Agentic AI “not even a teenager” – Orange exec

• Agentic AI is stirring passions in the telecom sector but it is still immature, far from being a differentiator and shouldn’t be applied indiscriminately 
• It is “very far from being an adult – it’s not even a teenager”, says Orange’s Philippe Ensarguet 
• Successful deployment by network operators will happen only if they fully embrace the AI-native telco concept, he noted
• Real-world deployments remain thin on the ground but Deutsche Telekom has emerged as a pioneer

The term “agentic AI” is being bandied about the global telecom sector so frequently and loudly you might think it has already transformed the industry and helped usher in the much-vaunted and long-expected network automation era. However, the reality is somewhat different, as Phillipe Ensarguet, VP of software engineering at Orange, explained during his presentation at the AI-Native Telco Forum in Düsseldorf, Germany, recently.

Once properly trained, agentic AI has the potential to introduce levels of autonomous decision-making and process implementation that could, in theory, help telcos deliver the promises of 5G and introduce levels of efficiency that would not only be beneficial but could be vital for digital service providers in the coming years – it has long been noted by network operators that manual processes just won’t cut it in a cloud-native, AI-native world.   

But there’s a long way to go before agentic AI has any meaningful impact on telco operations, it seems. 

Agentic AI is a technology that is “very far from being an adult – it’s not even a teenager” and can only be successfully deployed by network operators if they fully embrace the “AI-native telco” concept, Ensarguet told attendees at the AI-Native Telco Forum. 

During his presentation, he noted that Orange has conducted multiple agentic AI “explorations” in the past 18 months, including using a large language model (LLM) integrated with multiple model context protocol (MCP) services in a 5G lab to troubleshoot customer quality issues. 

Tests and trials are one thing, but “moving to production is a different story,” he added, noting that there are many challenges standing in the way of successful agentic AI deployments by telcos. He concluded his presentation by asking his telco peers a question: “Do you really, really need it?” (You can catch up with his presentation in this on-demand video.) 

That’s a question that telcos, blinded by the latest set of vendor-driven promises, might not necessarily be asking themselves, but clearly should. Agentic AI may well have enormous potential, but the technology is nowhere near being fully developed. Service providers will need to be very careful when considering the pros and cons of the technology’s capabilities and must not get carried away by irrational exuberance and unrealistic expectations engendered by overexposure to flatulent marketing hype. 

Agentic AI isn’t a universal nostrum to be applied indiscriminately. It should be used as a precise tool where there is an evident business case indicating that the technology’s unique capabilities can solve real-life problems better than existing alternative AI models and techniques, such as machine learning (ML), symbolic reasoning and ‘traditional’ automation systems.

At the same time, it’s clear that many telcos are at least prepared to see what gains can be made from agentic AI deployments over the next few years. 

A new report from research house Omdia, Agentic AI: An Evolution with Transformative Potential for Telecom Operations, finds that it will, eventually, be a key driver of autonomous network operations. The report reveals that 41% of respondent communications services providers (CSPs) regard network management as the area where agentic AI will have the most impact, initially in improving customer experience and later in enabling “autonomous diagnostics, optimisation and fault resolution at scale.”

Omdia also notes that big telecom and IT vendors, such as Amdocs, Ericsson, Huawei, Nokia and Salesforce, are integrating agentic capabilities into their platforms. Roz Roseboro, senior principal analyst at Omdia, commented: “While customer care is a visible starting point, the 41% of CSPs targeting network operations are signalling the real revolution. This moves agentic AI from an outward-facing support tool to the core engine of the business, enabling the autonomous, self-healing networks required to manage future service complexity.”

The report’s analysis further finds that CSPs are “cautiously optimistic” and are “prioritising low-risk use cases while emphasising the importance of observability, explainability and governance in early deployments.” 

Take up of agentic AI is also being boosted by “multi-agent collaboration” and the emergence of new protocols, such as Model Context Protocol (MCP) and Agent2Agent (A2A). 

MCP is an open-source standard that allows large language models (LLMs) to seamlessly connect with external tools, data sources and services, essentially enabling AI agents to make decisions based on multiple, defined data sources. MCP standardises the communication between AI models and other systems, making it easier to build AI-powered applications. 

The A2A protocol is another open standard that enables AI agents to communicate and collaborate securely and efficiently across different platforms and vendors. 

