With the buzz around AI reaching a fever pitch, most conversations focus on bigger models and more impressive benchmarks. But today, we’re speaking with Dominic Jainy, an IT professional with deep expertise in AI, machine learning, and blockchain, who argues that the real revolution isn’t in the model, but in the architecture underneath. He suggests the next breakthrough lies in “agentic AI”—a system of smaller, collaborative agents that requires a fundamental rethinking of how we handle data.
This conversation explores the shift from single-answer generative AI to the continuous, looping intelligence of agentic systems. We’ll delve into why fragmented data moves from being an annoyance to an active danger, and how a unified data layer acts as the essential “shared memory” for these autonomous agents. Dominic will also explain why treating AI as a “plug-in” is a recipe for failure, what it means to design an AI-first architecture from the ground up, and how the human role evolves from giving direct commands to strategically refining an AI’s intent.
You describe agentic systems as running “ongoing loops” rather than simply answering prompts. Could you walk us through a real-world example of how a team of agents might observe, decide, and act over time, and what specific business metrics would demonstrate their effectiveness?
Absolutely. Imagine an e-commerce platform. Instead of a human analyst pulling reports, you have an agentic system. One agent monitors real-time customer browsing behavior, noticing a spike in searches for “raincoats” in a specific region. It doesn’t wait for a prompt; it observes this signal. It then communicates with an inventory agent, which confirms we have a surplus of raincoats in a nearby warehouse. A third agent, a pricing and promotions expert, then decides to launch a flash sale targeted only to that region. The results are tracked by another agent, and this entire loop—observe, decide, act—refreshes every few minutes. The effectiveness isn’t just a gut feeling; you’d see it directly in metrics like a higher conversion rate for that product, a reduction in slow-moving inventory, and an increase in regional sales revenue, all without a single human instruction.
The article calls fragmented data “dangerous” in agentic systems. Can you share an anecdote where agents with different data realities caused tangible damage? How would a unified, identity-resolved layer have prevented this, and what are the first practical steps for a company to build one?
I recall a situation at a large retailer where the consequences of this became painfully clear. Their marketing agent, working off a customer database that updated every 24 hours, identified a group of loyal customers and sent them a special offer on a popular new speaker. Simultaneously, the logistics agent was operating on a real-time inventory feed that showed the speaker had just sold out. The result was a wave of angry customers clicking a “special offer” link only to find an “out of stock” page. The damage wasn’t just lost sales; it was a breach of trust. A unified, identity-resolved layer would have served as that single source of truth. Both agents would have seen the same reality: the customer is loyal, but the product is unavailable. The system would then have made a coherent decision, perhaps offering a discount on a similar item instead. The first step for any company is to stop thinking in terms of siloed application data and start investing in a central platform that resolves customer identity across all touchpoints, creating that essential shared memory.
You compare agentic ecosystems to microservices that must interpret data identically. Beyond standard APIs, what specific architectural patterns or engineering practices ensure this shared understanding? Please share an example where two agents misinterpreting the same signal led to chaos instead of autonomy.
This is where the engineering gets really challenging, and it goes far beyond just having well-defined APIs. The key is creating a shared semantic layer or ontology. It’s a framework that defines what data means. For example, a signal indicating “inventory_level = 10” needs to mean the same thing to every agent. I saw a case where a supply chain agent interpreted that signal as “critically low, reorder immediately,” while a marketing agent, designed to create scarcity, interpreted it as “perfect time for a ‘last chance to buy!’ campaign.” You can imagine the chaos. The marketing agent triggered a sales surge, the supply chain agent couldn’t fulfill the orders, and the company ended up with backorders and furious customers. True interoperability means ensuring that when one agent sends a signal, the receiving agent understands the context and intent, not just the raw data.
You argue that agentic AI can’t be a “plug-in” and must be part of the core architecture. For a business with legacy systems, what does the transition to an “AI-first” data model actually look like? What are the key initial changes to infrastructure and governance?
For a business with deep-rooted legacy systems, the idea of an “AI-first” model can feel overwhelming, like you have to tear everything down. But it’s not about a big bang replacement. The transition is about fundamentally changing how you think about data flow. The first key change is building infrastructure designed for feedback loops. Instead of data flowing one way into a data warehouse for analysis, it needs to be able to flow back into operational systems to inform agentic decisions. It’s about creating a circular data economy. On the governance side, you have to move beyond rules for human access and start creating policies for autonomous behavior. This means establishing a new kind of oversight committee that asks questions like: What are the boundaries for this agent’s decisions? What happens if its actions lead to a negative outcome? It’s a shift from governing data access to governing intelligent action.
The article states that the human role shifts from “giving instructions to refining intent.” What does this oversight look like day-to-day? Describe the tools or dashboards a manager would use to spot drift or bias and “course-correct” an agentic system without micromanaging its actions.
The day-to-day changes completely. A manager’s dashboard in this world doesn’t show a list of individual transactions an agent approved. Instead, it visualizes alignment. You might see a high-level dashboard showing that the “customer retention” agents are successfully reducing churn, but perhaps their actions are disproportionately favoring one customer segment, indicating bias. Or you might see that the “inventory optimization” agents are hitting their cost-saving targets, but at the expense of delivery speed, which misaligns with the company’s broader goal of customer satisfaction. To course-correct, the manager doesn’t dive in and say, “Stop offering that discount.” Instead, they adjust the system’s priorities—they might increase the weight given to the “customer satisfaction” metric in the agents’ objective function. You’re not the puppeteer pulling every string; you’re the strategist defining the rules of the game and refining the ultimate goals.
What is your forecast for agentic AI’s adoption? Which industries will be first to successfully move beyond generative Q&A to fully autonomous systems, and what common hurdles will they face in the next three to five years?
I believe we’ll see the quickest and most successful adoption in industries where the feedback loops are fast and the data is abundant, like e-commerce, supply chain logistics, and financial trading. These fields are already driven by real-time signals—customer behavior, market shifts, and operational events. Their challenge isn’t a lack of data; it’s the inability to act on it at machine speed. The biggest hurdle they will all face, without a doubt, is architectural debt. For years, they’ve built systems that silo data and are designed for one-way transactions. The primary struggle won’t be finding the right AI models; it will be the massive undertaking of re-architecting their data foundations to support unified, interoperable, and continuously learning systems. The technology for agents is arriving, but the real work is in building a home for it to live in.