Guy Daniels, TelecomTV (00:14):
Right. Welcome back everyone to the AI Native Telco Forum 2025. Those of you here in the room with us, please come on in and take your seats. We just finished our networking break. We have one more session for you today and there you go. It's a good one. As you can see, embedding AI within the network. Plus we've got our second panel discussion of the forum that will follow this session and take us through to the close of day one. So let's get started and I'm going to invite our next two presenters, George and Philippe to come up to the stage and George is going to start us off. So Philippe and George, please come up big hand please, George. George is going to start us off and George is CTO for the TM Forum. George, over to you.

George Glass, TM Forum (01:09):
Thank you Guy and good afternoon and welcome to the AI Native Telco. I'm George Glass, I'm the CTO at the TM Forum. I want to start off with some figures, which I've written down in a piece of paper because if I didn't write them down, I probably wouldn't believe them. Let's say you're doing IP backhaul fault management and you're serving 300 million customers. Imagine if you could reduce your operation and maintenance costs by 30% and reduce your meantime to the repair on faults by 50% and reduce your number of tickets raised by 15%. Bearing in mind that you probably processed in the region of 18,000 tickets per month or the second case looking at your core network for fault management and complaint handling, serving 400 million customers, 40% reduction in your operational maintenance costs, 87% reduction in your meantime to repair for faults, 65% reduction in the complaint handling, meantime to repair, and that's against 4,000 faults and 1300 complaints per month. Those are real world use cases delivered by China Mobile, implementing an AAN level four solution based on the TM Forum, open digital architecture and autonomous network patterns.

(02:35):
We've developed the TM Forum as an industry association and we are focusing on three missions, composable IT and ecosystems, autonomous networks and supported by AI and data innovation. All of that runs in our open digital architecture, an architecture which is initially designed to help with the transformation of the IT estate, but as the network become software, increasingly we are applying the same patterns from a software engineering perspective to the management and operation of the core network, the transport network, the access network, and all of the network resources. It's developed through industry-wide collaboration. I have a very small team, but we work with our members who are from the service provider and the vendor side, implementing solutions that are then taken and put into production and fed back to us in terms of what works and what doesn't work. We refine those solutions and they become the design patterns that our members build.

(03:39):
We've seen the evolution of the telco estate. When I started in telco, I would describe it as the traditional telco era. You had a network, you bought a set of systems, you managed it with the USS and BSS. The systems very often dictated the processes that you use to manage your business. We moved into the digital layer of telco. We developed ODA decoupled architecture that's layered, that has a set of components that expose their business capabilities via industry standard open APIs, and effectively the business logic from the components then can be encapsulated within the APIs and used multiple times not only to support the telco, but also to move into the ecosystem to adjacent industries and actually build solutions that will enable you to grow your revenues and adjacent industries. We've now taken AI and applied it into ODA as we've started to design what we describe as the AI native telco, and this is not taking AI and bolting it onto existing processes.

(04:46):
This is embedding AI into the functions, into the processes, into the interfaces of the architecture from top to bottom and applying it appropriately, whether it's customer management, whether it's product and service orchestration or whether it's network resource configuration. We've developed the autonomous network architecture and that has stood the test of time. We published this in 2018 or 2019 and it's built around the concept of intent-based operations. So you specify what you want as an outcome as opposed to specifying technology from the network. You take that requirement as an outcome basis and the intent ontology translates it into a set of service characteristics and then at the intelligent orchestration layer, you take those service characteristics and through discovery you select the appropriate network technology to deliver against those service characteristics. And the intent on SLA coming from the customer, it leverages concepts such as the closed loop controls, self-healing domains, and wrapped network resources with industry standard open APIs, we developed the methodology to measure your autonomous network levels from level zero through to level five, and we've seen a large number of our members now publicly stating that they're going to build level four capabilities.

