From Trust to Autonomy: Why AI Must Earn the Right to Act
There is a persistent misconception about autonomous networks – that achieving greater autonomy means removing people from network operations. That’s the wrong objective. The goal should be to determine which decisions can safely be delegated to machines, under what conditions, and with what evidence.
Whether man or machine - an operator would be foolish to hand responsibility for a national network to a new engineer (virtual or otherwise) on their first day. Responsibility is earned through observation, validation and repeated success. AI should be expected to meet the same standard.
Autonomy is a progression, not a switch
Consider how this might work in everyday network operations.
Initially, an engineer may use an AI system to investigate a problem and recommend an action while retaining full approval authority. The system identifies the likely cause, presents the supporting evidence and proposes a remediation. The engineer reviews the recommendation and decides whether to proceed.
Now imagine that process repeated hundreds or thousands of times. If the recommendations prove consistently accurate and the outcomes can be verified, the operator may become comfortable allowing the system to execute certain low-risk changes automatically.
A Wi-Fi channel adjustment might be one example. A broadband profile optimization could be another. A nationwide routing change is something entirely different. This is why autonomy can’t be implemented as a single policy across the network. Different actions carry different levels of operational and business risk. The level of human oversight should reflect that reality.
Give AI a defined operational mandate
Operators need more than intelligent agents, they need governance around those agents. That means defining what an AI system is allowed to observe, what decisions it can recommend, what actions it can execute and when human approval is required. The boundaries should change as confidence grows.
Low-risk, high-frequency decisions are natural candidates for early automation. Higher-impact actions remain supervised until the system has accumulated enough evidence to demonstrate consistent reliability. This creates a progressive model of autonomy.
The progression is important because it turns autonomy into something measurable. Instead of asking, "Do we trust AI to run the network?" operators can ask a much more useful question: “Which operational decisions have the system demonstrated that it can safely manage autonomously?” That’s a question engineering teams can answer.
This is also where the architecture behind autonomous operations becomes important.
At AXON Networks, turning this principle into an operational architecture is enabled by AXON Maestro, where autonomous operations are built around a real-time Digital Twin, operational intelligence and agentic AI rather than treating AI as a separate layer sitting above existing OSS systems. The AXON Digital Twin provides the operational context by maintaining a current representation of the network across infrastructure, services, telemetry and customer experience. AXON Neura, a sub-system of Maestro, uses that context to support operational reasoning and actions for different network and service personas.
AXON Maestro also provides the environment in which those decisions can be orchestrated, governed, executed and verified. The distinction is important. An AI agent shouldn’t be given authority simply because it can generate a plausible recommendation. It should earn authority because its decisions are grounded in trusted network data, executed within defined policies and verified against actual network outcomes.
That is the foundation required to move from AI-assisted operations toward closed-loop autonomy.
Confidence becomes the control mechanism
All this changes how we should think about the path toward AN-4 autonomous operations as defined by TM Forum. The objective is appropriate autonomy, not maximum automation.
Some operational decisions may become fully autonomous very quickly. Others may remain human-supervised for years. A small number may always require human authorization because the consequences of failure are too significant.
That isn’t a limitation of autonomous networking – it’s good engineering.
The operators that succeed will therefore build autonomy progressively. They’ll start with bounded use cases, measure outcomes, verify actions, expand authority as evidence accumulates, and retain human judgment where the risk justifies it.
The future of network operations won’t belong to organizations with the most AI, it will belong to those that know when to trust it, how far to trust it and how to prove that trust is justified. That is how autonomous operations move from demonstration to everyday network management.
Next in the series: Measuring What Matters: Defining Success in Autonomous Network Operations
Many operators measure automation by the number of workflows executed or tasks eliminated. Those metrics say very little about business outcomes. The next article looks at how autonomous operations should be measured—from customer experience and operational resilience to engineering productivity and business value—and why choosing the right metrics determines long-term success.