The Math Driving Agentic AI Adoption in NetOps – Harvard Business Review

Every day, enterprise monitoring systems generate more than 4,000 alerts, about half of them concerning the network. An experienced engineer can investigate and resolve about 20 alert events a day. Clearing that queue by hand would take roughly 100 IT specialists—far more than most companies employ—so nearly half of alerts are closed without investigation.
This isn’t the sign of a bad team; it’s the sign of a team that can’t keep up with an increasingly complex network that is generating alerts faster than the teams can respond. For years, network operations (NetOps) teams have been expected to deliver stronger performance while managing a growing mix of applications, devices, dependencies, security controls, and telemetry. That mounting complexity is compounded with AI and agents, driving teams to look beyond manual operations.
A 2026 study by Omdia and Cisco of 1,000 IT and NetOps leaders finds that more than half of their organizations already use agentic AI to act on NetOps processes and tasks in production. Networks change constantly, feedback arrives within minutes, and mistakes ripple visibly across security, applications, and customer experience. “AI workloads are bringing new traffic patterns, increasing the data exchange, and also creating more dependencies across the environment,” says Joe Vaccaro, Cisco’s senior vice president and general manager of Network Platform and Assurance.
Under those conditions, the math leaves little room for any other answer: Agentic AI is no longer optional but necessary. As network operations teams hand real authority to AI agents, though, the challenge now is to decide what to trust the agents with, how to verify their work, and who pays for it.
The Risks of Doing Nothing
When most executives evaluate AI autonomy, they tend to tally the risks of acting and treat the status quo as neutral. NetOps executives see AI autonomy differently. The failure to automate isn’t simply an oversight for NetOps decision makers. Without it they anticipate more security gaps and risk, increased frequency and severity of outages, and increasing numbers of failed AI initiatives.
These fears describe the present as much as the future: Teams already spend 45% of their investigation time chasing false positives, while nearly half of network alerts go unexamined. An agent making a bad call is a real risk, but so is leaving a genuine threat sitting unread in a queue of thousands of alerts.
The Long Tail of Network Incidents
While most network incidents may be manageable, a handful of more complex incidents can have an outsize impact on resolution times. They can become multi-day or multi-week ordeals and could bring business operations to a halt. One in 10 respondents say incidents typically take more than a week to resolve.
These drawn-out incidents rarely stay in one domain. Survey respondents say performance problems commonly span multiple domains, such as networks, cloud, security, and applications. A latency complaint might originate in a cloud misconfiguration, surface as an application slowdown, and trigger a security alert all at once.
Two implications follow. First, a business case focused only on making routine fixes faster could miss the value of shortening those lengthy investigations. Second, visibility should be a condition for giving an agent authority. An agent that sees only one part of the environment might fix the symptoms in front of it while making the underlying problem worse.
Before granting an agent authority, decision makers should determine whether an agent can see enough of the system to understand the consequences of its actions.
Governance Built for Stable Systems Is Breaking
Change review boards and sign-off procedures assume that systems change occasionally. But networks now change constantly. Close to a third of surveyed organizations make multiple changes to their production networks every day.
Executives are responding to this challenge not by abandoning oversight but by unbundling it. While most still want humans to approve agentic actions, the overwhelming majority are comfortable letting AI make production changes without permission for at least some categories of action, such as rerouting traffic, adjusting wireless parameters, or isolating suspicious endpoints. Decision makers grant authority action by action much as a manager gives a new hire more latitude as their track record builds.
The design challenge is subtle. If every action requires human sign-offs, the alert queue simply becomes an approval queue. Oversight itself must become partly automated, with policies settling in advance which actions an agent takes alone, which require a person’s decisions, and which are off-limits to agents.
“Agentic autonomy doesn’t mean transferring accountability to the system,” Vaccaro says. “The more authority we give these systems, the more important it becomes to know what they did, why they did it, and when to step in.”
At modern furniture and home decor retailer Room & Board, a three-person network team supports 34 locations, delegating routine monitoring, radio resource management, and logging to agentic platforms. Senior network engineer Mark Rodrigue is emphatic about the terms: Checking an agent’s work must cost less than doing the work.
“Show your work,” Rodrigue says. “I can’t emphasize that enough. We don’t really have the time and resources to spend managing management tools. If your tools require a full-time babysitter, you’ve already lost.”
Toil Is a Talent Problem, Not Just an Efficiency Problem
Most decision makers say alert volume prevents their teams from doing other critical work. And experienced engineers are scarce and expensive to replace; endless triage is a retention risk.
Rodrigue’s team is navigating workforce constraints even as Room & Board’s network grows, and he describes agentic AI as “a force multiplier.” The most AI-mature organizations in the study report close to 30% higher productivity per practitioner. The more durable return may lie in what experienced people stop doing and what they are free to do instead: designing guardrails, building systems, and solving problems that require human judgment.
Practitioners See What’s Coming First
NetOps practitioners are twice as likely as line-of-business executives to see AI’s role shifting from advising to acting. But half of those business leaders still view AI in operations mainly as a source of insight rather than as an agent of autonomous action. And the line-of-business executives often control the budget.
That gap leaves technical leaders with a business case to make, not just a technology to deploy. “How do we engage with technology in a way that allows us to maintain that efficiency and leanness but still keep up with the demands?” Rodrigue asks.
NetOps teams are not waiting for a perfect framework before delegating action to agents. The alternative is to leave thousands of alerts unanswered.
What they have instead is a working framework for deciding when agents act. They start by asking what is already going unexamined, focus on where the hardest problems concentrate rather than where the numbers look good, insist that verification cost less than the work itself, and push budget holders to understand that AI now acts as well as advises.
On these terms, teams don’t hand over autonomy—the AI agents earn it, one action at a time.
For more information on how AI is impacting network operations, read here.
Harvard Business Publishing:
Copyright ©2026 Harvard Business School Publishing. All rights reserved. Harvard Business Publishing is an affiliate of Harvard Business School.

source
This is a newsfeed from leading technology publications. No additional editorial review has been performed before posting.