(06:11):
We've got the design patterns, we've got the solution packages in production now, and we've identified what we describe as a series of high value scenarios and our members are picking off those scenarios one by one and implementing those and actually getting the sorts of benefits that I was describing at the start of the talk. They're going to IP fault management, they're going to ran of energy efficiency, they're going to complaint handling, they're doing network planning and taking those subsets of processes that they need to run and operate and manage their business and taking those to a and level four. It's our ambition because there's a lot of focus on fault management to effectively go wall to wall fault management, enable our members to build a and level four fault management by June of next year across all processes and technologies associated with managing the faults and across all network demands and network technologies.

(07:11):
If we actually look at overlaying the capabilities and the functionality that we've been describing in our AI and architecture, we've got concepts like autonomy and intent, intelligent orchestration, but this is where I want to just get to the crux of the matter. Some of those functions today are rules based and that is perfectly sufficient for what you need to do for the next two to three years. Other functions actually use ai and I was off work for three months and I came back after three months and suddenly this concept of ag AI had appeared from nowhere. I was told that our open APIs were dead because everybody was going to use MCP. I'll tell you now, that's a myth. You need to apply AI and you need to apply the correct AI at the right point in your architecture where it's going to deliver business value.

(08:08):
At the moment, when we survey our members, what we actually see is that they are playing with ai, particularly agentic AI in the laboratories. They've got some proof of concepts running, but nobody has agentic AI running at scale. In the sort of 300 million type scenarios that I was describing, those AI capabilities are built on predictive AI and in some cases generative ai. So if I look across the architecture I've tried to highlight in the green, purple and blue robots, AI robots where I think based on our understanding and knowledge of what our members are doing, where the appropriate use of AI could be applied to different elements of your architecture. So at the front end, whenever you're doing intent-based operations, whenever a customer is going to describe the outcome that they want from the network, obviously you're looking at chat GPT type generative ai, you're taking a natural language, you're going through an intent ontology, and you're describing a set of service characteristics from it.

(09:15):
When I'm looking at something like intelligent orchestration that works really, really well with predictive ai, but whenever I'm actually trying to do the discovery of the demands that are available to deliver the service, then agentic AI potentially has a role to play. But be careful if I actually discover that I've got five potential domains and I'm going to discover those through agentic API with an MCP message, and then I'm going to select one of those domains because it's the one that most likely or is closest to the characteristics that I have to deliver. Whilst I'm delivering that because I've got an autonomous network, I want to keep an eye on the other four domains because if the one that I'm operating on starts to degrade or fail, I want to swap to the others, but I cannot afford to have a chatter going on between those four domains and my intelligent orchestrator and the current domain that I've selected because I'll tell you this, the amount of admin traffic that you're going to produce from your MCP and your agent to agent will actually stop your network working.

(10:22):
So be pragmatic and actually design it. Whenever you use an event based architecture, you discover the domains, you establish their status and you leave them alone saying, tell me whenever your status changes, select the domain you're going to use and run that as your session during your session. If things change, then you can switch domains, you can enhance or increase the domain you're using, but that will be in a controlled manner. We are letting danger out of the box if we just let agents run wild over our network and over our IT estate without proper architecture, proper engineering. You have been warned, we are building these solutions with our members. We take them at scale. China Mobile have 300 million customers using and level four solutions. We will build those with our members and prove that they work before we publish them as a guide for you to adopt. Thank you very much.

Guy Daniels, TelecomTV (11:22):
Thank you very much. Round of applause. Come and sit down with us, George. Right, well first of all, congratulations on the timekeeping there. I think that was 10 minutes and zero seconds. You win the prize for that. But on a more serious note, strong words, strong message, well strong message. I'm sure we've got some feedback from our audience on what George had to say there. I'm just shooting myself on the lights and we have one in the front row, Alex, in the front row in the middle here. Thank you very much.

Basma Driss, Deutsche Telekom (11:53):
Thank you. Basma Driss from Deutsche Telekom. So to which extent will TM Forum try to standardize AI? We had in the morning the opening question about how much standardization do we need around AI? So what's your view on that and what's your strategy also as standardization body?

