In the rapidly evolving world of enterprise AI, the line between transformative potential and overhyped promises can be difficult to see. We sit down with Dominic Jainy, an IT professional with deep expertise in AI, machine learning, and blockchain, to cut through the noise. In our conversation, he demystifies the current state of agentic systems, offering a pragmatic look at the battle between flashy Large Action Models and more mature, reliable approaches. We’ll explore why the smartest CIOs are shifting their focus from building temporary technological “plumbing” to creating lasting competitive advantages through data and governance. Jainy also breaks down the game-changing economic shift in physical AI, the foundational challenge of data quality that can stall even the most ambitious projects, and the critical balance between innovation and data privacy.
The market seems divided between the hype around Large Action Models (LAMs) and the steady progress of more established agentic systems. Could you share a specific example of where the limitations of a LAM could create real-world risks, and walk us through how a current agentic system handles that task more safely?
Absolutely. It’s easy to get swept up in the vision of a self-driving browser, but the reality is that current LAMs lack crucial memory and contextual awareness. Imagine an enterprise using a LAM to automate procurement. A user gives a natural language command: “Order 500 high-tensile bolts for the new assembly line.” The LAM, acting like an action transformer, navigates to the supplier’s website and places the order. But let’s say it orders the wrong size. The next week, the user tries again, saying, “Order the correct high-tensile bolts this time.” Without a robust memory system, the LAM might repeat the exact same mistake because it has no memory of the previous failure or the context of what “correct” means. This isn’t just inefficient; it can halt a production line.
A more mature agentic system, on the other hand, is built with safeguards and orchestration. It would likely use an LLM for understanding the initial request, but the action is governed by a workflow with checks and balances. The system would log the first failed order, and when the second request comes in, it would flag the ambiguity. It would then trigger a human-in-the-loop control, asking the user to confirm the exact part number from a pre-approved list. This approach is less flashy, but it’s built for reliability and avoids costly, repetitive errors because its architecture acknowledges the need for memory, context, and often, human oversight.
You advise treating low-level agent orchestration as a “temporary advantage.” For a CIO building an internal agent platform today, can you outline a step-by-step strategy to focus on lasting assets like golden data sets and governance, rather than on plumbing that will soon be commoditized?
That’s a critical mindset shift. We’re seeing about 10% to 20% of leading firms building their own agent platforms right now, primarily because off-the-shelf tools lack the needed reliability and policy control. But this is a short-term solution. The big cloud providers will commoditize this plumbing within a year or two. To build a lasting advantage, a CIO should follow a clear, strategic path.
First, stop focusing on the agent’s “brain” and instead meticulously document and expose your proprietary business logic. This means creating a library of high-quality, secure, and well-documented agent-callable APIs. This is your unique business operating system, something no vendor can replicate for you. Second, invest heavily in creating “golden data sets” and robust evaluation suites. These are curated, high-quality datasets specific to your domain that you can use to train models and, more importantly, to test and validate the behavior of any agentic system, whether you build it or buy it. This ensures safety and effectiveness.
Finally, formalize your governance. This involves defining security policies, establishing guardrail layers, and integrating the agent lifecycle into your existing software development and security operations workflows. The real, defensible asset isn’t the specific planning or routing algorithm you use today, but the rich domain knowledge, rigorous evaluation data, and strong governance framework that will allow you to safely swap in better vendor-provided “plumbing” as it matures.
With platforms like Nvidia’s Omniverse shifting physical AI to a cloud-based OPEX model, how does this change the game for smaller competitors? Please provide a concrete example or metric showing how this new economic model lowers the barrier to entry for advanced robotics or simulation development.
This shift is nothing short of revolutionary for competition in the industrial space. Historically, developing advanced robotics required massive upfront capital expenditure, or CAPEX. A startup wanting to design a new autonomous warehouse robot would need to build a multi-million-dollar physical mock-up of a warehouse just to test and train its AI. This was a huge barrier that protected large, established incumbents.
