- Microsoft's Telco Industry CTO said AI anomaly detection use cases haven't panned out as expected
- Network operations and BSS applications are still fair game and key for telcos
- Level 5 autonomy will take time and require building trust with agents
Telecom operators chasing AI‑driven automation are starting to separate hype from reality, Microsoft Telco Industry CTO Rick Lievano told Fierce, noting the technology is already paying dividends in network operations and business support systems (BSS) — but has fallen short in at least one widely promoted area.
While AI‑driven anomaly detection and outage prediction have been widely promoted across the telecom industry, Lievano said Microsoft’s internal experience with this use case has been sobering.
The problem with massive network anomaly detection models is they’re incredibly complex and only deliver minimal improvements over traditional statistical models, he explained.
“The reality is that the amount of complexity that they add to your overall working set, plus the amount of false positives that you're constantly essentially playing Whack a Mole with, turns out that, you know, that wasn't the best investment in AI,” he said. “It didn't lead to the kinds of ROIs that we were expecting. The amount of complexity that introduced kind of wiped out any potential benefit that they provided.”
He said less than six months ago they still were promoting that use case as they thought it was viable. But time has shown it's not, so they're pivoting to promote rapid response rather than predictive capabilities. They're now focused less on predicting every possible failure and more on resolving problems quickly and automatically when they happen.
“It’s more important if there’s an outage that I’m able to completely automate, you know, and route around that outage as quickly as possible, and then resolve that outage as quickly as possible as well,” Lievano said.
Those kinds of lessons – where AI doesn’t work – were largely absent from the broader telco AI conversations at MWC last week.
Network ops in action
But that doesn’t mean all network AI use cases are a bust. He pointed to network operations and BSS modernization as two areas that show great promise for telcos. On the network operations front, Microsoft is again speaking from its own experience.
The cloud giant has spent roughly three years building what it calls Net AI, an internal system used to help run its global Azure network. In June 2025, Microsoft debuted its Network Operations Agent (NOA) framework and made the same agent‑based automation it uses internally available to telecom customers.
Rather than relying on a single monolithic AI model, the framework uses multiple specialized agents — each designed to act like a subject‑matter expert — that can work together to diagnose issues, interrogate network devices and take corrective action. Think field operations, ticketing, telemetry, troubleshooting and SONIC agents, all overseen and orchestrated by a Network Operations Center (NOC) Manager agent.
In the nine or so months since its NOA announcement, Lievano said Microsoft has made several changes to the platform – adding model routing, data management, broader context windows and MCP capabilities. The latter two are key for delivering the consistent, deterministic responses from AI agents that are needed for telco operations, he said.
“We’ve essentially built a platform for automating a lot of these network capabilities for really being able to more reliably run our network and maybe most importantly being able to address issues as quickly as humanly possible, using our non‑human friends,” he said.
Internally, Microsoft has an agent called Miles that independently works on fiber outages. Obviously, humans are required to repair a cut in the physical world, but Miles is able to dispatch, coordinate and direct the field operations teams required to fix things.
According to Lievano, telcos are very interested in Miles and AI for field operations.
“This is a huge one for them because it’s just a very time-consuming, expensive process,” he said.
At MWC last week, Microsoft pointed to live customer AI deployments, including with Vodafone and Taiwan's Far EasTone Telecom (FET), as evidence its approach can scale in real‑world telecom environments. FET, for example, has used NOA to embed agentic AI in its NOC and already nearly 60% of its NOC operations are AI-assisted.
Importantly, Lievano said operators do not need to retrain or fine‑tune large language models with telecom‑specific data to get started. Instead, Microsoft relies heavily on retrieval‑augmented generation and model selection at runtime via the aforementioned model routing capability. The latter helps balance performance and cost depending on what task the AI has been asked to perform.
“If you’re doing, you know, root cause analysis, and you really want to be able to churn through a lot of data, make good decisions and good reasoning, GPT‑5‑dot‑X models will be the best ones that you’ll use for that,” Lievano said. “But if you’re summarizing a case, some of the SLM models are exceedingly good, fast and cheap for that type of workload.”
AI agents take aim at BSS complexity
Beyond network operations, Lievano said AI agents are also showing strong potential in BSS, an area long plagued by complexity and customization.
“AI agents really have the potential for totally flipping the BSS world upside down,” he said.
Rather than forcing operators to rip and replace legacy platforms, agents can sit on top of existing systems and interact with them using natural language and open APIs. That opens the door for non‑technical employees — from product managers to customer service reps — to automate workflows without deep IT involvement.
For instance, an “order fallout” agent can monitor stalled or failed orders, identify the cause and take corrective action.
“It basically looks at the order queue,” Lievano said. “If the order fails or if it’s stuck, the order fallout agent can automatically take action.”
If the problem is simple, the agent can fix it directly. If not, it can let the order manager know and have them step in to take corrective action.
Lievano said the impact of AI for telcos goes beyond efficiency gains.
“We’ve also found that the quality of work improves, the job satisfaction improves,” he said. “People are not doing essentially mindless tasks and are focused on higher‑order tasks now.”
Road to autonomy
Like many operators, Microsoft is looking to AI to advance its journey to autonomous network operations – which has the potential to significantly lighten the company’s load given its network spans 600,000 kilometers of fiber.
Lievano noted that Microsoft faces many of the same challenges as traditional operators, particularly when it comes to managing multivendor, multigeneration infrastructure.
“We’ve got, you know, as much legacy [infrastructure], particularly on the optical side, as any telco out there,” he said. “We have hundreds of SKUs across the network that we have to manage our own network.”
He also stressed that a jump from Level 1 to Level 5 autonomy won’t happen overnight. A lot of the progress that will be made will be based on trust that is slowly built with agent use.
Internally, Microsoft has certain use cases across support, troubleshooting and field operations that are clearly at Level 4 today, but Lievano said the company has a “strict policy across our own network that no change to the network is to be introduced by an AI agent without human supervision.”
That said, it is eyeing fully autonomous Level 5 capabilities. “As our trust in our agents grows, maybe there’s going to be a subset of potential workflows that we may allow them to eventually do on their own,” he said. “We haven’t made any commitment or announcement related to that specifically, but that is within our current roadmap.”
As for where it might start, Lievano said network repair seems to make the most sense. After all, there’s less risk when the network is already down.
Read all of our coverage from Mobile World Congress 2026 in Barcelona here.