- AI agents are moving from telecom network recommendations to real-time actions, raising new questions about operator control
- Nokia, Ericsson and Samsung say telcos must remain in charge of network AI, setting policies, guardrails, permissions and audit trails
- Co-developing with hyperscalers while avoiding lock-in is another critical balancing act
After a few years sailing the roiling AI waves, telecom operators are finally getting their sea legs. But as they experiment with powerful AI agents that recommend and act on network changes, a key question is cropping up: who is actually in control – and responsible – when autonomous agents tinker with live infrastructure?
The answer from major network vendors is strikingly consistent. Operators, not vendors or hyperscalers, must govern the policies, authorization boundaries and risk thresholds that determine what an AI agent can do. But the mechanics of that control are becoming more complicated.
“The operator's policy must govern the outcome,” Nokia’s Oguz Sunay, CTO of AI and Autonomous Networks, told Fierce. “AI can help identify the best action; the operator defines the conditions under which it can occur…The goal is not to remove people from the loop, but to place human judgment where it creates the most value; at the governance and approval layer, rather than at every routine decision. “
Sunay described the target model as “glass-box governance,” where agents are authorized for specific actions, constrained by policy and overseen by a governance layer that prevents local automation logic from overriding broader service intent.
Ericsson Americas’ Head of Technology Strategy Office Akhil Gokul made a similar point, framing agentic AI as a capability inside an autonomous-network operating model, not an independent authority. Vendors and hyperscalers may provide infrastructure, software, models and tooling, he said, but those capabilities should operate within the operator’s boundaries.
“Where a model or workload runs is not the same as who governs the network,” Gokul said. “Operators can retain control by defining the intent and policies, governing data access and residency, controlling agent permissions and orchestration, retaining authority over network actuation, and maintaining visibility and auditability.”
Keeping operators in the driver’s seat
That distinction matters because center of competition between hyperscalers, operators and vendors is shifting. The emerging question is less about where software runs and more about who owns the intelligence that interprets network state, decides which actions are safe and determines when an agent is allowed to touch the network.
Arjun Nanjundappa, staff engineer for system structure design in Samsung’s networks business, told Fierce that operators do not need to own every component of the AI stack to retain control. But he argued they should own the reasoning, action and observation loop; tool integration; context and memory management; guardrails; and governance functions.
“Ultimately, the key question is not who owns every underlying technology component, but who has the authority to determine how the network is operated,” Nanjundappa said.
Practically speaking, that authority comes down to guardrails. Nanjundappa pointed to bounded action, authorized restrictions and guardrail validation in a digital twin environment as mechanisms for keeping agentic automation from becoming open-ended execution authority.
In his example, an agent could be limited to modifying a specific network slice or restarting a cloud-native network function, keeping a bad remediation step contained rather than allowing it to ripple across the network.
The question of liability when something goes wrong, however, is a tougher nut to crack.
Asked who is liable when AI makes a change that results in a service outage, both Nokia and Ericsson said responsibility will depend on deployment design, commercial agreements, the authority granted to the agent, applicable law and the facts of the incident. Sunay and Gokul also stressed the importance of auditability and audit trails for AI decision-making to help operators understand what actions were taken, why and under what authority.
Where do hyperscalers fit in agentic AI strategy?
In this complicated equation, hyperscalers are both essential partners and strategic tension points.
All three vendors described hyperscalers as AI co-development partners who offer infrastructure, scale, toolsets and model and developer ecosystems. But they still all flagged lock-in as a risk.
Sunay warned of an “intelligence gravity well,” where lock-in shifts from hardware to the models, policies and context graphs that drive network decisions. In that scenario, the provider that controls the agent context layer could exert more influence over network behavior than the provider hosting the core workload.
Both Ericsson and Nokia stressed the need for openness and interoperability to combat this risk.
“Standards-based, governed interfaces — across bodies like O-RAN, 3GPP, TM Forum and the AI-RAN Alliance — are what keep agentic automation portable across ecosystems, turning co-development into a durable model rather than a one-way dependency,” Sunay concluded.
Read more about AI developments on Fierce Network
GSMA warns telcos against outsourcing AI future to hyperscalers
What is sovereign AI? Only 13% of enterprises know.
SK Telecom's AI model push offers a blueprint for telcos seeking sovereignty
Dell builds a sovereign AI stack for the on-prem era