Dominic Jainy stands at the forefront of the next great leap in enterprise technology, bringing years of deep-seated expertise in artificial intelligence, machine learning, and blockchain to the table. As organizations move beyond the initial excitement of generative AI, Jainy has become a pivotal voice in guiding leadership teams through the transition from reactive chatbots to proactive “digital workers.” His perspective is shaped by a rigorous understanding of how autonomous agents can be integrated into complex business architectures to drive genuine value. In this conversation, we explore the nuances of agentic AI—systems that do not just talk but act—and examine the strategic shifts required in governance, data fabric, and team structure to ensure these technologies deliver on their promise without spiraling into chaos.
The discussion delves into the evolution of AI execution authority, the technical barriers that keep nearly half of all projects trapped in the proof-of-concept phase, and the critical need for a hierarchical orchestration layer to manage a growing digital workforce. We also address the looming challenge of “bill shock” caused by unbounded agent behavior and the fundamental redesign of the IT department as it moves from prompt engineering to complex harness engineering.
Traditional chatbots generate content based on prompts, whereas agentic AI can trigger workflows and modify system states autonomously. How do you distinguish between these two in a production environment, and what specific reasoning capabilities must be present for a system to be considered a true digital worker?
The fundamental distinction lies in the transition from reactivity to autonomy, or what we call execution authority. While a standard generative AI chatbot is essentially a sophisticated autocomplete engine that follows a prompt to generate content, an agentic AI system uses a reasoning engine to decompose a high-level goal into a series of actionable steps. For example, at Malaysia’s Ryt Bank, we see agents that don’t just answer questions about balances but actually queue up transactions, pausing only for a human to confirm the final transfer. This ability to call APIs, modify system states, and self-correct if a specific step fails is the hallmark of a true digital worker. To be considered truly agentic, a system must possess the cognitive capability to track its own state and adapt its actions based on the feedback loops it encounters during execution. It is no longer about just “reasoning and responding”; it is about “planning and executing” within a live production environment.
Many organizations struggle with “innovation theatre” where pilots remain isolated from core business systems. What specific technical or organizational gaps usually prevent a proof-of-concept from reaching live production, and how should an enterprise prepare its data fabric for this transition?
It is a sobering reality that roughly 44% of organizations remain stuck at the proof-of-concept stage, a phenomenon often referred to as innovation theatre. The gap between a successful pilot and a live production environment is usually a lack of deep integration with core enterprise systems like ERP or logistics frameworks. Many companies build impressive agents in isolated sandboxes, but these agents fail because the surrounding ecosystem—the data flows and business workflows—isn’t ready to support autonomous action. To bridge this, an enterprise must move away from treating AI as a standalone project and instead embed it into the very fabric of the organization. This requires a digital backbone that allows agents to access reliable, real-time data while ensuring that sovereignty extends beyond just the data layer into the decision-making layer. When we see successful implementations, such as the document handling systems in Japan that reduced processing time by 90% and dropped error rates by 80%, it is because the data fabric was prepared to handle the agent’s operational needs from day one.
Scaling multiple digital workers often leads to agent sprawl and disconnected silos without shared context. How can a hierarchical orchestration layer prevent agents from entering infinite loops, and what parameters are necessary to ensure effective agent-to-agent communication across different business functions?
As you scale from one agent to dozens, you inevitably face the risk of agent sprawl, where disconnected digital workers operate in silos and potentially conflict with one another. To prevent this, enterprises must implement a hierarchical orchestration layer where a “manager agent” coordinates the activities of various specialized worker agents. This layer is responsible for task decomposition and, perhaps most importantly, establishing clear termination conditions to ensure that an agent doesn’t get caught in an infinite loop of retries. Technical parameters like state tracking, schema constraints for tool invocation, and fallback logic are essential to maintain control. Effective communication between these agents requires a shared context that is decoupled from core systems via APIs and events, ensuring that the entire workforce acts as a unified entity. Without this level of governed orchestration, you aren’t building a workforce; you’re just creating a series of uncoordinated scripts that can break the moment they encounter an unexpected variable.