George Glass, TM Forum (12:17):
Yeah, so we're building an AI framework effectively the guardrails that allow you to use AI safely at scale. That's what we do in the forum. We take the best practice from all of our members that is working or not working in their environments. We distill that and then we publish it as either an information guide or a guidebook documentation from the forum. It's the collective output of the industry and we are working on, we have the outline of an AI framework that works with generative AI and predictive ai. We're in the middle of actually developing the agent AI framework at the moment.

Guy Daniels, TelecomTV (12:55):
Great. Thanks very much indeed. Thanks for the question. That was great. Do we have any more questions for George here? Otherwise I will jump in with a little question myself. Don't think so, George, we've been, one of the themes or the threads that's emerged this morning has been around timescales and whether or not it's not zero to one kind of step towards the air native turco or there's iterative steps or there's processes or there's different parts. Does the team forum have a view on this yet?

George Glass, TM Forum (13:24):
Yep, I thought it might. Everything we do in the forum assumes it's a Brownfield site, so we assume you've got legacy that isn't depreciated. Every solution that we publish has to be able to accommodate that and it has to be incremental to steps. I cannot take or ask a CTO or a chief network architect to go to the CEO or CFO and say, give me 10 billion pounds to my build my autonomous network. I've got to build it step by step, which is why we've got the 50 plus high value scenarios or overall scenarios and 20 of those are high value scenarios and we are implementing those one by one at a time, and that is managed, it's controllable, and you can build out your autonomous network attacking the areas where you've got pain, where you've got problems, but where the return on that investment is going to be maximum for your organization.

Guy Daniels, TelecomTV (14:15):
Great. Thanks very much, George. Well, what we'll do is we'll move on now. So round of applause for George and then Philippe, we'll take over. Philippe, do you want to go to lectern or in this middle of the stage? And Philippe. Philippe Ensarguet is VP software engineering at Orange.

Philippe Ensarguet, Orange (14:31):
Alright, so I hope that everyone took spacesuit to jump with me on the adjunct carry rocket. It'll be the purpose of my presentation. So most of the time when I'm talking about agentic AI am always starting by saying that agent I for me has been born basically it's very far from being an adult and it's not even a teenager. So we don't reach the rebellion, I would say age. So we must be very, very cautious. So in my presentation, I'm very sorry, but I won't talk about production grade scale, deployment of agentic AI because it's not the truth. I will only speak about some explorations that we are managing right now since almost 18 months. And perhaps the most interesting for you guys here is what are the reality check. So to start with, I think that when you are having a look to the definition of what's behind agentic AI, I think that three words are very, very quickly popping to mind.

(15:40):
It's about autonomy, it's about adaptability, and it's about orchestration. But I think that we need to keep in mind that it's a very, very new topic on the timeline. We could also just look at the 2022 chat GPT and the story about the agentic AI is just after and during all the talks, all the great talks we got since this morning, I would say that a lot of technology has emerged. We talk about MCP, we talk about hway and I really perceive a kind of, I would say, situation that is not so comfortable from some of the speakers of the audience. But yes, I think that we must take for granted that all this is in highly dynamic ecosystem and of course it'll change. But for me change is also a lot of opportunity when I'm talking about AI agent, according to me, it's an autonomous system that basically perceive, take decision, make action with a limited amount I would say of human intervention.

(16:52):
An AI agent basically for me is made of three very, very important components. The reasoning loop that is basically managing the breakdown and how the task will be split the brain that is basically one or several models that are bringing, I would say the thinking and something that is truly, truly extremely important. It's about the tools and I use the hands and the eye because it's really about what are the tools in this situation. And then we got five very, very, very important characteristic, autonomy, goal orientation, reactive and proactive. The fact that it's interactive and something that is super, super, super important and you will understand this when I'm reaching the conclusion, is the task specificity. So with AI agent, if we want to enter into the agent care ecosystem, I think that we need to add extra layer. And moving from AI agent to agentic AI system is I would say in particular an architectural evolution because to have agent care happening, there are for me four pillars that are extremely, extremely important. Multi-agent collaboration and a total different kind of architecture. We know on the generative AI that is massively centralized, very, very big GPU racks and so on and with I would say agentic, I would say it's small, highly specialized agent using specialized model.

