Editor's note: This is the second of a two-part series on autonomous network operations. Read part one, "IT leaders test the waters of autonomous network operations."
Enterprise IT organizations expect significant near-term growth in agentic network operations as they cope with AI-driven upheavals, though market research differs on how quickly and how far agentic autonomy will advance.
Of 1,000 IT and network operations leaders surveyed by Cisco and Omdia, a division of Informa TechTarget, 84% said they expected to "reach an AI-led operating model" within 12 months, according to a survey report released last month. A Cisco exec described an AI-led operating model as one in which AI agents conduct initial triage of network alerts and, with human approval, take action on them, in a TechTarget interview this month.
Looking further ahead, 85% of respondents said they expect agents to manage most operational tasks in five years – and responses predicted that specific operational tasks will increasingly be handled by agents in two years.
For example, while 18% of respondents currently use agentic AI to take remediation actions, 36% expect to do so over the next 24 months. Similar increases are expected for 14 other specific NetOps tasks, including performance optimization, system updates and changes, application and network troubleshooting, and capacity planning.
"The types of use cases that I've heard of really run the gamut from very simple things like resetting an interface or restarting a device that seems to have hung up," said Jim Frey, an Omdia analyst who led the survey, in a September interview. "At the other end of the spectrum is … [making] a major configuration change or setting up a new VLAN where you've got to distribute the definitions of that across a whole bunch of devices."
IT learns to stop worrying and embrace AI agents
Comfort with agentic network operations is also growing significantly, according to the Cisco/Omdia report. While 56% currently favor approaches where a human approves actions before AI takes them, and 24% favor full autonomy with no human approval, 80% said "they are comfortable with AI taking a high or fully autonomous role in NetOps today," the report reads.
"82% are comfortable letting AI make changes to their production networks without seeking human permission first for at least some categories of action," the report continues, "including rerouting traffic, adjusting wireless parameters, isolating suspicious endpoints, and resolving incidents end-to-end."
Research from another analyst firm, Enterprise Management Associates (EMA), also points to growing adoption of autonomous network operations, though without a specific time frame and in smaller numbers than the Cisco/Omdia survey reported.
"Cisco's view of agentic, autonomous operations is a bit more aggressive than what I'm seeing, but not by much," wrote EMA analyst Shamus McGillicuddy in a September email.
The mid-year EMA survey of 356 IT professionals about unifying IT observability tools, including those for network operations, asked, "Which AI systems or interfaces do you anticipate will consume your organization’s unified IT observability data?”
In response, 55.6% of survey participants chose "Agentic AI systems that take autonomous action based on observability signals." One early adopter of HPE's Self-driving Networks said his teams still use a copiloting approach when rolling out agents, but he can envision a point where agents handle that stage autonomously as well.
"If we are able to define QA criteria well enough for a human being, we can also do it for agents," said Sune Tornbo Baastrup, senior vice president and CIO at Danfoss, a global industrial manufacturing company based in Denmark. "And if we can provide enough learning material for the agents, that will also be something we can [add to] our change management and release process."
AI traffic ratchets up network pressure
Already, enterprises face an alert volume from fragmented network and infrastructure management tools that "cannot be staffed" by humans, according to the Cisco/Omdia report, which estimates the average organization generates around 4,100 monitoring alerts and events per day, 51% of which are network-related.
"This is no longer a staffing problem that better recruiting can solve," according to the report. "It is a structural mismatch between the volume of information environments produce and the capacity of humans to process it."
The problem stands to worsen with increased traffic volume from agentic AI apps on enterprise networks: Cisco testing has shown that AI agents can generate up to 450% more network traffic than human-driven applications.
"[AI adoption] is just going to continue to make the network more important, from both a guardrail enforcement perspective as well as an availability and performance perspective," said Mark Rodrigue, senior network engineer at Room & Board, a furniture retailer in Golden Valley, Minn. "It's already looked at as a utility. We have electricity, we have water, we have network connectivity. There's no difference, from a business perspective … and tolerance for performance-related or outage-related issues is just going to continue to decrease."
Adding to the pressure, cloud-native networking is already complex, and AI agents running on corporate networks will compound that complexity to the point where it will be "mind-numbingly hard for people to wrap their arms around," said Rob Strechay, lead analyst at VB Intelligence.
"AI just made it that much more difficult, especially when you talk about the different silos of automation that are going to have to be connected together in some way, shape, or form that weren't before, or isolated with harder boundaries to limit the blast radius [of issues]," Strechay said.
At the very least, AI agents will be used to analyze large, complex networks and propose architectural plans that account for details humans would likely miss, Strechay said.
Another AI-driven trend is a wave of software security vulnerabilities being newly uncovered by AI models or newly exploited at scale by agentic attacks, which puts patching at the top of the list for autonomous network operations at Room & Board, Rodrigue said.
"Once upon a time, quarterly patching was good enough. Quarterly patching is not good enough anymore," he said. "But monthly patching and maybe even weekly patching, that's just not sustainable. To be able to do that, we really need that bug scrub, that code evaluation, the upgrade, the validation, and if needed, the rollback to be handled by agents."
Obstacles to autonomous network operations at scale
Despite his optimism about autonomous network operations, Baastrup said there are limits to how broadly he'd trust agentic orchestration, particularly across different levels of incident investigation, from frontline response to and third-line forensic analysis.
"Segregation of duties would prevent you from giving a human access rights across those things. So, do we want to orchestrate that at a higher, aggregated level? Probably not," Baastrup said.
"Just like, even though I lead all of our IT, I don't have access to all of our data centers because we have a strict policy that you cannot access primary and secondary nodes and backup locations. You can access one of those, but not all three of them. That's the same thinking I think we need to develop around agents," he said.
Baastrup said he has also been unimpressed by agentic orchestration platforms that aim to centralize and coordinate agents' workflows across an entire enterprise – one reason Danfoss has still kept humans in the loop for QA during agent rollouts. .
"What most of the vendors at the moment actually try to sell you is that they can do that quality assurance on agents, but what they really are is more like an inventory of where the agents are and what they are used for," he said. "But at least to my knowledge, also being part of the Gartner Research Board … there are not a lot of platforms that can do this across use cases from an agent perspective."
Fragmented tools and data pose another obstacle to autonomous operations, according to EMA's observability report:26.1% of respondents said AI requires unified data, and that fragmentation of systems severely limits its effectiveness, and another 64.3% said AI works better with unified data, though it can still deliver value across fragmented tools.
Technical obstacles aren't the only barriers to autonomous network operations. The Cisco/Omdia study also found that IT teams' growing comfort with agentic autonomy might not extend to the business leaders who influence budgets.
"Practitioners, who feel the operational strain daily, see the transition [to AgenticOps] clearly," according to the survey report. "Line-of-business leaders with budget influence are far more likely to view AI in operations as a matter of insights and recommendations, the older 'AIOps' paradigm, rather than agent-driven autonomous action."
Beth Pariseau, Lead Editor, Infrastructure News at Informa TechTarget, is an award-winning veteran of IT journalism. Have a tip? Email her or connect on LinkedIn.
