Agentic AI is driving strong business interest despite a lack of solid return-on-investment (ROI) models, a financial challenge one Cisco networking executive said could be tempered by smart planning that takes into account advantage of small language models (SLMs) running in edge environments.

Lawrence Huang, SVP and GM of network platform and wireless at Cisco, explained that agentic AI is evolving toward a model that can allow enterprises to structure more relevant workflows for building “specific outcomes.”

“And what that means for the campus and branch, and the enterprise infrastructure ultimately, is it's less about pre-defined types of data and users that we historically have seen, but it becomes much more, call it fragmented if you will, or increase in the number and types of people and services and agents that actually need to access the network infrastructure,” Huang said.

This parsing out of real need is important for enterprises that requires a thoughtful look into what exact business challenge or outcome needs to be overcome or is desired before plunging blindly into the agentic AI hype pool.

“There's definitely use cases that are emerging that customers are interested in,” Huang said, pointing to areas around compliance and surfacing insights. “One way to do this today is you send everything to some cloud service. But, one way that you can do it with these small language models is to process things more locally to reduce the number of bits and bytes that you have to transmit to the cloud. I think there's definitely a play here for customers to gain more efficiencies, but also to get faster insights into the things that they care about within their infrastructure.”

That efficiency is increasingly important as enterprises grapple with their agentic AI plans. A recent Gartner report predicted more than 40% of agentic AI projects will be canceled by the end of 2027 due to cost uncertainty.

“Most agentic AI projects right now are early-stage experiments or proof of concepts that are mostly driven by hype and are often misapplied,” Anushree Verma, senior director analyst at Gartner, explained. “This can blind organizations to the real cost and complexity of deploying AI agents at scale, stalling projects from moving into production. They need to cut through the hype to make careful, strategic decisions about where and how they apply this emerging technology.”

Huang said he has seen this challenge in the real world.

“When I talk to some enterprises who are trying to do this, they are still so early in the process and not really sure,” Huang said. “They have an idea of what the use cases, but they get kind of confused on how to make the next steps. Most of them already have a vendor that they work with who is providing them with some sort of insight or a platform, so they're dabbling with it, but they're not really sure what that ROI is going to be on this.”

This confusion is being further pressured by overarching cybersecurity concerns, “where people are getting nervous about throwing their information out in these large pools of data.”

What does the SLM future look like?

SLMs could alleviate some of these concerns as they can be structured into easier-to-manage databases running in a localized or edge environment. This model is being propelled by more capable edge devices and what Huang said would be a “rise of more edge processing devices in the infrastructure.”

“That's what I believe as you think about the different use cases … and this is only going to explode the number of devices that IT teams need to think about,” Huang added.

These devices will challenge IT teams by requiring both a strong device-to-device connectivity that could be through advanced Wi-Fi technologies or even private networks, as well as connectivity back to a centralized cloud environment to access broader corporate policies.

Huang noted that this will require enterprises to construct networking infrastructure with a focus on wide-area connectivity, edge computing, network visibility tools for data management, architecture designed to latency requirements, and embedded security to oversee data access.

“I think there's going to be a spectrum of different deployment options out there, and it comes down to what are the latency and security requirements you have, what are the control points that you need or want to have, and I think that's going to determine if it is truly local, at the edge, closest to the data, the application, or can you move it a little bit further out to [edge environments],” Huang said. “I don't think there's a one size fits all.”

This does not mean organizations should attempt to try on all of those sizes but instead should focus on specific outcomes.

“Depending on the type of business problems you're trying to solve, it still goes back to the basics: Have I defined that business problem clearly? Do I have an architecture that I can actually build toward that? If it's just a bunch of random experiments, if I'm going to bring in 20 different tools to play with, that is not a recipe for success,” Huang said. “I think what we're going to see over the coming year is that more and more companies are going to find that the hodgepodge approach to these proof of concepts is honestly just wasted effort and time, and they're going to get smarter about it. And I truly believe that people are adaptable, and they can figure out how to do this better.”