Autonomous networking is moving from controlled trials into live operations, where every decision carries greater weight. Agentic AI can analyze conditions and act at speed, yet operators still need to know what an agent recommends, why it recommends it and what resources it can use. Glass-box autonomy addresses that tension by making agent activity visible, traceable and open to intervention.
Operational guardrails turn that visibility into control. Policies define what an agent can do, when it can act and where its actions apply, while logs connect each decision and tool call to the agent's identity. As Vivek Jaiswal puts it, "Operators can understand what an agent can do before execution, address the expected impact and intervene when required." That oversight creates a practical route toward Level 4 autonomy, where automated action can advance without removing human authority from high-impact decisions.
Nokia's partnership with NVIDIA brings this model into a sovereign agentic AI platform for telco autonomy. NVIDIA Nemotron models are combined with telecom-specific data and deployed through an architecture that can run on premises, giving operators more control over operational data, models and policies. That approach delivered more than 40% higher answer accuracy for radio alarm, fault resolution and troubleshooting compared with a model that had not been fine-tuned. As agentic AI moves deeper into network operations, progress will depend on combining that level of performance with the visibility and control operators need to act with confidence.
To learn more about Nokia & NVIDIA’s partnership, visit: www.nokia.com/autonomous-networks/#sovereign-agentic-autonomy-nokia-nvidia
Steve Saunders:
Everybody's talking about autonomous networking amongst my audience of carriers, and you've just had a big announcement with one of the largest companies in the world. What's going on?
Vivek Jaiswal:
We are announcing a partnership with NVIDIA where we are bringing a production grade agentic AI platform for telco autonomous networks use cases, and we are combining NVIDIA's Nemotron models to bring what we are calling as telco models for our customers out there.
Steve Saunders:
Now you are with NVIDIA making it possible for them to actually move forward with this at a very high level. What's their response been? Are they excited about it? Are they worried about it?
Vivek Jaiswal:
All of these capabilities which we are bringing to our operators is providing them immense amount of flexibility in terms of what they want to choose. And also very importantly, with all the agentic AI capabilities, it is providing them something real in terms of getting to level four autonomous networks and not something which is just inside labs. So this is real, Steve, and as real as it gets, as customers see this, there is definitely a huge sense of excitement on this journey.
Steve Saunders:
The concerns, I guess, at the moment from carriers that I talk to almost every day is about how to control network autonomy in a way that they're going to be comfortable with. And I understand that there are solutions to doing that. There are ways that Nokia can help carriers do that. One of them is glass box autonomy. And then also I hear a lot about agentic guardrails. I mean, can you start by just explaining those concepts a little bit and then give us a sense of how that can help operators maintain oversight?
Vivek Jaiswal:
Glassbox autonomy and agent guardrails as we call it, maintains operators oversight. Operators can understand what an agent can do before execution, address the expected impact, and intervene when required. Think of it like this team. Every agent decisions, the actions the agent is doing, the tool calls it is making, everything is logged. It is traced and measured against the acting agent's identity. So you exactly know this particular agent, this is what it is recommending, why it is recommending, and what is the audit trail as you use the word.
Now when you come to the guardrails themselves, you essentially define operational policies, the policies that govern what agents can do, when they can do it and where they can be applied. All of this in some way helps limit the blast radius. On top of it also puts a control on, provides a mechanism to put a control on what resources the agent can use. And as you rightly said, when a human approval is required.
Steve Saunders:
It sounds like we're shooting for level four here at the highest level and not level five because level five, I think I haven't met a carrier yet that wants level five. Is that accurate?
Vivek Jaiswal:
We are shooting for level four. But I think, Steve, there are two parts of what we are doing with NVIDIA here. There is of course a part of trying to hit level four and level four autonomy. However, I think what we are trying to do is this, we are trying to create an agentic AI platform which is available on premises. And the moment you have something on premises, it gives you the capability of having sovereignty, which is becoming very important for many operators.
Steve Saunders:
Give me an example of a specialized telecom model.
Vivek Jaiswal:
Well, we picked up a Nemotron 3 Nano, four billion model. And what we did was we tuned this particular model with some amount of data from inside Nokia. And using this, when we fine-tuned it, we kind of measured the answer accuracy of this model as compared to a non-fine-tuned model. And we find that in that specific use case, the accuracy of what the model gives us back is almost 40% plus better than as compared to a non-fine-tuned model. And this is for specifically we did this exercise for Nokia radio alarm, fault resolution, and troubleshooting use cases.