(18:30):
I think that on specialized agent, they don't need to know how to bake a cookie, they only need to have the reasoning capability on the task that is managing the task, the composition super important, the orchestration super important and something that has become absolutely totally incremental with the agentic AI is the shared context and if I want to move a step further, it's really about, I would say context engineering. It's really something that we learn during our exploration that the way you are managing your context at the scale of your adjunct system is basically what make your agentic AI system relevant or not. So I just wanted to bring you a picture where you could just scratch the surface of the level of complexity. We may deep dive in. So here, just take a simple system one that we use, check and fix my network issues.

(19:35):
We could imagine to have multiple AI agents specialize on different parts and domain of the network. They all need to talk, I would say with local resources, whether is it prompts, whether is it tools, whether is it, I would say datas, so I need to have access to, for instance, my network topology, my issue knowledge base, multiple network management system where we will wrap the API system. And also a lot of other, I would say tools that we are using in production, the GitLab, the slack, and of course the Kubernetes clusters that are the runtime of the network function. And here is just a simple, I would say extract. We could imagine a lot of agent that need to communicate all together to make the use case happening. And you will see that it'll be one of my key learning when you are in front of an AI agentic AI decision think twice, think 20 times before really jumping into the topic because I think that you don't understand the level of complexity in which you are basically jumping.

(20:44):
And what is extremely hard here is that if the issues we need to manage and will be only, I would say limited to the adjunct care system, fine, but it's far surrounding what is adjunct. And I will come back on this on my conclusion. So basically during the last 18 months we did multiple exploration and I extracted three use cases that I think could be relevant for you. The first one is in the lab. So in the lab we have basically a complete network system, 5G, we have a core, we have transportation, we have the run. And what we did is quite simple. We implemented multiple MCP that are exposing the services from every network domain that we directly integrated into an LLM. And we are entering with a simple question, my customer with this phone number has a quality issues near Dusseldorf center.

(21:42):
And then the model is using, I would say the semantic that is exposed to the MCP to basically have all the breakdown thinking to try to understand where he need to look first on the core, then on the run, then on the transportation to try to understand where basically the quality issue is sitting. And what is absolutely mind blowing is that it's a total a hundred percent large model without no telecom specification. And when you are extracting the chain of thought of the scenario that we implemented, I would say that it could be really like the way a team of network architect should have thought and honestly it's a little bit disrupting. The second cases is really about agentic AI for operations. So you know that at orange we decided to move for an horizontal model based on the silver architecture. And here the idea is really to bring tools for the guys that are making the support and support level one.

(22:51):
And basically we implemented what is called MCP silver, that is basically a bunch of eight, 10 different agents that are, I would say understanding what's happening inside the cluster and then connecting to the different knowledge base and to be able to identify is similaar problem have emerged and where the solution. And then the last one is TAIA. That is basically an exploration. We started the beginning of 2024, that is using agentic AI on the change and on the monitoring process with agent dedicated on the infrastructure, on the network life cycle management and on the observability. So basically what we learned from our exploration, so the first challenge, the first and the biggest one is about security. AAA is a must have. We need to have authentication, authorization, accountability, and in most of the time due to the regulation, we need to have a sovereign runtime environment to be sure that we are compliant with what the regulator is asking.

