Dominic Jainy stands at the forefront of the modern shift toward agentic AI, bringing years of deep-trench experience in machine learning and blockchain to the table. As an IT strategist, he has watched the initial wave of artificial intelligence hype settle into a far more complex reality: the era of production-ready AI agents navigating intricate corporate ecosystems. His expertise is rooted in the belief that the “intelligence” of a model is only a fraction of the battle, with the real challenge lying in the governing structures and legacy workflows that these agents must inhabit. Today, we delve into how leading global enterprises are moving past the “black box” approach, focusing instead on the messy, necessary work of orchestration, conflict resolution, and the human-centric guardrails that keep autonomous systems from veering off course.
Through our conversation, we explore the specific lessons learned by major financial institutions and tech giants as they transition from isolated AI pilots to interconnected agent workflows. We examine the critical “confidence gap” where organizations feel ready to govern but remain ill-prepared to handle multi-agent conflicts. We also discuss the philosophy of keeping authority within hard-coded orchestration rather than delegating it to the agents themselves. Finally, we look at how the deployment of these agents is serving as an unexpected diagnostic tool, exposing long-hidden inefficiencies and “shadow workflows” within the enterprise, and what this means for the future of operational excellence.
When an AI agent mistakenly applies rules from one subsystem to another, as we saw in recent high-profile production environments, what does that tell us about the risks of “all-in-one” agent architectures?
The situation we saw at Meta is a perfect case study for the dangers of architectural over-simplification. When they first deployed an agent to manage data-column onboarding across recommendation systems, they gave it access to instructions from every subsystem simultaneously, which sounds efficient on paper but created a massive cognitive overlap. In production, the agent started hallucinating or misapplying rules meant for one specific area to an entirely different one, which is the digital equivalent of a mechanic trying to fix a jet engine using the manual for a lawnmower. To solve this, they had to pivot to a much more granular structure, breaking that single entity into 14 specialized agents. By using a small, deterministic orchestrator to handle the routing and reliability checks, they didn’t just fix the errors; they slashed the preparation time for new data configurations from a grueling three or four days of manual labor to less than an hour. It proves that narrow focus is almost always superior to broad, unfettered access when you’re dealing with high-stakes production workflows.
A recent industry survey of over 800 executives showed that while 82% are confident in their ability to govern agents, only 26% feel ready to resolve conflicts when multiple agents and workflows clash. How do you explain this massive disconnect?
This gap between confidence and readiness is exactly where most enterprise AI projects go to die right now. It is relatively easy for a CIO to say they have a governance framework in place—usually a set of static policies or high-level ethical guidelines—but those documents rarely survive contact with the “messy middle” of integration. When you have 858 executives being surveyed, and such a small fraction feel prepared for conflict, it highlights that we are still thinking about AI agents as individual performers rather than a collaborative orchestra. The real friction happens when the controls for one agent are insufficient for three or four agents doing interconnected work, such as when a change in a CRM record ripples through billing, commissions, and sales forecasts. We are realizing that “governance” isn’t a checkbox; it’s a living, breathing technical challenge of monitoring how these agents talk to each other and preventing them from making a bad situation worse before a human even realizes something is wrong.
There is a growing sentiment among IT leaders that authority should never sit with the agents themselves, but rather within the underlying code. Why is this distinction between “orchestration in code” and “orchestration in agents” so vital for security?
Tiago Azevedo at OutSystems hit the nail on the head when he argued for keeping orchestration in code rather than letting the agents decide the “how” and “why” of a process. In a low-code or traditional development environment, the code sets the hard boundaries—the “guardrails”—that define exactly which options are available at any given step, even if the agent is allowed to choose between them. If you give an agent the authority to define its own permissions or call other systems without a deterministic check, you are essentially creating a security vacuum. We need identity and permission checks to trigger every single time an agent acts, especially for high-risk actions that require a human to step in, investigate, or even reverse a transaction. You want your agents to be the muscle that completes the task, but the “brain” or the authoritative logic must remain in a layer that humans can inspect, audit, and hard-code to prevent runaway processes.
We’ve seen cases like Trust Bank Singapore where AI struggled because the “runbooks” it was following were outdated. How much of an agent’s success depends on the quality of existing business documentation?
The hard truth is that if your runbook is incorrect, it doesn’t matter if you’re using the most advanced AI model on the planet; the final output is going to be wrong every single time. Srinivas Patil at Trust Bank learned this the hard way when their automated procedures for investigating IT problems became stale as their systems evolved, leading to inconsistent and unreliable AI analysis. They eventually had to pivot back to a model where prescribed runbooks gather the evidence, the AI analyzes it, and then human engineers—not the machines—make the final call on what action to take. It’s a sensory experience for these teams to realize that their “ground truth” documentation is often a house of cards. This is why testing runbooks against reference data is now a mandatory step, ensuring that the signals the AI is looking for actually still exist in the current environment.
Some organizations, like Barclays, are finding that deploying AI agents actually exposes hidden problems in their manual workflows that have been ignored for years. How can this “mirror effect” be turned into a strategic advantage?
Deploying agents is like shining a high-intensity floodlight into the dark, dusty corners of your enterprise architecture. At Barclays, for instance, the corporate customer due diligence process can take months and involves a labyrinth of manual information gathering across dozens of teams and systems. When you try to build specialized agents to automate parts of that, you quickly discover that the “architecture diagrams” everyone has been following are often just polite fictions, hiding a mess of informal employee workarounds and broken integrations. Julie Linn Teigland at EY pointed out that an agent might complete its specific assignment perfectly, but because the broader business process isn’t coordinated, it could actually accelerate errors across billing or HR. The strategic advantage here is that the effort of preparing for AI forces a company to finally clean up and understand processes they’ve ignored for a decade. It turns an automation project into a total business transformation, where the byproduct is a leaner, more transparent organization.
What is your forecast for the evolution of AI agents within the enterprise?
I believe that over the next few years, we will see a shift away from the “chat-bot” mentality and toward a “multi-agent fabric” where the primary role of IT is no longer building models, but managing the inter-dependencies between them. We are moving toward a world where the most successful companies won’t be those with the smartest agents, but those with the most robust deterministic orchestration layers that can catch a conflict before it impacts the bottom line. I predict we will see the rise of “agent auditors”—both human and automated—whose entire job is to stress-test the runbooks and business rules that these agents rely on to ensure they haven’t drifted from reality. Ultimately, the biggest win won’t just be the hours saved, like Meta’s reduction from days to minutes, but the fact that AI will finally force enterprises to achieve the operational clarity they’ve been promising since the early days of digital transformation. It is a future where the machine does the heavy lifting, but the human remains the indispensable architect of the logic that governs it.