Dominic Jainy has spent years at the intersection of machine learning and enterprise infrastructure, witnessing firsthand the transition from experimental AI pilots to the complex, agentic ecosystems of 2026. As organizations integrate more autonomous agents and embedded capabilities, the primary challenge for leadership has shifted from simple model performance to the intricate web of control layers—orchestration, identity, and governance—that keep these systems in check. In this conversation, we explore the architectural hurdles CIOs face when managing a multi-vendor AI environment and how to prevent a fragmented control landscape from undermining enterprise security and efficiency.
Our discussion covers the emergence of the “control plane” as a critical architectural concept, the risks of making isolated technology purchases without a holistic view, and the tactical ways specialized tools like agent harnesses are redefining model flexibility. We also dive into the necessity of maintaining visibility across overlapping permissions and the shift toward establishing authoritative sources for data and identity in a world of autonomous AI.
How has the role of the CIO changed now that AI control is no longer just about selecting the right model, but about managing a multi-layered architecture?
It is a profound shift from a “single-product” mindset to a systemic governance challenge. We are seeing CIOs move beyond simply asking which Large Language Model provides the best response to managing multiple distinct control layers: orchestration, model selection, data access, identity, permissions, and observability. If you look at how companies are deploying today, they aren’t just adding tools; they are accumulating behavioral systems. The real headache is not the performance of a single agent, but the friction that occurs when these layers collide and create a fragmented environment where responsibility is split between security, data, and infrastructure teams. It requires a much more holistic architectural view than we’ve ever needed for standard SaaS deployments.
You mentioned that sensible product decisions can lead to an irrational whole. Could you explain how this fragmentation happens in a typical enterprise environment?
It happens almost invisibly because no single purchase feels like a mistake at the time. You might choose Salesforce for its robust orchestration and permissions, then integrate GitHub for its specialized model routing, and finally use Alteryx to govern your business logic across specific data sets. Individually, these are high-impact, rational choices that solve specific pain points for your developers or sales teams. However, the aggregate result is a “control sprawl” where your identity rules might behave one way on one platform and completely differently on another. Without a centralized architectural vision, you end up with a mess of overlapping authorities where nobody knows which system is truly authoritative when a conflict arises during a runtime operation.
We are seeing tools like GitHub’s HydraFusion introduce dynamic model routing at runtime. How does this change the way we think about the lifecycle of an AI application?
This is a massive departure from the “set it and forget it” deployments we saw in previous iterations of software development. Instead of locking a project into one specific model during the development phase, HydraFusion allows the system to choose the best model—or even a combination of models—dynamically based on the work being done at that exact moment. This turns model selection into a real-time control decision, allowing for better cost optimization and performance tuning without needing to rewrite significant portions of the codebase. It adds a level of fluidity to the stack, but it also increases the need for rigorous observability to understand why a certain model was chosen for a specific query. It feels like we are finally moving away from static configurations toward truly intelligent, self-optimizing infrastructure.
Several vendors are pitching the idea of an “AI Harness” or a “Control Plane.” How do these components help stabilize an environment with multiple agents?
Think of a harness as a protective shell that surrounds a model with essential enterprise features like context management, memory, tools, and safety guardrails. Salesforce’s Enterprise AI Harness, for instance, is designed to orchestrate agents across multiple systems, ensuring that permissions and identity are consistently applied even when dealing with third-party tools. This approach makes it significantly easier to swap out underlying models without having to rebuild the entire support structure from scratch every time a newer, cheaper model hits the market. By using a model-agnostic harness, you gain a level of infrastructure independence that protects you from being overly reliant on a single provider’s roadmap. It provides a sense of security and consistency in an otherwise volatile vendor landscape.
Mistral and Boomi are taking different approaches to the control problem. How do these infrastructure and integration perspectives add to the CIO’s architectural burden?
Mistral is making a very strong case for deployment flexibility and infrastructure control, arguing that where your data sits and which infrastructure it runs on is a primary control lever. Meanwhile, at events like the Boomi World Tour, we see a focus on the integration layer, where governance includes routing queries to the right models and carrying a human’s specific access rights through to an agent acting on their behalf. This brings identity and permissions directly into the conversation with model routing and agent behavior, which are often handled by separate teams. These approaches aren’t necessarily competing for the same technical space, but they overlap in terms of function—context, orchestration, and governance—making it harder for a CIO to draw a clean line between where one vendor’s responsibility ends and another’s begins.
If a single enterprise-wide control plane isn’t the answer, how should CIOs determine which system is “authoritative” for specific functions like identity or data access?
The solution lies in defining clear boundaries rather than trying to find a “one-ring-to-rule-them-all” software package, which is often a pipe dream in a large enterprise. An organization might decide that its core identity and permission rules must remain centralized in its primary directory service, while allowing application-specific orchestration to remain local to Salesforce or Microsoft environments. You have to be very deliberate: perhaps one platform governs access to trusted business data, but it doesn’t necessarily have the final say on how an agent routes work across different models. What matters is that your architecture document clearly states which control takes precedence when two layers try to make a decision about the same piece of data. Without this “precedence map,” you are essentially leaving your security posture up to chance.
Beyond just taking an inventory, what specific telemetry or visibility should leadership be looking for across these control layers?
An inventory is just a list, but true visibility is about understanding the interaction between autonomous systems. CIOs need to dive into the telemetry produced by each layer to see how work is being routed and where potential bottlenecks or security gaps exist in the workflow. You need to know exactly which data systems a specific control layer depends on and what happens to a person’s access rights as they move through a complex agentic process. If a tool like Boomi is carrying a user’s permissions into an agent action, we need to be able to audit that transition in real-time to ensure no privilege escalation occurs. This level of granular visibility ensures that the sophisticated management tools developed by AI vendors don’t just become another layer of opaque complexity that we can’t troubleshoot.
What is your forecast for the evolution of AI control architecture over the next few years?
I forecast a major industry shift toward “federated governance” where we stop looking for a single pane of glass and start focusing on standardized protocols for control layer communication. We will see the “AI Harness” become a standard requirement for any enterprise deployment, likely with three or four dominant frameworks emerging as the preferred connective tissue for the Fortune 500. By the end of this decade, the most successful CIOs won’t be the ones who picked the “best” model, but the ones who built the most resilient architecture—one that allows them to plug in any agent while maintaining a single, immutable source of truth for security and business logic. The complexity of these systems is only going to increase, so building that foundation of visibility today is the only way to stay ahead of the coming sprawl of autonomous agents.