0

Telecoms is deterministic. AI is probabilistic. So what’s the path forward?

TM Forum’s annual DTW Ignite event in Copenhagen was, as always, excellent. Between keynotes, breakouts and interviews, I spent plenty of time in the media and analyst work room. There was something of a common refrain: if everyone is saying the same thing, what’s the story? That, I think, is the story.

The telecommunications industry has reached the strange moment when strategic ambiguity is giving way to executional anxiety. The direction of travel is no longer seriously contested. Operators need more automation, more AI-native operations, more agentic capabilities and, ultimately, autonomous networks. They need this because cost structures remain under pressure, network complexity keeps compounding and the next platform shift will not wait for the industry to feel entirely comfortable.

But comfort is not the point. Søren Kierkegaard, Copenhagen’s great philosopher of anxiety and action, wrote that “anxiety is the dizziness of freedom.” The phrase fits the mood of DTW Ignite because telcos are no longer anxious because they lack choices. They are anxious because they have to choose. The industry has spent years debating whether AI belongs in network operations. That phase is ending. As Nokia Chief Technology and AI Officer Pallavi Mahajan put it in an interview, “It’s no longer a question of should we or should we not.”

The harder question is what telecom becomes when it accepts that answer.

The central tension is easy to state and difficult to resolve: telecoms is deterministic; AI is probabilistic. Networks are engineered around availability, repeatability, standards, policy, rollback, accountability and provable behavior. Generative AI, and especially agentic AI, works differently. It reasons in probabilities and creates plausible outputs. It can infer, summarize, triage and recommend. But in its raw form, it does not satisfy the traditional telecom requirement that a given input must produce a predictable and safe operational action every time.

As AWS’s Telco CTO Ishwar Parulkar framed it, root cause analysis is one class of problem. If an AI assistant offers a wrong hypothesis, that is not categorically different from an engineer starting with the wrong hypothesis. The answer can be tested, refined and corrected. But network configuration is another class of problem. “You can’t be wrong,” he said. In other words, probabilistic intelligence can be useful before it is fully trusted, but it cannot be allowed to act everywhere in the same way.

This is where the DTW conversation became technically interesting. The industry is not trying to make telecom probabilistic. It is trying to build deterministic envelopes around probabilistic systems.

George Glass, CTO of TM Forum, described this as a shift from bolting AI onto existing processes to embedding AI into the way the business operates. The point is not to sprinkle AI across legacy workflows and hope for productivity. It is to reimagine operations around dynamic, AI-assisted systems while preserving reliability and availability. Glass pointed to digital twins as a critical control mechanism — model the network, test the AI-recommended change against the model and only apply that change to the live network when there is confidence it can be made safely.

That same logic is visible in the broader DTW program. In recent literature, TM Forum positioned autonomous networks as moving from concept toward “high-value use cases, real-world deployments, and a roadmap from concept to reality,” with Level 4 autonomy “closer than ever.” The organization also used DTW Ignite to push an AI-native extension of its Open Digital Architecture, including a governed execution layer for AI agents, secure agent interactions, guardrails, managed access to data products and LLMs, and Level 4+ demonstrations across fault management, 5G slicing and sovereign AI inference.

This is the practical answer to the deterministic/probabilistic problem. Agentic AI in telecoms cannot be an unbounded agent wandering through the network with root credentials and a vague instruction to optimize things. It has to be bounded by intent, policy, observability, simulation, approval thresholds, audit trails and domain-specific knowledge. It has to know when to recommend, when to act, when to escalate and when to stop.

Nokia’s DTW announcements fit this pattern. The company introduced agentic AI capabilities across its autonomous networks portfolio, including an Autonomous Networks Agent Library, updates to its Autonomous Networks Suite, enhanced RAN automation, and AI-driven frameworks for IP, fixed and optical networks. Nokia described the goal as letting operators phase in AI and agentic automation while maintaining operational control and trust in live network environments.

The “phase in” language matters. The leap of faith is not blind. It is staged, governed and use-case specific.

