What Agentic AI Really Means for Telcos
What is Agentic AI?
We hear about Agentic AI everywhere these days. Major tech companies are racing to integrate it into their operations, and large consultancies are releasing shiny reports claiming it’s the next big thing!
Over the years, many transformative technologies have reshaped industries. In retail, for example, online shopping was once controversial. Many brands debated whether it was worth the risk. Today, having an e-commerce channel is non-negotiable. The same happened in media: streaming was doubted as uneconomic and niche, and now it’s the standard. Entire supply chains, from content delivery to network capacity planning, revolve around it.
A similar optimism surrounds Agentic AI. According to Cisco’s report on how Agentic AI will shape customer experience, most surveyed executives believe that vendors adopting AI agents will gain major efficiency and personalisation benefits, while those who delay risk customer churn.
But what exactly is Agentic AI? What should we expect AI agents to actually do in the context of network operations and, more importantly, what should we not expect them to do?
The Buzzword Evolution From “AI” to “Agentic AI”
At this year’s Mobile World Congress and Connected Britain, “Agentic AI” or “AI Agents” appeared on banners and slides everywhere. Just a couple of years ago, “Generative AI“ or “Gen AI” dominated the stage. Before that, it was simply “Artificial Intelligence“ or “AI.”
This rapid rebranding reflects two paradigms:
Genuine exploration: as the industry better understands AI capabilities that are seamlessly evolving at a fast pace.
Marketing refresh: making old ideas sound new.
Understanding what an AI agent actually does is essential to evaluating how it might improve operations or deliver a return on investment. It’s easy to see why companies have embraced the term Agentic AI: it suggests a more advanced, purposeful form of intelligence — not just any AI, but a smarter, more autonomous one. The shift in vocabulary gives the impression of technical maturity. Yet, in practice, Agentic AI remains largely an evolution in branding rather than a fundamental change in capability.
As a matter of fact, there’s no universal definition of Agentic AI. Broadly, it describes multiple AI agents, most often powered by large language models (LLMs), that collaborate via natural language to perform complex tasks. It contrasts previous terminology such as (deterministic) AI — a more general term referring to systems capable of performing tasks that normally require human intelligence like pattern recognition, prediction, or decision-making, and GenAI — a subset of AI focused on creating new content like text, images, code, or designs using patterns learned from large datasets and possibly conditioned on user-defined input or exogenous context variables.
Agentic AI builds on these same foundations but adds a layer of autonomy and coordination: instead of a single model producing output, multiple models can plan, communicate, and act together toward defined goals. Conceptually, this represents a step forward. However, since individual AI models are still not fully trusted or deployed at scale in telecom operations, letting multiple agents “cooperate” only amplifies the existing challenges of reliability, interpretability and control in critical infrastructures such as large-scale, operational telco systems.
Telecom Reality — ML Adoption Is Taking More Time Than Expected
Despite the buzz, large-scale AI deployments remain rare in the telecom ecosystem.
Agentic AI may sound like the next big thing, but in practice, its goals overlap with the same automation challenges that telcos have been wrestling with for years. Machine Learning research evolves quickly because it lives in an algorithmic domain, where new code, models, and data can make breakthroughs almost overnight. Telecom players, however, operate in a systems domain: they are highly regulated, risk-averse, and slow to evolve. Integrating any new paradigm into live networks is very expensive and time-consuming. That gap between AI’s pace of innovation and telco’s pace of adoption explains why, at major industry events like Connected Britain, “Agentic AI” dominates the headlines even though real-world deployment remains limited. It’s less about what’s already working and more about showing that the industry is keeping up.
Again, there is no clear definition for it, but in telecom, Agentic AI could mean diverse agents monitoring network KPIs, diagnosing faults, or triggering reconfigurations independently to then coordinate autonomously. This is an evolution, rather than a revolution. More importantly, it is not an answer to the big problem: while AI is used in production for a few, specific, long-timescale, and non-close-loop tasks such as customer support, analytics, or radio planning, the large-scale deployment of AI in mission-critical network control is not dependable enough to be trusted by the major 5G network operators.
