layer, “model gravity” seems to be losing its pull compared to governance. Why are we seeing security and permissions take such a front-seat role in the selection process?
We are seeing a fundamental shift in the “buying logic” of AI, where the allure of a state-of-the-art model is no longer enough to win the contract. Only 10% of enterprises are picking an orchestration platform based on native alignment with a specific base model, which is a staggering decline when you consider the hype surrounding frontier models just a year ago. Instead, 17% are prioritizing security and permissions, while another 30% combined are looking for production reliability and execution control. This tells me that the “developer convenience” phase of AI is ending, and the “operational governance” phase has begun. Enterprises are terrified of the “black box” nature of provider-resident control; 37% specifically fear the security and permissioning limits of these providers, which is why we see a majority of 53% moving toward a hybrid control plane. They want the power of the model, but they want to keep the “kill switch” and the permissioning logic within their own walls, or at least in an environment they can audit and constrain.
There is a notable gap between the current leaders in usage, like Microsoft and OpenAI, and the platforms being considered for future adoption, with Anthropic taking a significant lead. How do you interpret this disconnect between the installed base and the forward-looking pipeline?
This is perhaps the most telling metric in the current landscape: while Microsoft currently anchors the installed base, appearing in seven out of ten stacks, it is Anthropic’s Claude platform that leads the forward consideration set at 43% for those planning to move. This is a massive jump, especially when you compare it to the 17% consideration for Microsoft or the 25% for OpenAI among those looking to change. This disconnect suggests that the first wave of adoption was driven by “convenience and existing relationships,” but the second wave is being driven by “intentionality and specialized performance.” Companies are realizing that the default seat of record they got through their enterprise agreement might not be the most agile tool for complex, agentic work. There is a sense of “planned re-platforming” happening, where 67% of enterprises expect to add or replace a platform within the next 12 months, signaling that the market is still very much up for grabs.
The industry often uses the term “agent” loosely, but your data suggests many of these are still just glorified chatbots. How are enterprises actually measuring the maturity of their agent portfolios, and what does it take to move into “true” orchestration?
The “chatbot trap” is real, and the data reflects a sobering truth: 35% of enterprises admit that only a quarter or fewer of their deployed agents are performing genuinely orchestrated, multi-step work. We are currently in a “bridging the gap” phase where 47% of organizations have managed to get about 26% to 50% of their portfolio into the realm of true orchestration. To move beyond a simple prompt wrapper, an organization must optimize for task completion reliability (30%) and multi-step workflow management (27%), which is far harder than it looks on a demo slide. It requires a move toward stateful systems that can handle memory, tool-calling, and error recovery across multiple turns of conversation. Interestingly, maturity isn’t tied to company size; a 50,000-employee behemoth is just as likely to be stuck with basic assistants as a 500-person firm. What actually correlates with maturity is the number of platforms used—those furthest along run an average of 3.5 platforms, proving that sophistication in this space currently requires a diverse toolkit rather than a single, all-in-one solution.
Investment seems to be shifting away from the tools used to build agents and toward the infrastructure used to watch them. What does this tell us about the current “pain points” for directors and VPs of AI?
The money is following the fear. When 31% of planned budget growth is dedicated to agent monitoring and debugging, and another 30% is going to security and permissions enforcement, it sends a clear signal: the builders have built, and now the governors are worried. Only 19% of the growth is earmarked for workflow tooling, which suggests that the basic “how do we make this?” question has been answered, and has been replaced by “how do we keep this from breaking?” and “how do we know what it’s doing?” For a Director of AI, the nightmare scenario isn’t a slow agent; it’s an agent that has a 3.63 “value for money” rating and is hallucinating through a sensitive customer database without an audit trail. This emphasis on observability and governance is the hallmark of a technology that is finally being integrated into the core business, where “operational stability” is valued more than “end-user experience,” which only garnered 7% of the optimization focus.
Fiscal control remains a significant blind spot, with 21% of enterprises still relying on reactive monitoring. What are the risks of this “metered by hope” approach, and how are the more advanced organizations solving the cost problem?
It is genuinely alarming that one in five enterprises only finds out their agent went into a runaway execution loop when the bill arrives at the end of the month. This “reactive only” posture is a fiscal ticking time bomb in an era where token costs can escalate exponentially in seconds. Another 30% are leaning on native platform controls, which is essentially trusting the person selling you the fuel to also manage your fuel gauge. The most sophisticated players—the remaining 49%—are taking matters into their own hands by building custom gateway plumbing or using dynamic routing arbitrage to offload heavy tasks to low-cost models in real-time. These organizations are treating token consumption as an engineering constraint to be managed deterministically, rather than a variable cost to be lamented. Given that “value for money” is the lowest-rated attribute of these platforms, those who fail to move toward proactive, programmatic cost-cutting will likely find their AI initiatives defunded before they ever reach true autonomy.
What is your forecast for the future of agentic orchestration over the next two years?
By the end of 2026, I expect the “hybrid control plane” to be the absolute standard for any organization of consequence, as the current 53% expectation continues to climb. We will see a massive consolidation of the “chatbot” tier into a few standard utilities, while the “orchestration” tier remains a pluralistic, highly customized battleground where 78% of companies keep a firm hand on the steering wheel outside of the model provider’s reach. The 21% of companies currently lacking a “kill switch” for their budgets will either adopt custom middleware or be forced out of the agent space by unsustainable costs. Ultimately, the winners won’t be the companies with the smartest models, but the ones with the most robust governance and the ability to switch models as easily as they switch cloud providers today. We are moving toward a world where AI agents are treated less like magic and more like any other piece of critical enterprise infrastructure—governed by intention, monitored in real-time, and strictly metered by the bottom line.
