Cautious Optimism Drives The Tone of Agentic AI Conversations

If there’s one term that’s all over the place – from product roadmaps, to boardroom conversations, to coding tools and even predictions about the future, it has to be ‘Agentic AI’. With IIITH’s TechForward series approaching its second anniversary, the institute hosted a roundtable to take a closer look at the ‘Agentic AI’ theme that has consistently drawn attention across several editions and to summarize where the broader AI conversation needs to be.

If you use the term ‘agent’ interchangeably with ‘bot’, you’re not alone because there exists some ambiguity between the two and it’s not just semantic. While all AI agents are bots, not all bots are AI agents. Unlike chatbots that simply respond to prompts, agents can reason, plan, use tools, remember context, and execute tasks with a degree of autonomy. That opens the door to automating complex enterprise workflows, assisting software development, and creating systems that behave more like collaborators than interfaces. But according to Prof. Karthik Vaidhyanathan whose research group has published over 20 papers on Agentic and GenAI in the last three years at the Software Engineering Research Centre in IIITH, the term itself is widely misunderstood. “It is neither a chatbot nor a model. Agentic AI is more of an ecosystem,” he says. In other words, the challenge is no longer about just building better language models. It’s designing entire systems around them: memory, tools, governance, evaluation, safety, and human oversight. 

Agentic AI has received significant attention in IIITH’s TechForward series because this is viewed as one of the next major transitions after generative AI for both academia and industry. To separate the signal from the noise, researchers at IIITH convened an industry-academia roundtable in March 2026. Thirty practitioners from financial services, healthcare, manufacturing, pharmaceuticals, and software joined faculty to wrestle with three deceptively simple questions: What really is agentic AI? What are the real-world use cases beyond chat? And are enterprises actually ready? 

The Autonomy Slider
One of the biggest misconceptions surrounding Agentic AI is that it is a new kind of language model. But it is more than just that. Agentic AI is best understood as an engineered software system that combines models, tools, workflows, memory, external environments, and human oversight. While popular narratives often portray AI agents as fully autonomous entities capable of replacing human workers, enterprise leaders, however, are taking a more measured approach. Prof. Vaidhyanathan lets us ponder over a hypothetical scenario concerning a financial workflow. An agent may be capable of identifying investment opportunities, analyzing risk, and recommending actions. But should it be allowed to execute transactions independently? “Most organizations would say no,” he states.

Enterprises are instead increasingly adopting a model where agents perform analysis, propose actions, and automate routine tasks, while humans retain authority over critical decisions. “Think of it as an autonomy slider rather than an on-off switch. An agent can draft an important email but not send it. It can recommend a supplier but not approve the purchase. It can assess a loan application but not make the final lending decision. This balance between automation and oversight is rapidly becoming a cornerstone of enterprise Agentic AI adoption,” says Prof. Vaidhyanathan. 

Real World Use Cases
While public attention often focuses on conversational interfaces, organizations are already deploying Agentic AI in more operational settings. During the roundtable discussions, participants highlighted use cases across multiple industries such as banking agents that assist with collections, customer outreach, and fraud detection; healthcare systems that reduce prior-authorization processing times from nearly an hour to a matter of minutes; insurance workflows that streamline document handling and revenue-cycle management; manufacturing and procurement systems that analyze inventory, identify dead stock, and support supplier decision-making and software engineering agents that assist with requirements analysis, testing, verification, maintenance, and architectural documentation.

What makes these examples significant is that they are not standalone chatbots. They are embedded within business processes, helping organizations automate complex workflows while keeping humans in control where it matters most. As one participant observed, the greatest value may not come from doing existing work faster, but from enabling decisions and analyses at a scale that human teams simply cannot achieve on their own.

The Problems Evaluation
If Agentic AI sounds like such a panacea, why aren’t all enterprises rushing to embrace it? The answer lies in the fact that the most difficult challenges have little to do with model performance and everything to do with engineering. Traditional software testing assumes deterministic behavior. If the same input produces the same output every time, validation is straightforward. Agentic systems however do not work that way. An agent may successfully complete a task while using the wrong tool, relying on outdated information, or taking an unsafe path to reach the result. This has prompted researchers to rethink evaluation entirely.