Qualcomm’s Alex Teper, senior director and head of network management products, made that point in the context of agentic RAN management. “That leap of faith needs to happen in a focused manner through use cases,” he said. “But we believe that by the time 6G is out there, it can happen at scale across the various engineering use cases that networks are dealing with.” Note: The conversation with Teper didn’t take place at DTW but was topically similar enough that I’ve included it here.

The 6G reference is important because the time horizon can be misleading. Operators cannot wait until 6G to learn how to operate with agents. They need to start now with bounded use cases that create operational muscle memory around troubleshooting, service assurance, energy optimization, customer-impact analysis, anomaly triage and closed-loop remediation in constrained domains. These are not glamorous, but they are how the flywheel starts.

Parulkar saw that shift clearly at DTW. Operators, he said, recognize they are in an inflection point and are not simply waiting to see what happens. They are talking to hyperscalers and chipmakers. They are hiring data scientists. They are building centers of excellence. And, crucially, in 2026 he saw “active deployments of AI picking the low-hanging fruit, picking use cases that really deliver value.” His advice was blunt: pick something, build something, expand. Do not get stuck in the paralysis that traditional telcos know too well.

That is the Kierkegaardian turn. Freedom creates anxiety because choosing creates responsibility. But refusing to choose is also a choice. Or, as another Kierkegaard line has it, “There are two ways to be fooled. One is to believe what isn’t true; the other is to refuse to believe what is true.” What is true is that the AI-native network is no longer an abstract future state. It is becoming the industry’s organizing principle.

Verizon’s June 24 articulation of its Level 4 autonomy ambitions is another example of the shift from rhetoric to operational framing. The operator said it is targeting Level 4 cognitive automation in parts of its core network and reported that closed-loop automation platforms executed more than 70 million autonomous network configuration changes in 2025. Verizon’s framing explicitly notes that AI is becoming “the operational control system of the network itself.”

But the hard blockers are not only technical. They depend deeply on operating model and business model — you can buy all the AI in the world but if your people can’t use it or sell it, all you’ve done is spend money. TM Forum’s own research highlighted the trust gap: 72% of CSPs surveyed said their AI is trustworthy, but only 14% could produce evidence to prove it. That is an execution problem. Trust has to be evidenced in architecture, process, governance and measurable outcomes.

The same applies to monetization. Mahajan’s point was that technology itself is never the value. “It is humans and their adoption that is value,” she said. The question is whether AI-native operations simply lower cost or whether they also let operators sell something new; things like assured outcomes, differentiated performance, resilience, sovereignty, quality on demand, etc…TM Forum’s Autonomous Networks Leadership Forum similarly framed Level 4 autonomy as moving beyond cost-cutting toward revenue growth tied to network quality, resilience and customer experience.

That is why Telstra’s Kim Krogh Andersen offered the right counterweight to shallow incrementalism — the automation journey. “I’m against hunting short term small use cases in telecoms,” he said. “You need to go for the foundational cases. If you don’t go for the big ones we will be in the same mess in five years. We need to move with conviction, or we won’t make use of the AI supercycle.”

Said another way, operators need small wins, but not small thinking. They need use cases that are bounded enough to deploy and foundational enough to compound. Root cause analysis is a starting point. Intent-based operations, service assurance, cross-domain orchestration and closed-loop remediation are the compounding path. Demos are fine and serve an important role, but the goal is a new operating model.

Djakhongir Siradjev, CTO of PI Works, put the moment in historical context. A decade or more ago, he said, even strictly defined closed-loop automation was difficult for operators to accept. “Just the fact that something is done in the closed-loop was already scary enough,” he said. It took “bravery” and “leaps of faith” to adopt those systems.

That is where DTW Ignite landed for me. The industry does not need another generalized sermon about agentic AI. Everyone has heard it. The story is that everyone has heard it, everyone is saying it, and now the industry has to act.

Kierkegaard’s leap of faith was not a leap away from thought. It was a leap after reflection. Telecom’s leap should be the same. No one is asking us to place blind trust in probabilistic systems, but adopt a disciplined commitment to build the architectures, operating models and business models that make probabilistic intelligence useful inside deterministic networks.

The anxiety is real. So is the freedom. Now comes the choice.