Telecom has lived through more than one transformation cycle that promised to reinvent the business and, instead, left it more complicated. The Mobile Network’s new AI-Native Telco Market Update makes a case that AI risks becoming the latest example — unless operators change how they think about the problem. (Spoiler: It’s not just which tools they buy.)
The report is broadly framed around a comment from TM Forum (News - Alert) CTO George Glass — the difference between an AI-enabled operator and an AI-native one is the difference between building a faster horse and designing a tractor. What it means is if you bolt AI onto an existing workflow, the result may be a faster horse with real, measurable improvement, but one still constrained by the shape of the original process. On the other hand, redesigning the workflow around what predictive, generative and agentic AI can do creates an entirely different outcome. Judging from the report, much of the industry is still deciding whether they want a horse or a tractor.
AI-Native Is More Than AI-Enabled
Interestingly, though not surprisingly, the industry lacks consensus on what AI-native actually means. Vodafone Three’s Director of Network Strategy Valeria Baiamonte describes two definitions. The older, incremental version layers AI onto an existing network function, while the newer and more radical version speaks to self-optimizing, self-managing and self-healing networks with level 4 automation.
The latter, Baiamonte notes, “is more difficult to achieve and really requires strong foundations.”
Deutsche Telekom’s Outmane Laaroussi makes a related point that people routinely use the term AI-native while describing fundamentally different perspectives.
But, Laaroussi’s own definition seems to fall more along the lines of the newer model: “An AI-native network is built around intent and autonomy. Overall it’s about redesigning the network with the assumptions that autonomous agents are part of how it is built, planned, and operated.”
Here’s why the difference in definition matters. Vendor pitches, board-level strategies and engineering roadmaps can all carry the AI-native label while referring to very different operating models.
But, based on the idea that an AI-native network is built on foundations that allow AI to discover, reason, and act — that AI agents are part of the network planning, not afterthoughts — operators’ working definition seems to be setting a higher bar than most AI-enabled roadmaps currently clear.
That does not mean AI should be applied randomly or indiscriminately. Glass argues that deterministic automation, orchestration, predictive analytics, machine learning and agentic AI should remain complementary tools. A workflow with only a handful of predictable outcomes may not need an agent at all. The goal is intelligent design, not AI saturation.
That said, agentic AI points to a relatively clear dividing line between incremental improvement and a genuinely different operating model. Orange’s Philippe Ensarguet characterizes generative AI as essentially passive: It can explain, summarize and reason, but it cannot act. On the other hand, agentic AI introduces autonomy, and autonomy changes the nature of the system.
Practically speaking, it’s a shift away from the rules and scripts that have been created for anticipated scenarios, and a move toward systems that pursue goals, evaluate changing context and adapt to situations that haven’t been explicitly programmed for. That may seem self-evident, but it’s a very relevant different in the telecom context, where networks are becoming too dynamic and complex for humans or deterministic workflows to manage.
It’s also important to note this isn’t merely a sandbox conversation anymore. Deutsche Telekom’s RAN Guardian Agent, built using Google’s (News - Alert) Gemini models, has been operating in Germany since November 2025 and reportedly reduced event-management time from hours to roughly a minute. During Germany’s Carnival season, it monitored hundreds of mobile sites supporting more than 100 major public events, and Deutsche Telekom has since expanded the approach through MINDR, a multi-agent system spanning RAN, transport and core networks.
Verizon (News - Alert), Orange, EOLO and operators in China are pursuing comparable deployments, while IDC suggests roughly half of telcos surveyed claim to have more than 10 agents in production. Most of those agents, however, are concentrated in BSS and customer-facing functions rather than inside the network itself.
There’s some logic to that, if the thinking is to deploy AI agents initially where data and workflows are more tractable and the consequences of an error are easier to contain. Live networks inherently present the harder test.
Production Exposes the Real Obstacles
Some of the most interesting insight from the report comes when looking at what happens when a pilot moves into production. Laaroussi notes candidly that performance gains are relatively easy to demonstrate on curated data in a controlled environment. Production is where inconsistent APIs, unstructured information, legacy systems and years of multi-vendor accumulation compound to create challenges.
“The reality is that an agent is only as strong as the environment it operates within,” Laaroussi says.
That’s consistent with what most have said over the past few years about AI.
Scaling compounds the problem. A gap that one market can compensate for manually becomes unmanageable when agents are deployed across multiple markets, vendors and regulatory regimes.
