As telecom operators explore the use of agentic AI, one challenge is becoming clear: autonomous agents that pursue different or competing goals need coordination. Abe Nejad, Publisher of The Network Media Group moderated a panel discussion on this issue, bringing together industry perspectives on how operators can safely use agentic AI in support of moving to autonomous networks.
Speakers
Watch the full interview.
In mission-critical voice networks, autonomous AI agents cannot operate unless service requirements are clearly prioritized and have no competing objectives. A cost-optimization agent may try to reduce compute resources during low-traffic periods, while an assurance agent may simultaneously respond to KPI spikes or call-drop risks. In networks that carry emergency 911 traffic, those conflicts are unacceptable because packet loss and degraded voice performance are not tolerable outcomes. The central requirement is that agents must be coordinated around non-negotiable service-level agreements, understand when to act, and avoid creating a tug-of-war that could compromise essential network services.
Effective guardrails for AI agents need to operate across three distinct layers: intent, policy, and action. At the intent layer, organizations must define which agents are allowed to request specific outcomes and what types of instructions the system should accept. At the policy layer, they need a standard policy engine that can interpret those intentions in the context of the operator’s network, business rules, and operational priorities. At the action layer, agents must work within a controlled environment rather than having unrestricted freedom to execute changes. A feedback loop back to the operator is essential, allowing each layer of control to reinforce the others and ensuring that agent behavior remains aligned with operational expectations.
Operators are likely to grant AI agents more autonomy only in the few scenarios where the expected response is already well understood and governed by a proven playbook.For example, if a network node is generating alarms, the relevant criteria are clearly met, and the prescribed remedy is to remove that node from the network, an agent may be trusted to act without human intervention. That confidence does not yet extend to dynamic production changes such as altering router settings, routing logic, or subscriber database configurations. In those higher-risk situations, the agent can recommend an action, but a human operator should review and authorize the change before it is executed.
The path to agentic network operations depends as much on accountability, regulation, infrastructure, and trust as it does on technology. Operators and enterprises still need to clarify where responsibility sits when agents take action and how human teams remain accountable for outcomes. Regulatory requirements, such as emerging telecom AI rules, may also shape how quickly and broadly agents can be deployed. At the same time, organizations may need to rethink network operations center (NOC) infrastructure so that agents are supervised appropriately. Building trust will require giving agents defined areas where they can safely act, much as earlier automation gained acceptance only after teams saw it perform reliably within controlled boundaries.
On overcoming barriers: “More than a nything else, I feel it's (AI) not a technology transformation. It's more of people and cultural transformation and trying to break those silos that we work in.”