The introduction of such protocols is driving the development of scalable, interoperable ecosystems that connect vendors and platforms across the telecoms landscape. 

The Omdia report concludes that CSPs should begin deploying out-of-the-box agentic solutions whilst simultaneously building internal expertise to maintain control over data and development. It also advises that telecom vendors prioritise transparency, flexibility and portability in their agentic offerings to strengthen customer trust and long-term adoption.

Questionable potential  

Claims that agentic AI is already a “competitive differentiator” rather than an interesting concept that may have considerable validity at some point in the future should be taken with a pinch of salt and for those that do take the plunge, there may be trouble ahead. 

In its agentic AI report, Gartner warns that by the end of 2027 some 40% of planned agentic AI projects will be cancelled because of escalating costs, uncertainty about the technology’s business value and a lack, if not a complete absence, of controls sufficient to mitigate and minimise risk. Gartner, along with other critics trying to peer through the all the marketing fog, repeatedly makes the same point as Orange’s Ensarguet in that the agentic AI market is immature and has a lot of growing up to do before it can be deployed in mission-critical network situations.

As reported by the Harvard Business Review, large numbers of agentic AI projects fail because of “misalignment between the technology’s capabilities and the business problem at hand. Current agentic AI models lack the maturity and agency autonomously to meet complex business goals or follow nuanced instructions over time. Many deployments are little more than advanced chatbots or robot-process automation (RPA) tools with a conversational interface. This may improve user experience, but it does not deliver the transformative value that justifies the investment.” 

What’s more, integrating agentic AI into legacy systems and established workflows is very difficult. 

When it comes down to it, unfortunate realities, such as lack of compatibility between systems, the continuing problem of information being isolated in data silos and the need to rejig established process routines, can quickly combine to increase costs and extend project timetables. Attempting to retrofit agentic AI into a telco’s processes regardless of their suitability for the task can be an expensive and resource-heavy mis-step. Currently the market is confused, fragmented and verging on incoherence.

Organisations adopting agentic AI without paying proper attention to the expense of what is an extremely complex task, can be – will be – taken by surprise, become disappointed, lose their enthusiasm, cut their losses and cancel their agentic AI plans. After all, very well tried and tested automation techniques, machine learning and “traditional software” may themselves deliver equal or even greater value than agentic AI, and at a mere fraction of the cost, and prevent massive organisational disruption by stopping a telco from blundering down a technological cul-de-sac.

Deutsche Telekom – an agentic AI pioneer in telecom 

That said, some telcos have made their first pioneering forays into the commercial deployment of agentic AI. 

For example, Deutsche Telekom (DT) says its RAN Guardian Agent, developed in collaboration with Google Cloud, is now in use in its production network. 

Ahmed Hafez, DT’s senior VP of network strategy and data & AI in networks, told the AI-Native Telco Forum audience that RAN Guardian is part of the operator’s new wave of generative and agentic AI deployments – and “is a multi-agent system that self-heals and optimises radio networks in real time, dynamically adjusting configurations ahead of congestion or local events.” 

He also emphasised that the German telco carefully limits the “decision scope” of each agent to a set of “pre-defined, non-destructive actions” and “maintains human oversight with full logging and feedback loops. 

Stressing the importance of transparency and continuous verification, Hafez added, “We protect from any damage by this kind of limitation”. The limitations are, by definition, structures of constraint but at least DT evidently has carefully considered how it will exploit agentic AI and the order in which it will deploy it.  

In summation, deep within its currently immature existence, agentic AI should eventually grow up to be transformational for telcos, but only if they have a formalised and ratified plan in place together with the strategic intent and the organisational discipline needed to follow change through from beginning to end.

When, in its 2025 report on agentic AI, IBM’s ‘Responsible Technology Board’ wrote that “agentic AI introduces new risks and challenges”, it was right on the nose. Agentic AI has now reached the second phase of the five-phase Gartner Hype Cycle, which is the ‘Peak of Inflated Expectations’. When that dizzy height is reached it takes no more than to trip over a marketing brochure to slip headlong into Phase Three – ‘The Trough of Disillusionment’. 

Telcos shouldn’t wait for that trough, though, to ask themselves that key question posed by Ensarguet: “Do you really, really need it?” 

– Martyn Warwick, Editor in Chief, TelecomTV