Generative AI may be having a moment, but in enterprise IT, the success rate tells a different story. Gartner estimates that by the end of 2025, 30% of generative AI projects will be abandoned after proof of concept due to security risks, data quality, or costs with unclear business value.
That failure rate isn’t surprising to Kshitij Mahant, Senior Manager of Technical Marketing at Cisco. “Too many AI projects in networking collapse under their own weight,” he says. “Implementation needs to be a deliberate, careful process.”
Mahant’s perspective comes from more than a decade of translating Cisco’s enterprise networking strategy into functional, deployable solutions. He’s worked on everything from wireless reliability to SD-WAN assurance and is a Senior Member of IEEE, a distinction recognizing his contributions to the field of engineering and enterprise networking, including applied AI in network systems and infrastructure optimization. His central argument: AI needs to make its value obvious. If the payoff is buried in a dashboard or deferred to a future release, it’s already failed.
The Networking Case Study
Enterprise wireless networks now serve far denser environments than they were ever designed for: multi-floor office buildings, hospitals packed with critical life saving devices, and universities with thousands of concurrent users. Static configurations and manual tuning, the long-time default for network engineers, are no longer sufficient in high-interference conditions.
And the cost of this outdated model adds up quickly. Lost productivity, dropped calls, application timeouts, and user complaints are as expensive as they are nuisances. The resulting instability can cost businesses tens of thousands to millions per hour.
Compounding the issue is a mistaken belief that networking issues will fix themselves with the next hardware refresh or software patch, bundled with the latest so-called intelligent, self-sufficient AI. “That model might work for consumer technology,’” Mahant says. “But enterprise networking doesn’t run on autopilot. It really demands coordination between people, processes, and the tools they’re using.”
Focusing on the Problem
Instead of chasing a one-size-fits-all AI solution, Mahant’s team went narrow. They focused on one specific, high-cost issue: radio interference in wireless networks. The result was AI-Driven Radio Resource Management,a targeted system trained on real deployment data to dynamically manage channel selection, signal strength, and power levels in enterprise networks.
By narrowing the problem space, the team avoided the typical pitfalls of general-purpose AI: expensive compute needs, ambiguous outcomes, and poor integration with existing systems. Instead, they shipped something verifiably effective.
“Having smart algorithms is great,” Mahant says. “But the real test is whether they stop you from getting that 2 a.m. call. When interference drops and complaints go silent is when AI starts to feel real.”
Across deployments on over 200,000 access points and more than a million client devices, AI-RRM delivered notable improvements: an 80% drop in disruptive configuration changes, a 65% reduction in interference events, and an 8 dB improvement in signal-to-noise ratio.
Visibility Is the Value
AI-RRM’s success offers a lesson that extends far beyond consumer networking and into the product development space: AI has to make its success communicable. Stakeholders need to see, in plain terms, how AI is solving a costly problem. That clarity is often what separates a successful product from another abandoned pilot.
“Getting the tech right is only half the equation,” Mahant says. “The value has to be legible to the business.” That clarity translated directly into revenue: more than 85% of customers opted for Cisco’s premium AI-enabled tier because they could understand what it delivered.
Compare that to many AI offerings where benefits are speculative or abstract. Without a tightly scoped problem and a direct line of sight to resolution, even technically sound projects stall—first with hesitation, then with abandonment.
Building a Smarter Baseline
Deploying systems like AI-RRM sets the stage for more ambitious automation down the line. Once foundational issues like interference and signal quality are addressed, it’s possible to layer in tools like AI-based deployment assistants, NLP troubleshooting interfaces, or even commercial-grade self-healing SD-WAN functionality for consumers. These systems fall under the broader umbrella of AIOps, a space projected to exceed $32.4 billion in market size by 2028.
Still, Mahant cautions against overpromising. “I think we’ve matured past the point where we treat AI as a catch-all solution,” he says. With investor enthusiasm cooling and buyers growing more skeptical, his message is well-timed: “At this point, effective implementation should speak for itself.”
