Why Orange is treading carefully with agentic AI

“My brain is bubbling because every morning or every week you have a major turning point [in AI]. And then you have the reality check of day-to-day life.”

Philippe Ensarguet, VP Software Engineering, Orange, sums up how many in telecoms feel about the speed with which AI is advancing, and its potential to impact everything from network architectures and operations, to how customers will use apps and services.

When it comes to the everyday, Ensarguet and his teams have been busy getting to grips with the pros and cons of agentic AI by running early explorations. Results demonstrate agentic AI’s potential for transforming telco network operations, he says. However, while the technology works, there are several obstacles to scaling it within commercial operations, some of which Ensarguet believes will be best tackled through industry-wide collaboration on standards.

Securing magic

A primary concern is security. In one of its proofs of concept, for example, Orange explored agentic AI to perform end-to-end network troubleshooting.

“The model has an ability to generalize based on the semantic description of the service you're exposing to the MCP [model context protocol] that looks like magic,” says Ensarguet. (MCP is an open-source standard for connecting AI applications to external systems).

Behind the scenes, of course, instead of magic there was a traceable agentic AI chain of thought. Faced with a customer suffering from poor network quality, the model asked for the customer’s phone number, then went to the core on the user data plane to identify the radio access network (RAN) cell to which the customers had connected. Once the agent checked the RAN, it returned to the transportation layer to isolate the link affecting network quality.

“Basically, you have troubleshooting opportunity with root-cause analysis,” explains Ensarguet. And “agentic AI … is a direct entry point … into the very deep core of our network asset.”

As a result, “security in agent AI time will be even more important than today,” emphasizes Ensarguet. “And it's a triple A topic: It's about authentication, authorization and accountability.”

The fact that multiple agents may be communicating via MCP with data sources or tools wrapping network ports adds to the headache.

“You just scratch the surface of how complex the system is that we have in front of us,” says Ensarguet. “It’s a question of ingress/egress management. It's a question of observability. It's a question of having a kind of ‘gateway-fication’ of the different potential MCP exposure you may have on the different network domains.”

Data sovereignty is another concern. Here, Ensarguet, however, makes the point that AI agents work as a squad in which each one has a narrow specialized role. This means “we will have the opportunity to better fine tune the module for our own purpose on much more smaller models, not LLMs.” These much smaller, specialized language models, he hopes could run on CPUs rather than GPUs. This would allow telcos to create, on-premise, a level of sovereignty for the most sensitive data that surpasses what a hyperscaler can realistically offer.

When it comes to environmental impact, Ensarguet emphasises that Orange is aiming to use AI thoughtfully - selecting appropriate models for specific uses, optimizing deployment to balance efficiency and frugality, and minimizing environmental impact. One example of this he gives is how Orange is using AI to analyze its systemic effect on thousands of operator clients and millions of end users, thus reducing the carbon footprint of its networks.

Agents aren’t always needed

In the meantime, adopting agentic AI represents a shift for telcos that is much greater than that of moving from virtual to cloud-native networks, according to Ensarguet.

Beyond the technology, adoption is hard because it demands a huge change in mindset and ways of working.

Every major technological shift, such as the move from mainframe to client-server computing, has shown that “instead of embracing the paradigm of the new era, you start by migrating and porting your cornerstone from the last era,” says Ensarguet, adding: “It's where all people make the same mistake.”

He emphasizes that agentic AI is about much more than automation. It’s “about providing the right context for AI agents to operate effectively in complex network environments and reimagining workflows that enable agents and humans,” says Ensarguet. “It's about the intent. It's about the ability to choose the right tools you need to troubleshoot.”

Ensarguet particularly stresses the importance of context engineering; without context agents cannot be trusted to solve problems securely and accurately. Orchestration is another critical function that needs to be mastered as it determines how well companies manage agents' interactions with each other and humans.

And none of it is easy. Indeed, it is so complex that Ensarguet recommends carefully weighing where to use agentic systems, pointing out that predictable, structured tasks or single model don’t need them.

The cherry on the cake

Despite the complexity of deploying agentic AI, there are areas where it serves. Orange is in the process of building a horizontal, cloud-native network infrastructure. Today, between 15% and 20% of the operator’s network functions are cloud-native. As deployment grows, agents could have an important role to play.

“When you have to manage the level of complexity of cluster nodes across multiple locations … being able to rely on agents to help us grab the information, do the correlation, ask knowledge base, ask Slack channels, ask ticket and issues database, is extremely important,” says Ensarguet.

At the same time, “true cloud-native infrastructure will be important in an agentic future,” he adds.

“You can't be successful at agentic AI if you are not successful at automation. You are not able to be successful at automation if you're not successful at data. You cannot be successful at data if you're not successful at telco infrastructure,” Ensarguet explains. “So, I would say it's like a cake with multiple layers, and you don’t have the [agentic AI] cherry on top if you don't manage the rest.”

Agentic AI promises to usher in changes for vendors, too.

“Whether we like it or not, the vendors are a little bit reluctant [to advance] automation, because at the end of the day, the more we are asking for automation, the less they have opportunities for professional services,” says Ensarguet.

However, given the complexity of deploying, managing and securing AI agents, vendors may not have to worry about being displaced anytime soon.

“The move from the exploration side up to the production grid is for me, still quite a long journey.” In the meantime, when it comes to addressing issues around data, observability, or security, for example, “industry-wide protocols and best practices are vital.”