Friday, 18 September 2026
Artificial intelligence in telecommunications is moving beyond customer-service chatbots and experimental projects towards a much larger ambition: networks capable of identifying problems, predicting failures and eventually taking corrective action themselves. At Ai4 2026 in Las Vegas, a panel examining how AI is transforming telecom brought together perspectives spanning network infrastructure, machine learning, customer experience and autonomous-network research. Moderated by Cathy White, founder and CEO of CEW Communications, the discussion included Sahil Yadav, Head of Software and AI at Applied Optoelectronics; Sundar Krishnan, Director of AI and Machine Learning at Optimum; and telecom veteran and Bell Labs Fellow Anne Lee.
The discussion suggested that telecommunications may become one of the industries where the transition from generative AI to agentic AI has particularly visible consequences. Rather than simply helping employees generate information, AI systems are beginning to interact with the physical and digital infrastructure responsible for keeping millions of customers connected. For telecom operators, some of the clearest opportunities are already emerging in network operations. Predictive maintenance, anomaly detection and automated diagnostics allow operators to identify potential failures before customers necessarily experience them. The next step is self-healing, where AI not only identifies the problem but recommends or ultimately executes the solution.
Yadav described this transition from the perspective of network infrastructure. Applied Optoelectronics supplies communications equipment and associated software, and its AI work includes systems intended to support network optimisation and increasingly autonomous operations. The potential economic impact is considerable. Network outages create direct repair costs while also affecting customer satisfaction, retention and an operator’s reputation. AI capable of identifying deteriorating equipment before failure could allow maintenance teams to intervene earlier, while better diagnostics could reduce unnecessary field visits.
However, fully autonomous telecommunications networks remain some distance away. The panel drew an important distinction between specialised automation and an entire network capable of continuously managing and evolving itself. Lee, whose career has covered successive generations of telecommunications technology, argued that the nearer-term opportunity is likely to involve AI agents designed to handle specific network problems. An agent could recognise a recurring failure pattern, determine an appropriate response and potentially resolve it automatically within clearly defined limits.
Over time, larger numbers of specialised agents could work together across different parts of the network. This is consistent with research into self-operating networks, where agentic AI, digital twins and other technologies could eventually support infrastructure capable of adapting dynamically with substantially less manual intervention. The progression will probably be gradual because the consequences of an incorrect decision can be significant. Yadav described an approach based on risk. Low-risk and easily reversible actions can increasingly be assigned to AI. Medium-risk actions can operate under human supervision, while decisions capable of causing significant network disruption should continue to require human approval.
This could become one of the most important principles governing AI deployment in critical infrastructure. The objective is not necessarily to make every network function autonomous as quickly as possible. It is to expand autonomy as confidence, testing and operational experience improve. AI is also changing the economics of field operations. Telecom providers routinely dispatch technicians to investigate customer and network problems, but sending a vehicle and engineer to every suspected fault is expensive. Better prediction and remote diagnosis can help operators determine which problems genuinely require physical intervention.
Customer applications can increasingly perform initial diagnostics themselves, while AI can guide subscribers through basic troubleshooting. Visual AI could eventually allow customers to photograph equipment and receive instructions based on what the system sees. More sophisticated network models can analyse signal quality, connected devices and historical faults to determine whether a problem originates inside a home, elsewhere in the local network or within wider infrastructure. This creates a direct connection between AI investment and operating expenditure. The value is not the number of AI queries generated but whether operators reduce outages, unnecessary service calls, customer churn and the amount of engineering time spent on routine problems.
The customer side of telecommunications is undergoing a parallel transformation. Krishnan described how recommendation systems have progressed from relatively simple rules towards machine-learning and AI systems capable of providing customer-service representatives with more contextual guidance. Instead of waiting for a subscriber to explain the entire history of a problem, an AI system can potentially analyse the customer’s relationship with the operator, recent service issues, products and previous interactions before the conversation begins.
That could allow the representative to understand whether someone is calling because they want to upgrade a service, complain about unreliable broadband or potentially cancel their subscription. The system can then suggest relevant information, offers or potential resolutions. AI could also make customer retention more proactive. Rather than waiting until a dissatisfied subscriber contacts the operator to cancel, predictive systems can identify behavioural patterns associated with potential churn and allow the company to intervene earlier.
Digital twins could extend this further. Krishnan described experiments involving simulated customers that allow companies to test how different types of subscribers might respond to offers or service decisions before those approaches are deployed more widely. The underlying objective is hyper-personalisation, but the commercial test remains straightforward: does the technology improve retention, service quality or revenue sufficiently to justify its cost?