Moving beyond simple data privacy, agentic AI requires behavioral and cognitive control to manage operational risks. What does a proactive, real-time governance framework look like in practice, and how can teams implement adversarial reviews to prevent failure cascades between agents?
In the world of agentic AI, sovereignty must evolve to cover three distinct domains: behavioral, operational, and cognitive control. Traditional security measures that only protect the data layer are no longer sufficient when an agent has the power to execute tasks that impact regulatory compliance or brand reputation. A proactive governance framework treats trustworthy AI as an engineering discipline rather than a vague philosophy, mapping ethical guardrails and enterprise rules directly into the AI pipeline as version-controlled constraints. We implement adversarial reviews to stress-test the system for “failure cascades,” a dangerous scenario where one agent’s erroneous output becomes the corrupt input for the next agent in the chain. By creating these interceptors—where a privacy rule acts as a hard stop and a fairness requirement becomes a scoring constraint—we ensure that every action is logged and every decision is traceable. This level of transparency is the difference between a program that survives a production incident and one that is quietly shut down after its first major error.
Agentic workflows introduce high cost variability due to retries and multi-model orchestration. Since simple calculators are often unreliable for these budgets, what specific guardrails or iteration limits should a leader implement to shift from open-ended token consumption to a predictable cost-per-outcome model?
One of the most significant “bill shocks” for CIOs comes from the unbounded behavior of autonomous agents. In a traditional SaaS model, costs are predictable, but an agent tasked with resolving an IT ticket might take five steps or, if it encounters a bug, it might loop 50 times, consuming tokens at an exponential rate. To manage this, leaders must implement hard usage guardrails, such as strict iteration limits and token budgets per specific workflow, effectively treating AI costs like cloud economics. We need to shift the financial perspective from open-ended consumption to a “cost-per-outcome” model, where the value of the completed task is weighed against the total cost of ownership. This TCO must account for not just token usage, but also multi-model orchestration, data access fees, and the governance overhead required to keep the agent running safely. By modeling these full-system costs and using continuous optimization, organizations can turn a highly variable expense into a predictable operational line item.
The role of the IT department is shifting from simple prompt engineering to complex context and harness engineering. How should leaders restructure their teams to support these autonomous workers, and what strategies help employees transition from performing manual tasks to overseeing AI-driven systems?
The era of the standalone prompt engineer is already fading, as the role is far too narrow for the complexities of agentic AI. Success now depends on cross-functional teams capable of “harness engineering” and “context engineering,” which involve building the environments in which agents operate and ensuring they have the right business policy encoded into their logic. Leaders must restructure their departments to focus on workflow redesign and risk governance, moving away from manual task execution. For the broader workforce, this is a profound “people transformation” where employees shift from being the ones who do the work to being the ones who oversee and guide the systems doing the work. Change management is critical here; resistance is highest when AI acts without transparency, so we must emphasize how these agents reduce “toil” rather than replace jobs. The goal is to elevate the nature of human work, focusing on decision accuracy and strategic oversight while the agents handle the high-volume, repetitive execution.
What is your forecast for agentic AI?
My forecast is that within the next few years, we will stop talking about AI as a separate “tool” and start viewing it as a standard component of the enterprise workforce, but only for those who master the governance gap today. We will see a shift away from measuring success through headcount reduction and toward outcome-based metrics like reduced cost-to-serve and faster cycle times. The most successful organizations will be those that move beyond the “innovation theatre” by building a robust digital backbone and a culture of oversight. Eventually, the friction of manual business processes will seem as archaic as paper filing systems, as autonomous agents become the primary interface through which work is executed across the global economy. Durable success will be defined not by who has the most advanced model, but by who has the most reliable and governed integration of those models into their core business fabric.