Imagine a telecom network where autonomous agents resolve outages before customers notice, optimize costs in real time, and generate $60 billion in savings by 2030. This is the promise of agentic AI, but getting there is far from simple. While my last blog outlined the broad range of exciting, real-world implementations of agentic AI, there are still technology, governance, and trust barriers to surmount to deploy agentic AI at scale. To close the gap between potential and reality, senior executives must first grasp the complex challenges of moving from initial pilots to scalable, enterprise-wide deployment.
Fluid agents vs. rigid legacy
Perhaps the biggest obstacle to enterprise-wide roll-out of AI agents is the imperative to seamlessly integrate innovative agentic AI approaches with legacy systems and siloes of data created over decades of operations. A recent survey by consultancy Deloitte found that this organizational barrier ranked among the top challenges for nearly 60% of AI leaders it surveyed. Agentic AI systems thrive in connected, dynamic environments. Conversely, much of the legacy infrastructure that large telecom providers rely on is siloed and rigid, making it difficult for autonomous agents to plug in, adapt, and orchestrate processes across the network.
Today, telcos rightly deal in deterministic data as the foundation of decisions. Predictability, certainty, and accuracy are the hallmarks of good data. AI, however, by its nature is probability-based and its outputs are uncertain. Squaring the circle to bridge this gap, and simultaneously reduce the capacity for AI hallucinations, is vital to open the doors to wider adoption.
Reliable, high-quality data is the bedrock of safe agentic AI deployment. Without it, telcos risk not only ineffective AI decisions but also regulatory, governance, and operational pitfalls. Autonomous agents amplify both the benefits and the risks of the data they consume, making robust data management non-negotiable. Establishing a strong data foundation that ensures data is accurate, comprehensive, and readily available is an essential prerequisite for at-scale deployment of agentic AI in telcos.
Wild robots
As self-governing, continuously improving processes, AI agents act autonomously based on what they know and, critically, what they learn. As Duarte Begonha, Partner at McKinsey, colourfully noted at a recent Ericsson event, “AI Agents are like water, they go where they want to go.” The current generation of agents lacks the reliability and traceability necessary to allow them to be unleashed in network operations. For instance, if an AI agent misinterprets a network anomaly it could trigger a cascade of automated responses, potentially amplifying a minor issue into a major outage.
Overcoming the black-box issue — understanding why agentic AI made decisions and what data informed them — is a key step toward enabling widespread rollout.
Security is a related area of concern, acting as a hurdle to wider adoption. The multi-step nature of agentic AI means that one action, correct or incorrect, can trigger many others, leading to risk cascades across the whole business. So called “chain vulnerabilities” can propagate flaws and errors rapidly leading to critical situations. Corrupted data could also be replicated and amplified as subsequent agentic processes build on results derived from poor data. Data may leak unnoticed or through unrecognized mistakes, and there are numerous novel vectors for malicious attacks, for example, using prompt injections to make an agent act in a manner deleterious to the organization or its customers.
Human agency
The third area of concern hindering agentic AI adoption relates to the people working with and alongside the technology. Deploying agentic AI requires advanced technical capabilities in areas such as adaptive learning, agent orchestration, realistic simulation and enterprise integration. These are skills which few telcos have, certainly at the scale to support widespread implementation. Nvidia’s recent study into AI agent readiness in telco showed availability of AI experts as key obstacle to adoption; 43% of respondents cited it as a major issue in 2025, up from 34% in 2023.
Beyond the skills shortage there are also cultural issues that need to be addressed. The media is full of stories of how AI will take white collar jobs. But those working with AI agents will also need to adjust. The current mix of high expectations and low trust could limit their ability to reach full potential. As the TM Forum points out “[T]here will be a need to address operator culture to maximize potential use cases and ensure optimal collaboration between agents and humans.”
Then there are the governance and compliance aspect to consider. Telcos are uncertain as to the extent of their latitude to delegating decision-making power to AI systems, particularly because regulatory frameworks specific to agentic AI are still evolving. Addressing risk and compliance is a top concern for many AI leaders in the sector that is not only heavily regulated, but where even a single misstep, like a misconfigured router or a false positive in fraud detection, can disrupt services for millions of users, breach service level agreements and trigger regulatory penalties.
Build the foundations now
Despite these challenges, the long-term value of agentic AI is undeniable. Those who address these issues earliest will cement their leadership in the next wave of telecom innovation and growth. While the technology evolves, telcos can start to build the foundations that can deliver safe, trustable agentic AI that delivers rapid return on investment. Approaches that address today’s limitations are already being explored. Techniques including retrieval-augmented generation (RAG) plus technology architectures and governance guardrails can be deployed now to reduce risk, improve efficacy, and help streamline the use of agentic AI alongside existing processes. In the next blog, I’ll explore some of the specific building blocks that telcos can begin to implement now to pave the way for practical, scalable adoption of agentic AI.