Would you let an AI agent change your company’s network before a human approves it? In new Cisco and Omdia research, 82% of surveyed organizations said they are comfortable with AI making at least some production network changes before a person signs off. Another 51% said they already run agentic AI that acts in production.
The Cisco report, announced September 23, suggests network operations could become an early test case for enterprise AI autonomy. Some 84% of respondents expect an AI-led operating model within 12 months.
Omdia independently surveyed 1,000 IT and network operations leaders at organizations with at least 500 employees across three regions. The results describe reported deployments and willingness to allow autonomy. They do not establish how reliably those systems perform in production.
Network teams have more work than they can clear
The full Cisco and Omdia report describes organizations that generate roughly 4,100 monitoring alerts and events daily, more than half network-related. The report estimates that manually clearing the daily network-alert backlog would require about 100 specialists.
Nearly half of network alerts are closed before anyone investigates them, while 57% of respondents said their change processes cannot keep pace with required updates. These pressures help explain why companies are willing to delegate work to agents.
Network operations also contain tasks with measurable outcomes. If an agent reroutes traffic to resolve a performance problem, operators can check the result and reverse the change if necessary.
Joe Vaccaro, Cisco’s SVP/GM of Network Platform & Assurance, argues that this ability to test and reverse an action should determine an agent’s authority.
“Machine-to-machine activity changes the network in a fundamental way,” Vaccaro told The Neuron. As agents continually contact models, tools, data sources, and other agents, he argued, “the network becomes part of the AI runtime.” A delay or failure can affect the context an agent receives and the action it takes next.
Autonomy depends on what the agent is allowed to change
“I wouldn't think of it as one fixed line,” Vaccaro said. Companies should consider the consequences of an incorrect action. They should also assess whether the result is predictable and how quickly the system can undo it.
Rerouting traffic around a verified performance problem is one example. Vaccaro said an agent could act independently if it shows supporting evidence and operates within policy. It would then measure the result and roll back the action if performance did not improve.
Major topology changes or security-policy updates require greater human involvement because the consequences may spread further or prove harder to reverse. Vaccaro’s approach assigns permissions by activity and domain, with risk and the approved change window also determining authority.
The survey reflects similar conditions: 69% require explanations for agent-driven actions. For 36%, the minimum is full observability, which includes tracing and a summarized rationale, plus post-action audits.
These controls document what the agent saw and changed, while policy limits define where its authority ends. The same question appears in the wider debate over who controls an AI agent’s next move.
AI agents are also creating more network traffic
AI agents consume the same infrastructure they may increasingly help manage, and Vaccaro describes the network as becoming “part of the AI runtime.” Agents communicate continuously with models and data sources, so network conditions affect when information arrives and whether it remains relevant when an agent acts.
Cisco’s telemetry analysis puts direct-to-AI traffic on a trajectory to double every six months. Separate Cisco traffic testing found that agent-performed tasks can generate up to 450% more total network traffic.
The 450% figure comes from Cisco testing; it is neither an Omdia survey result nor a universal increase across agent workloads. Cisco also says AI inference remains a relatively small traffic category compared with video.
Even so, enterprises may face a growing operational burden: they are adding agents that increase network activity, then considering other agents to manage the additional workload.
More agents could recreate the problem they are meant to solve
Giving every network tool its own agent could reproduce the fragmentation NetOps teams already face, Vaccaro warns. Problems often cross operational domains, so agents need shared context.
"The deeper change is that the network becomes both infrastructure for agentic work and a source of real-time context about that work," he said. "Agents add complexity, while other agents help operators manage it. The risk is deploying isolated agents in every domain and recreating the fragmentation enterprises already face. That’s why shared context matters."
The survey found that 92% of respondents encounter performance issues across multiple domains, while teams use around ten tools for visibility. An agent that sees only part of the environment could make a decision based on incomplete information.
Cisco has a commercial stake here: its AgenticOps strategy promotes cross-domain management through Cisco Cloud Control. The research supports a market Cisco wants to serve.
NetOps may offer an early view of how enterprise AI autonomy develops. Companies appear willing to delegate narrowly scoped actions whose results can be checked and reversed, with audit records available for review. The harder test comes when consequences spread beyond one task and operators must decide how much authority software should have over critical infrastructure.