There is also a more fundamental shift happening. Telecom networks were engineered around predictability, while agents introduce probabilistic decision-making. That shouldn’t be considered a defect — it’s a different operating model that forces operators to decide where agents may act autonomously, where humans must remain involved, how conflicting agent objectives will be reconciled and how feedback loops between adjacent systems will be prevented.
Again, that’s consistent with what other industries have been saying about AI agents. There are things AI can do entirely autonomously and others where human intervention is required. The key is to define each and design the AI appropriately with the right guardrails.
Orange’s VP of Software Engineering Philippe Ensuarget points to security, observability, testing, debugging, runtime isolation, and trust as major challenges. (Yet again, for anyone who has followed the AI space, this is a common thread.) He also raises the problem of coordinating intent when agents operating in different domains have overlapping skills or conflicting instructions.
His conclusion makes the case for open standards, common architectures and interoperable agent frameworks: “No company alone can manage all of these problems; it is only an ecosystem that can address all these, and here open source is important, and I truly believe it is only thanks to open source that we can succeed.”
The underlying AI models are increasingly capable, and agents are producing credible results in pilots and selected production environments. What remains unresolved is how to operate them safely, consistently and economically across the full complexity of a telecom business.
Given that, it’s not surprising that the report seems to repeatedly fall back to the idea of foundations more than the agents themselves. Operators already possess enormous quantities of network, service, customer and business data. Their problem is understanding the relationships within that data quickly enough to make safe decisions and take effective action.
Netcracker’s Susan White believes operators will need a telecom-specific knowledge layer that gives agents semantic understanding of the relationships among assets, services, customers and business outcomes. Raw telemetry may show that a component is behaving abnormally, but a knowledge layer should help an agent understand what services depend on that component, which customers may be affected and what downstream consequences could result from a proposed change.
The Human Side of Autonomous Networks
Swisscom’s (News - Alert) experience looks at the AI journey from a very different perspective that doesn’t necessarily start with the technology. The operator found that technology modernization alone was not producing the expected results. So, it first reorganized around small, self-sufficient teams with domain ownership, shared governance and clear accountability. That people-first restructuring created the foundation to automate operations across the organization. The model is a direct rebuttal to other AI strategies that address organizational design after the technology has been deployed.
Engineers will remain central to this operating model, but their roles will change. BT (News - Alert), VodafoneThree, and Deutsche Telekom all describe a transition from executing individual tasks to defining intent, designing guardrails, supervising agent behavior and validating outcomes. Importantly, accountability does not transfer to an agent simply because the agent performs the work. Rather, it remains with the teams that establish the boundaries within which the system operates.
Standardization, which Ensuarget already points to as an important piece of the AI puzzle, should be an enabler of speed rather than an impediment to it. Operators may be tempted to move quickly by building isolated agents for individual problems, but that risks creating a new generation of silos, incompatible interfaces and technical debt. VodafoneThree argues that standards work is necessary now precisely because it can prevent operators from having to rebuild fragmented AI environments in the future.
“In this phase of autonomous networks they are important, because the risk is that we try and innovate by riding the wave of AI-Native and we introduce the technology in a way that becomes not interoperable in the future,” Baiamonte explains.
Current Reality, Future Promise
The evidence is currently strongest around operational improvements. Fault management, service assurance and troubleshooting are attracting the most automation investment because they have direct effects on costs and customer experience. They are also easiest to implement quickly. Operators report reduced repair times, fewer tickets, faster anomaly detection and more efficient network operations. That’s all good news.
The commercial promise is broader but less proven. Some suggest operators will evolve from connectivity providers to intelligence providers. That could include context-aware services that adapt to users and environments in real time, enterprise offerings built around automation and decision-making, AI-optimized network experiences and intelligent edge ecosystems supporting machines, robots and vehicles.
In that future, operators will not merely sell bandwidth more efficiently. They will participate in the creation, delivery and monetization of digital intelligence. Whether they can convert that ambition into substantial new revenue remains uncertain, but it is the larger prize behind the operational work that AI-native commands.
So, the near-term business case for AI-native telecom is largely about making complex networks more reliable and economical. The longer-term strategic case is about changing what a telecom operator sells. In that sense, this isn’t so much about the AI-native telco having arrived, but that there is a difference between AI-native ambition and AI-enabled reality. Operators are moving agents into production, but the heavy lifting is yet to come around data architecture, organizational design, governance, interoperability and control.
Here’s the catch: Without that heavy lifting, telcos won’t get their tractor.
