Agentic AI’s Promise and Perils

Creators, business leaders, and researchers convened at the 2025 CATT Summit to explore the hopes and cautions around emerging technologies

Agentic AI, an evolution of artificial intelligence that promises more autonomous AI systems (dubbed “agents”) that will more ably manage systems and learn from mistakes, is a technology so promising that companies from Microsoft to Google to Meta to Amazon are spending billions to harness its power in a tech arms race.

But because the technology is still emergent, there are many questions about how these powerful systems will be managed and governed, what value they’ll bring, and how to turn their potential into production.

These topics and many more were discussed at the 2025 Agentic AI Summit, a two-day virtual event hosted by the McCombs School of Business’ Center for Analytics and Transformative Technologies (CATT). CATT’s annual analytics event — which has previously focused on areas such as ethics and AI and differential privacy — had its most successful outing yet, with more than 9,700 registrants and participation that drew an average of 100 questions per session for its panelist mix of researchers and industry practitioners.

Michael Sury, CATT’s managing director and the host of the summit, said the event’s topic comes at a time when researchers and those in the corporate world are building use cases for agentic AI and experimenting. “Now the question is: Can we put the pieces of that puzzle together in ways that add tangible value for organizations?” Sury said.

The implications for answering that question are wide-ranging, explored through the conference’s workshops, panel discussions, and live technology demonstrations. While each session covered specific aspects of agentic AI use or research, multiple themes became evident across the event.

Embracing unexpected technology

CATT faculty director Kumar Muthuraman started the conference with this foundational truth: Agentic AI wasn’t a goal of technologists, but a byproduct of large language models (LLMs).

“AI agents were not something we set out to build,” Muthuraman said in the preconference workshop on the technology’s history. “They were an unplanned inheritance — and like any unplanned inheritance, they come with enormous risks.”

AI agents, and the kinds of generative AI tech such as ChatGPT that led to them, have become so rapidly embraced and adopted that they’ve come to define machine-learning technology that has been around for decades. “AI meant many things for decades, and within a year it came to mean just one thing: large language models,” he said.

Enterprise and AI agents: harder than it sounds

While agentic AI could be as revolutionary and game changing for business as the internet was, businesses are struggling to justify the investments they’re making while ensuring that the technology is scalable, governable, and using reliable data.

But many are finding that agentic AI introduces tools and frameworks that can be difficult to manage and that the base language models are not always the right fit for solving some problems.

KamalikaDas, principal AI researcher at Intuit, said in her keynote session on smart Model Context Protocol, “Large language models are not reliable calculators, tax engines, or accounting systems. In high-stakes domains like finance, agents must know when to defer to deterministic software. If your tool descriptions are weak, no amount of prompt engineering will save you.”

Others panelists pointed out that some companies are risk-averse and reluctant to invest in agentic AI use cases if they can’t foresee a quick return on investment.

Tools and techniques for using agentic AI explored

Companies that are embracing these challenges demonstrated how they’re doing it.

Sury hosted a panel on agentic AI in the enterprise with representatives from Dell Technologies, Alphabet (Google), Zebra Technologies, and Applied Materials that detailed the different ways each company is approaching and using the tech. The panel discussion included the challenges of defining metrics of success and measuring ROI.

TransUnion’s senior vice president of data science, Robert Stratton, discussed automation, and engineers from Amazon Web Services working on the open-source Project Jupyter — which is developing a suite of agentic AI tools — illustrated its capabilities with an invitation to use and contribute to the project.

Rohan Kodialam, a co-founder of Sphinx AI, demonstrated how his company’s work can help AI recognize data in ways that are not intuitive to LLMs. He stressed that these are early days in AI development, despite the investments and hype. “We are very early in the curve of scaling,” Kodialam said, “even as the conversation feels overheated.”

The ‘vibe coding’ era

One of the liveliest presentations at the summit involved a recent phenomenon so popular it was named the 2025 Word of the Year by the Collins Dictionary: vibe coding, or using natural language descriptions (typically spoken aloud) to generate code.

Panelists from companies including Informatica, PathPilot, GiddyUp Data, and Wisary, and visiting Imperial College Business School fellow Michael Schrage talked about their experiences vibe coding, but the talk quickly turned to discussion about the need for humans to stay in the AI loop even as automation comes to the fore.

“I think things that will remain human are intention, critical thinking, and giving direction to all these kinds of AI tools,” said Ala Stolpnik, founder and CEO of Wisary.

It echoed a welcome talk at the start of the summit from McCombs Dean Lillian Mills, who said that in the age of AI’s dominance in the discourse, McCombs is committed to that type of stewardship. “Business schools have a responsibility not just to teach the technology, but to help leaders understand its consequences,” Mills said. “Our role is to prepare leaders to ask better questions — not just deploy faster tools.”

Story by Omar L. Gallaga