For network operations teams that have long relied on scripting and automation, handing control to AI agents may seem like the logical next step. Yet most network teams remain cautious about ceding that control, less because of what agents can do than because of whether operators trust them to do it.
Recent survey data supports this. According to EMA’s AI-Driven NetOps research, based on a survey of 458 IT professionals, only 35% of enterprises say they have been completely successful applying AI to network management. Only 31% completely trust the AI tools their organizations use on the network, and EMA found that weak trust tracked closely with weak uptake among operators. The correlation is unsurprising, since engineers are unlikely to delegate work to a tool whose output they feel compelled to verify by hand.
Autonomy Arrives in Stages
The telecom industry has already given this problem a vocabulary. TM Forum’s autonomous network levels run from L0 (fully manual) to L5 (fully autonomous), and Nokia Bell Labs Consulting reports that most service providers sit between L1, assisted operations, and L2, partial autonomy. Enterprise networks aren’t carrier networks, but the lesson carries over: autonomy is something an operation grows into, not a feature switched on at install.
In practice, agentic network operations tends to unfold in three stages, each a bigger commitment than the last:
- Recognize. The agent observes, correlates and explains—spotting the firewall pair that has drifted from its baseline or the link that keeps flapping, and showing the evidence behind its conclusion. Humans still make every change.
- Act. The agent proposes or executes specific fixes on specific devices within narrow, preapproved boundaries, with a human approving the change or able to roll it back.
- Assure. The agent handles change at scale (the same configuration fix across 400 devices before a compliance deadline, for example), verifying state against a declared, known-good baseline before and after every change.
Organizations that attempt to move directly to the third stage take on considerable risk. Gartner predicts that more than 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls.
Trust Doesn’t Require Perfection
One of the more counterintuitive findings in EMA’s research is that trust and accuracy aren’t the same thing. Many respondents who trusted their AI tools told EMA they knew those tools made mistakes regularly. What mattered was whether the mistakes were easy to catch. When asked how vendors could earn their trust, respondents cited transparency about how the AI was built (56%), insights tied directly to the underlying data (50%), visualizations of the AI’s reasoning (47%) and manual confirmation workflows (45%).
Translated into operational requirements, that means a few things have to be in place before an agent moves from one stage to the next:
- Data the team already trusts. An agent’s judgment is only as good as its view of the network, and only 36% of IT professionals fully trust their network monitoring data, according to EMA’s Network Management Megatrends 2026 research. That problem predates AI, and automation only amplifies it.
- Verification on both sides of every change. Uptime Institute’s 2025 outage analysis found configuration and change management failures behind 50% of major network-related outages. An agent that pushes changes faster without validating them will simply produce those outages faster.
- A feedback loop. Every incident an agent diagnoses, and every fix a human approves or rejects, should improve how the next one is handled, moving the operation from reacting to problems toward preventing them.
- A complete audit trail. Every action should be traceable to the agent that took it, the data it relied on and the policy that allowed it.
Agents Are Privileged Users, Too
Trust also has a security dimension that network teams should address from the outset. An agent that can rewrite router configurations is, functionally, a privileged user, and it should be governed like one—with its own identity, permissions scoped to the stage it has earned and every session logged. The OWASP Top 10 for Agentic Applications, published in December 2025 with input from more than 100 security practitioners and researchers, catalogs what can go wrong when software acts with delegated authority, including goal hijacking, tool misuse, identity abuse, cascading failures and rogue agents. Each of those risks grows with the agent’s level of autonomy, which is one more reason to expand that autonomy deliberately.
The fastest route to autonomous network operations, then, is a measured one. Teams that let agents prove themselves in stages, starting with watching, then acting, then assuring, will end up handing over more of the network sooner than teams that attempt to skip stages and then absorb the cost of the failures that follow. Trust, like uptime, is earned one change at a time.
