AI’s Future Lies in Strong Data, Human Empathy, and Meaningful Outcomes

Honestly, I still consider myself a student of AI and data product management. I began with the logic of circuits and control systems at engineering school, and that love of orderly, dependable systems never left me. Over the years I realised that Generative AI and Agentic AI are deeply grounded in the discipline of data products—without strong data foundations, they simply fail to deliver, and that is exactly what much of the industry is now discovering.

My perspective has shifted from chasing the technology to chasing the outcome. Traditional AI and ML aimed for goal-oriented results too, but LLMs pre-trained on the world’s knowledge bring genuinely emergent capabilities. Over the last year, models have matured in reasoning to the point where we are entering an era of autonomous decision-making. I still believe we are some distance from fully autonomous decisions in complex environments—but the trajectory is unmistakable.

Agentic AI uses large language models to sense and understand data, reason over it, and trigger goal-oriented actions. That combination moves us from GenAI, a technology that understands to systems that can act on processes and goals to achieve the desire results. For example, a collections agent reasons through customers’ patterns and chases the money to achieve desired working capital results. Right now, we are orchestrating this learning and action through effective context engineering and that is where the meat is, effective knowledge and context management for Ai and Humans to use and work on. 

The real turning point will arrive when models can learn on the fly, in the flow of our interaction with them. For businesses, the shift is from automating repeat complex tasks to augmenting judgement and empowering humans to take on more complex problems with the help of Agentic AI. Decisions get faster and more consistent, routine work gets automated, and human effort moves up the value chain toward the exceptions and the strategy. Innovation accelerates because teams can test ideas against real data at a pace that was simply not possible before.

As recent MIT research shows, most AI initiatives fall short of real business outcomes. Companies pour resources into state-of-the-art technology, but the real work is understanding your customers, their pain points, and identifying the right problems that actually need AI, not necessarily GenAI or Agentic AI, to solve them. The gap between a compelling demo and a dependable product is where most programs stall.

As per one of the recent research I conducted and wrote a paper as well, the hard parts are rarely the model’s capability. They are clean data pipelines, robust usage of controls and human in the loop, traceable reasoning & observability, clear audit trails, and change management so that people trust and adopt the system. My guiding principle is deliberately narrow: build where it matters, scale only what serves outcomes and most importantly people are still your biggest enablers or barrier. Its your choice how you handle people around this technology. Replacing them blindly will not just work.

Software engineering to start with has already been heavily disrupted through Agentic AI, Finance and accounting, supply chain, and customer operations are where I see the earliest, most measurable impact—these are process-rich, data-rich domains where reasoning and action combine beautifully. In my own work, order-to-cash, collections, disputes and deductions, cash application, and supply-chain optimisation have all delivered hard savings and better customer experience.

R&D and knowledge-heavy functions will follow closely, because agents excel at synthesising vast context and surfacing the right insight at the right moment. Wherever a decision is repeatable, evidence-based, and today bottlenecked by human bandwidth, Agentic AI will make the deepest difference.

Start from the customer experience and work backwards. Tie every product decision to outcomes and OKRs, not to features or hype. And start small and big enough. In one healthcare engagement, aligning usage of Agentic AI in reasoning over complex data with business OKRs lifted customer satisfaction by 15% and delivered around USD 4 million in annual savings—because we measured what mattered and scaled only what served it.

Reliability comes from operational rigour: scenario-driven testing, strong exception handling, traceable agent reasoning, and clean audit trails. Scalability comes from strong data foundations and disciplined governance. If you get the data products and the guardrails right, value and reliability stop being a trade-off and start reinforcing each other. And never forget the power of human and process change management. If you want to achieve the value of Agentic AI, then change the processes and ways of doing things and invest in your human workforce to work collaboratively with Ai agents. That is where the value gets stuck as per one of my research paper on Agentic AI affordances.

They are the foundation - everything else is built on top. Without a strong focus on how you manage your data products and infrastructure, GenAI and Agentic AI fail to deliver. The magic never lies in using the biggest model; it lies in strong data foundations, clean pipelines, and a delightful customer experience that solves a real pain point. If you look at Agentic AI, effective context engineering and memory management is what is required for the best results and both context engineering and memory management are nothing but data products and engineering.

I often tell teams that Agentic AI is data product management wearing a new coat. Interestingly, recent NVIDIA research suggests small language models may actually be the future of Agentic AI—another reminder that thoughtful data and system design, not raw model size, is what creates durable value.

As systems gain autonomy, governance has to move from an afterthought to a design principle. I prioritise traceable reasoning so we can always understand why an agent acted, clear audit trails for accountability, and human-in-the-loop checkpoints for high-stakes decisions. Secure AI practices and ethical decision-making are themes I return to constantly in my lectures and my work. I recently wrote a teaching case that has been published in Ivey Cases and now being taught in business schools cross the world on this topic itself (https://www.iveypublishing.ca/s/product/axis-my-india-governing-a-generative-aipowered-citizen-platform/01tOF00000BXXuvYAH)

I propose a multi-tier governance model for Agentic Ai systems through guardrails in prompt engineering, controls through deterministic automations, and finally human in the loop as a reviewer. Responsible AI is ultimately about trust. An organisation only adopts what it trusts, and it only trusts what it can see, explain, and control. Robust exception handling and well-defined boundaries are not bureaucracy—they are what make autonomy safe enough to scale.

Today’s models reason impressively but do not truly learn from our interactions in real time; we compensate through context engineering, which is powerful but has limits. Reliability in complex, long-horizon tasks, consistency of reasoning, and the cost of scale all remain real constraints. We are firmly in an era of context engineering driven learning era, and not fully autonomous, self-learning LLMs driven decision-making.

The breakthrough I am most excited about is on-the-fly learning—models that adapt as we work with them, rather than being frozen at training time. Combined with stronger data foundations and, potentially, efficient small language models purpose-built for agentic tasks, that would be a genuine turning point for the field.

Five years back, a professional with great coding skills who can understand domain was more valuable than a professional with deep domain knowledge and average technical understanding. But with agentic engineering, the tables have turned. We are now in a world where deep domain knowledge becomes indispensable and english becomes the new programming language as said by Andrej Karpathy. Product leaders will spend less time specifying features and more time framing the right problems and designing the data, guardrails, and experiences around intelligent systems. As one of my professors says, research is not about answers, it is about finding the right problem—and that mindset is becoming central to product leadership too.

Technology professionals will need to be lean, fast, and experimental—comfortable learning, unlearning, and reinventing themselves as the tools evolve. They need to invest in becoming deep domain experts as well and at a fast pace. The differentiator will not be who uses the biggest model, but who best understands customers an domain at an indepth level, builds the cleanest foundations, and translates capability into outcomes people can trust.

See, Yann Lecun has left Meta to work on World models which wont be restricted by learning from content only so I see five years down the line, world models will be far more intelligent and we may hit the Artificial General Intelligence mark as that remain the north star for many of the top leaders such as OpenAI, Anthropic, Grok, and Google amidst others such as DeepSeek. In the current era, My PhD work at IIM Sambalpur excites me because it unites research and practice—studying how humans will work alongside AI agents and how businesses can extract real, sustainable value from GenAI and Agentic AI. My vision is enterprises where intelligent systems handle the repeatable reasoning and action, while people focus on empathy, creativity, and judgement.

I want to leave behind a framework that practitioners worldwide can adopt to use these technologies meaningfully and sustainably. The coming decade will favour those who are research-driven and customer-focused. When intelligence meets empathy, organisations find their way through even the most demanding digital journeys—and that is the future I am working to build.