That question is becoming more important as corporate attitudes towards AI expenditure mature. The first phase of the generative-AI boom encouraged companies to experiment widely, often measuring adoption through the number of users, prompts or tokens consumed. The next phase is increasingly about linking expenditure to business results. For telecom companies, that means distinguishing between an impressive demonstration and a system capable of operating reliably across millions of subscribers or network devices.
Moving from a proof of concept to production requires considerably more than a functioning AI model. Companies need data pipelines, access controls, monitoring, testing, governance, latency management, security and people capable of maintaining the resulting system. This helps explain why promising AI pilots do not automatically become commercial deployments. Building a demonstration may prove that a technology works. Scaling it requires proving that it remains reliable when exposed to the complexity of a real telecommunications network.
The industry has faced similar challenges before. Each major generation of telecommunications technology has moved through prototypes and limited deployments before reaching commercial scale. AI is different technologically, but many of the engineering disciplines required for deployment remain familiar: reliability, capacity, security and operational resilience.
Legacy infrastructure presents another major obstacle. AI depends heavily on data, but large telecommunications networks contain equipment installed across many different technological generations. Some devices can provide detailed real-time telemetry and be controlled remotely, while older infrastructure may provide relatively little information. Making networks more intelligent therefore has a physical investment dimension.
Operators cannot simply install an AI platform and expect decades-old network equipment to become autonomous. In some cases, equipment needs additional sensing and telemetry capabilities. In others, older devices may ultimately need replacement. That makes AI deployment partly a capital-expenditure question. Operators need to determine where upgrading infrastructure creates enough operational value to justify the investment.
The panel suggested that phased deployment offers one practical solution. An operator can modernise equipment within a particular geographic area, introduce AI capabilities there and observe how the technology performs before extending it across the wider network. This approach also limits the potential impact of failure. If an autonomous system behaves incorrectly, the affected portion of the network remains relatively contained while engineers gain real-world experience that can improve subsequent deployments.
Edge AI could accelerate this transformation further. Smaller and more efficient models increasingly allow AI processing to take place closer to the network device or customer rather than sending every request to a large central cloud environment. For telecom companies, edge processing can offer advantages including lower latency, greater resilience and potentially better control of sensitive information. Specialised AI agents could perform relatively narrow functions on network equipment without requiring the computing resources associated with the largest generative models.
The relationship between AI and telecommunications consequently works in both directions. Telecom companies can use AI to operate their infrastructure more efficiently, but the growth of AI itself is also increasing demand for faster, more responsive and more distributed communications infrastructure. Industry research in 2026 increasingly describes this as a shift from using AI within telecom towards telecom becoming part of the infrastructure enabling the wider AI economy.
The human consequences may be equally significant. Network engineers are unlikely simply to disappear as automation increases. Their responsibilities are more likely to change. An engineer who previously spent much of the working day identifying and correcting routine network faults may increasingly supervise automated systems, investigate exceptional situations and manage the AI agents operating parts of the infrastructure. Understanding machine-learning systems, evaluating AI recommendations and determining when automation should be overridden could consequently become important telecom engineering skills.
Lee argued that releasing engineers from repetitive network-management work could also create capacity for new services. Telecom companies have spent enormous resources maintaining increasingly complicated networks. Greater automation could allow some of those resources to move towards developing products that generate additional revenue or provide wider social value. Emergency communications provide one example. Modern networks are technically capable of transmitting far more than voice, including video, images and other information that could potentially assist emergency responders. Developing and operating such services requires investment, however, and efficiency gains elsewhere in the network could make these types of applications more economically practical.
Customer service could experience an equally visible transformation. Traditional interactive voice-response systems remain a frequent source of frustration because users navigate layers of menus before reaching the information or person they need. More capable conversational AI could eventually replace much of that structure. Instead of selecting options through a telephone keypad, customers could describe the problem naturally and allow the system to determine what information or action is required.
The important threshold will be whether customers stop wanting to escape from the automated system and speak immediately to a human. Reaching that point will require better contextual understanding, reliability and trust rather than simply more convincing synthetic voices.
Telecommunications therefore provides a useful illustration of where enterprise AI is heading. The industry is moving beyond the question of whether companies should experiment with AI. The more important questions now concern which functions should be automated, how quickly autonomy should expand and whether the economic return justifies the infrastructure required to support it.
The ultimate destination may be networks capable of continuously monitoring their own performance, anticipating failures, reallocating resources and resolving many problems without waiting for engineers to intervene. Getting there will require modernised physical infrastructure, reliable data, governance, human oversight and a gradual transfer of responsibility from engineers to machines.
The result would represent more than another efficiency programme. Telecommunications networks could evolve from infrastructure that humans continuously operate into infrastructure that increasingly operates itself, with engineers supervising the intelligence responsible for keeping it running.