The actual barriers that need to be addressed are:
Economic friction: unclear or delayed ROI — According to Forbes, 62% of executives expect at least three years to see ROI from AI investments, and one-third report being disappointed with early results.
Talent gap: limited in-house AI expertise and little intersection between the machine learning community and the telecom industry.
Strategic fatigue: overpromising and underdelivering pilots.
This slow pace reflects telecom’s structural complexity. Legacy infrastructure makes experimentation costly and cautious. And Agentic AI is not a solution that directly targets these issues — in fact, it risks having the opposite effect by complicating AI operations and further entangling their integration in network systems.
Promise vs Pitfall
Letting LLMs coordinate to achieve complex network management tasks also introduces two new major challenges that add to the existing burden of AI operation in telco systems:
Reasoning vs. mimicry: LLMs do not truly “reason.” When asked to explain their logic, they often reveal a lack of genuine understanding. They succeed by recognising statistical patterns, not by building causal models. Academic research takes the idea of “agency” far more seriously than most industry messaging suggests. A 2022 DeepMind and University of Oxford paper defines an agent not as something that produces output, but as a system that would behave differently if it understood the world differently. By that definition, most LLM-based “agents” today fall short: they generate responses, but do not adapt goals or models of the world itself.
Sustainability and scale: LLMs require substantial GPU resources, energy, and cooling infrastructure — possibly orders of magnitude more than “baseline“ neural network models. Running hundreds of models to control a nationwide network could drastically increase energy costs, at odds with telecoms’ paramount goal of reducing OPEX, energy bills and carbon footprints to increase the current, lean profit margins and align with green regulations.
Deloitte’s Agentic AI Blueprint for Telcos provides a useful example of how the industry is approaching Agentic AI in practice. Their agent, “Hoopie,” is presented as capable of responding to network incidents and supporting human analysts through coordinated agents. In the demonstration, Hoopie primarily assesses situations and provides recommendations, while decisions remain with human operators, illustrating that autonomy is currently framed as assistive rather than fully autonomous.
The publicly available material focuses on architectural concepts rather than technical implementation, and does not describe in detail how decision logic is generated or validated. At the same time, Deloitte highlights the significant infrastructure requirements of GenAI, noting that telcos will require “chips, energy, water and money” to compete, which brings attention to the scale and sustainability considerations involved in deploying such systems.
Similarly, one of the recent TM Forum catalyst projects focused on ‘Agentic & autonomous AI for business excellence’ illustrates how the industry currently envisions autonomy: multiple agents coordinating tasks within a shared framework. In the demo, a “master agent” orchestrates specialised agents such as the “business agent” and the “resource agent,” while an “assurance agent” oversees execution. Although this architecture looks futuristic, the publicly available information offers no insight into how decisions are actually made or validated, leaving the underlying reasoning elusive.
Looking Ahead — What’s Actually Worth Watching
So, where does AI genuinely create value in telecom today?
One of the sector’s most pressing issues is shrinking profit margins. The recent Vodafone-Three merger illustrates how operators are striving to expand their customer base while keeping infrastructure costs flat. AI agents may help streamline front-office operations like customer service, but the real efficiency gains will come from machine learning systems that improve network management.
Tools that predict traffic patterns and allocate network resources proactively, rather than reacting after problems occur, are more effective at reducing over-provisioning, improving energy efficiency, and delivering measurable ROI.
Telcos’ slower pace of AI adoption isn’t due to a lack of ambition but to the complexity of their systems. Integrating and maintaining any AI in live networks demands specialised expertise that many operators are still building. The real opportunity lies in closing that knowledge gap by focusing on technologies that deliver incremental, data-driven improvements, rather than chasing the next buzzword.
Sources:
https://www.ibm.com/downloads/documents/us-en/1443d5f5af4f532a