Now, with platforms like Omniverse and open standards like OpenUSD, that entire paradigm is flipping to an operational expenditure, or OPEX, model. That same startup can now build a photorealistic digital twin of a warehouse in the cloud. They can simulate thousands of scenarios—from navigating cluttered aisles to interacting with other machinery—on a pay-as-you-simulate basis. Instead of a $5 million initial investment, they might spend a few hundred thousand dollars a year in cloud simulation fees. This dramatically lowers the barrier to entry, allowing smaller, more agile teams to innovate and iterate on designs at a speed that was previously impossible. The competitive frontier is no longer about who has the biggest R&D facility, but who is best at managing their cloud simulation spend and leveraging open standards to avoid vendor lock-in.
The article states data quality is a major roadblock, recommending a focus on a semantic layer. What are the first three steps a company should take to build this semantic layer for its unstructured data, and what specific KPIs should it track to measure the project’s success?
Data quality is the elephant in the room for so many AI initiatives. You can’t build intelligent agents on a foundation of messy, unreliable data. Building a semantic layer is the right approach, and it begins with three foundational steps. First, you must conduct a thorough audit of your unstructured data landscape to identify and quantify the “data noise.” This means finding all the redundant copies, outdated files, and conflicting versions that clutter your systems. This initial cleanup is crucial.
Second, establish a centralized business glossary. This isn’t just a technical exercise; it involves working with business leaders to define key terms, entities, and relationships in a standardized way. This glossary becomes the shared language that both humans and AI can use to understand your data. Third, implement a metadata management strategy. This involves tagging unstructured data with rich, context-aware metadata derived from the business glossary. This is what transforms a chaotic data swamp into an organized resource that an LLM can effectively reason over.
To measure success, you need to track specific KPIs. Obvious ones include a percentage reduction in data storage from eliminating duplicate files. More importantly, you should track business-oriented metrics like the accuracy of AI-generated insights or the reduction in time it takes for a business analyst to find the right information. Another powerful KPI is the success rate of automated processes that rely on this data; as the semantic layer improves, the error rate of those processes should plummet.
Given the challenges of using sensitive data, techniques like federated learning are maturing. Can you describe a scenario where federated learning is the best choice for an enterprise, and explain the key compliance and security trade-offs when compared to using synthetic data for the same purpose?
Federated learning is a perfect fit for scenarios where collaboration is essential but data privacy is non-negotiable. A classic example would be a consortium of hospitals wanting to build a more accurate diagnostic AI model for a rare disease. Each hospital has its own sensitive patient data—scans, medical records, outcomes—that it cannot legally or ethically share in a central repository. With federated learning, a global model is trained by sending the model to each hospital’s local, secure server. The model learns from the local data on-site, and only the updated model weights, not the private data itself, are sent back to be aggregated into an improved global model.
The primary advantage here is a massive win for compliance and security; the sensitive data never leaves its secure enclave. However, the trade-off is increased complexity. It’s technically challenging to manage and orchestrate, and you can run into issues with data heterogeneity between institutions, which can degrade model performance.
Now, compare this to using synthetic data. The hospitals could generate artificial data that mimics the statistical properties of their real patient data and then pool that synthetic data to train a model. This is simpler to implement and avoids direct use of sensitive information. The trade-off, however, is fidelity. Synthetic data, no matter how good, may fail to capture all the subtle, complex patterns and outliers present in the real data. For a critical medical diagnosis model, those subtle patterns could be the difference between an early detection and a missed one. So, you trade the ironclad data privacy of federated learning for implementation simplicity, at the potential cost of model accuracy. There’s no silver bullet; it’s always a careful balancing act.
What is your forecast for the adoption of autonomous agents by mainstream enterprises over the next three years?
I believe we’ll see a steady, pragmatic, and increasingly focused adoption, rather than a sudden explosion of fully autonomous systems. The initial proofs of concept are now colliding with the messy realities of enterprise data and risk management. Over the next three years, the focus will shift from flashy demos to foundational work. Enterprises will be investing heavily in the unglamorous but essential tasks: cleaning up unstructured data, building robust semantic layers, and meticulously documenting business logic into agent-callable APIs.
We will see agentic capabilities become standard features bolted into major enterprise software suites, handling well-defined, supervised tasks within existing workflows—think of them as highly advanced copilots and automations. However, truly autonomous agents making high-stakes, independent decisions will remain on the fringe, limited to highly controlled environments. The primary barrier won’t be the capability of the AI models themselves, but the enterprise’s confidence in their data quality and governance frameworks. The winners in this next wave won’t be the ones who adopt the most agents first, but the ones who build the most reliable and trustworthy data foundation to support them.