The appetite for AI in telecoms is hardly surprising. Mobile networks are already complex to manage, and the economic case for automation is compelling. The top AI use case cited for return on investment among telecoms operators is autonomous networks, ahead of improved customer service and internal process optimisation, with around 90% of operators reporting positive impacts on revenue and costs.
The market is growing to match that enthusiasm. The global AI in telecoms market is anticipated to expand from around $2.9 billion in 2025 to $6.8 billion by 2031 (GSMA Intelligence, Q1 2026). Early adoption has understandably focused on customer-facing systems such as chatbots, churn prediction and personalised offers, where mistakes are unfortunate but recoverable. But operators are now moving towards integrating AI into the network itself. A recent GSMA Intelligence report highlighted that while current AI adoption on the network side is around 18%, it grew by 30% in the last six months of 2025 alone.
The prize operators are increasingly chasing is the concept of the Agentic NOC: an evolution of the traditional Network Operations Centre built around AI-driven autonomy, orchestration and closed-loop decision-making. The commercial incentive is difficult to ignore. NVIDIA estimates that telecom operators spent almost $295 billion in capital expenditure and more than $1 trillion in operational expenditure in 2024 alone, much of it on the planning, optimisation and maintenance of increasingly complex networks. Modern mobile systems contain vast numbers of interdependent parameters that require continuous tuning as traffic patterns, mobility and interference conditions fluctuate throughout the day.
The next logical step would be to remove humans from the loop altogether, an attractive prospect for operators looking to reduce operating expenditure, redeploy skilled engineers towards higher-value work, and eliminate the variability inherent in any human-dependent process. TM Forum estimates that a fully automated NOC framework could save between 300,000 and 500,000 person-hours per year in a network of 60 million subscribers. The case for handing that burden to machines is clear.
TM Forum and Red Hat have coined the term 'DarkNOC' to describe the ultimate vision for autonomous network operations, a highly automated operational environment in which AI agents perform the majority of monitoring, analysis and remediation activities with minimal human involvement. It is a compelling vision. But it raises an uncomfortable question: if something does break, will anyone still understand the network well enough to fix it?
The industry appears to be in something of a gold rush, while remaining firmly in the early stages of agentic AI maturity. Examples of the technology's limitations are not difficult to find. In May 2025, Waymo recalled more than 1,200 vehicles after it emerged that a software fault had been causing collisions with roadside barriers, a problem that had been occurring quietly for almost two years before triggering a formal recall. IBM's 2025 Cost of a Data Breach report found that 13% of organisations had already experienced a breach involving an AI model or application, while 97% of those affected lacked proper AI access controls. For an industry contemplating handing its critical infrastructure to autonomous systems, this should give pause for thought.
Traditional rules-based automation uses rigid, if-this-then-that logic that fails obviously when something falls outside its parameters. AI agents respond to changing conditions in a context-aware way, which is precisely what makes them useful. But where rules-based systems fail loudly, AI can do so quietly.
Models need to be thoroughly trained across a wide range of scenarios before they go anywhere near a live network, with proper guardrails in place to prevent overreach. In a multi-agent environment, the stakes are higher still. Errors don't occur in isolation, they compound. One bad decision shapes the next, and by the time the problem surfaces, tracing it back to its source can be extremely difficult.
At the same time, the industry must be careful not to mistake the availability of agentic AI for operational readiness. While relatively simple autonomous agents are already being deployed within tightly controlled environments and well-defined guardrails, more advanced systems capable of making high-stakes operational decisions remain in their infancy. Questions around trust, governance, observability and human oversight still need to mature alongside the technology itself.
The real challenge facing the autonomous telco network may not be whether AI can operate it, but whether operators can continue to understand the systems operating on their behalf. In an industry where network failures have direct and immediate consequences for millions of people, that is not a question the sector can afford to leave unanswered.
This article was originally published in ITP magazine and has been adapted for Mpirical by the original author, Kevin Moore, Mpirical’s Technical Trainer.
GSMA Intelligence, Telco AI: State of the Market, Q1 2026
NVIDIA, Network Operations Assist
TM Forum, Agentic NOC: AI-Native Operations for the Autonomous Telco, Catalyst Project C26.0.924
TM Forum, DarkNOC: GenAI Propels Insights Driven NetOps, Catalyst Project C24.0.693
Hernandez, F.P. (2024), DarkNOC: GenAI & Automation Propels Insights Driven NetOps, Red Hat Blog
IBM, Cost of a Data Breach Report 2025