Recent work by IIITH and industry collaborators proposes moving “beyond task completion” by assessing not only outcomes but also how agents use tools, memory, reasoning, and environmental context throughout execution. The challenge is no longer asking, “Did it work?” The more important question becomes, “Did it work correctly, safely, and efficiently?”

Trust and Accountability
Organizations understand how to hold humans accountable. They are still figuring out how to hold AI systems accountable. In highly regulated sectors such as banking, healthcare, and life sciences, explainability is not optional. Regulators, customers, and stakeholders need visibility into why decisions were made and who remains responsible when errors occur. The reality is that enterprises are unlikely to hand over complete control anytime soon. Trust must be earned and engineered.

Data Readiness
Many organizations are discovering that their biggest AI challenge is not AI at all. It is data. Agentic systems depend on accurate, accessible, and contextually meaningful information. Poor-quality data does not merely cause hallucinations; it undermines entire workflows. As several roundtable participants noted, successful Agentic AI initiatives often begin with improving data foundations before deploying agents.

Cost and Sustainability
Another recurring concern is cost. Most enterprise agents today rely on large language models from providers such as OpenAI, Anthropic, and Google. As adoption scales, so do token consumption and infrastructure expenses. Organizations are increasingly tracking AI usage, managing token budgets, and evaluating whether every task truly requires a large model.

Researchers are also exploring hybrid architectures that combine smaller, specialized models with larger ones to improve efficiency while reducing cost. Beyond economics lies another important consideration: sustainability. As AI adoption accelerates, so does demand for compute infrastructure and data centers. Enterprises are beginning to ask not only whether a system works, but whether it is environmentally sustainable at scale. 

“The technology is definitely moving faster than organizational readiness and a fool with a tool is still a fool,” notes Dr. Vaidhyanathan. Upskilling, therefore, may prove just as important as technology adoption.

Skills and Organizational Readiness
Perhaps the most overlooked challenge is people. Many organizations face pressure to become AI-enabled quickly, yet their workforce may not be prepared for the transition. Building and using Agentic AI systems requires new competencies in systems thinking, evaluation, governance, and AI-assisted workflows.

The Other Side: Vibe Coding
The Agentic AI conversation is not limited to enterprise workflows. A parallel transformation is happening inside software engineering itself. Tools such as GitHub Copilot, Codex, Gemini, and Cursor are increasingly enabling developers to describe what they want and have AI generate much of the code. This trend, often referred to as “vibe coding,” shifts the developer’s role from writing syntax to defining intent.

The implications are profound. Developers may spend less time coding line by line and more time reviewing, validating, architecting, and thinking strategically about systems. The future software engineer may resemble an architect more than a traditional programmer. But there is a catch. AI-generated code still requires human judgment. As Dr. Vaidhyanathan puts it, “If you are more skilled, these tools can make you even more productive. But if you do not understand the fundamentals, they can become dangerous.” The lesson is clear: AI amplifies expertise; it does not replace the need for it.

The challenge for enterprises is not deciding whether Agentic AI matters. It is learning how to harness its possibilities while responsibly managing its risks. The future will likely belong neither to organizations that dismiss the technology nor to those that blindly embrace the hype. It will belong to those that understand both. “We need to look at a marriage between possibilities and risk and take it from there. My only request to the community is to be mindful of the cost and the sustainability angle, which if ignored now will come to bite us later,” concludes Dr. Vaidhyanathan.

Between Hype and Reality
The roundtable’s most important takeaway was not that Agentic AI will replace people, nor that it is simply another technology fad. The reality lies somewhere in between. Organizations are already adopting Agentic AI in meaningful ways. Real productivity gains are being reported. New workflows are emerging. Entire categories of software are being reimagined. At the same time, critical questions around evaluation, governance, trust, accountability, operations, and sustainability remain unresolved. Perhaps the most useful perspective is to view Agentic AI not as an employee, but as software, albeit software with new levels of autonomy and complexity.