(23:58):
The third challenge, the second challenge is really about the trust. We need to have in mission critical system with the agent. We can read but we can write and are we how we can be sure about the action that an agent can make that is right in terms of purpose on our infrastructure and it's what is basically generating your revenue. So we clearly have a question about the trust and the reproducibility of the result. And the last one, in terms of challenges, yes, moving the POC was quite easy. Moving to production is a total, total different story and a total different scale. Three major learning, I already talk about the context engineering. That is basically a game changer if you succeed in your context engineering, you have chance to succeed in your agile QA scenario, human AI collaboration redesign. I think it's a total way to envi how basically the team on the support side, because it's in majority what we explore, we work with those kind of tools.

(25:03):
And the last one that is extremely, I would say connected with the topic that have been shared earlier on is really focused about the observability. We have multiple, I would say silos in terms of network domain, all fragmented, siloed, and it's super difficult to have something that is really consistent when we have so multiple vendors. And when you are moving into this area, the observability cross domain is absolutely critical. My last slide is basically just to say that I truly believe that agentic AI, all the transformative potential for technical networks, it's early days, it's promising, it's a super, I would say interesting journey and definitively we will continue moving forward, experimenting, exploring, because something that I really learned during all that those exploration phase is that if we want to be successful in the AI native telco world, it's super important to not reproduce what we did in the past.

(26:06):
Basically mainframe, pc, web, mobile AI, we all made the same mistake. Since we are jumping in a new era, we are starting by migrating and porting the system from the past into the new paradigm with all the weight and uncuff with what we did in the past. If we want to jump in AI native telco, we need to fully enter into the AI native telco. And the very, very major question you need to ask before starting an urgent care system is do you really, really need it? If you are on a very, I would say, task oriented process with that are well known, you don't need an agent AI system. If you are able to manage your AI system with a single model that is performing on what you want to do, don't jump into an AI system. And I just bring here a Q code to an article that I dropped last week where I review all the design patterns that are applying to agentic AI and the number zero is really think twice 10 times about jumping into AI agentic AI because we are early stage, we have challenge in maturity, data trust, regulation, standardization, and yes, I really enjoy all the discussion we got in the migration from cloud native to agentic AI because I truly think that cloud native is the foundation to support, I would say this AI native telco.

(27:32):
So basically here are the stories that I wanted to bring to you today.

Guy Daniels, TelecomTV (27:41):
Great. Absolutely fascinating. A terrific glimpse into the many work areas you are exploring there at Orange. Do we have it time for one quick question. I think for Philippe, if there's a quick question from the floor. There's one over here. Where's Alex? Over here? Front? Yep. Oh you have a mic do you? I've got some blindness.

Francis Haysom, Appledore Research (28:04):
Thank you Philippe. Thank you George. Philippe, we first met at KubeCon a little while back and one of the things about going to that type of conferences, it's very different from Telco, very sort of innovative, very fast, very iterative, et cetera. In terms of what you've gone through, how do we avoid what I would term is that you said get rid of the baggage behind us, but how do we both balance the sort of care and caution but also don't make that an impediment to us actually changing and changing anything? Do you have any sort of observations in that area?

Philippe Ensarguet, Orange (28:41):
I think that regarding your question, I would say that when we met at KubeCon last week, basically our discussion was about the standardization of the ecosystem. And I would say that Telco is a very highly standardized ecosystem, but siloed by siloed, if I may say so. That's why I would say a telecom operator like Orange. But a lot of other peers are like us currently looking for commonalities, commonalities, cross domain, and I truly believe that the cloud native infrastructure is one, the data layer is another one. Certainly the observability is another one. The lifecycle management, GITOps lifecycle management of the NF is also another one. So my point here is certainly something that I didn't have the time to tell in the conclusion, I'm really calling the network vendor ecosystem to truly rely on open and standardized. I would say approach, open source, standardized, open is extremely important. If we want to be successful into this direction, anytime we are deploying a new network system, if we need to do an adaptation, I would say it's basically wasted money because it's only for a single purpose. So here I think that we need more than ever to have this standardized approach, this horizontal model where basically we could really spend time and monies on what make us different instead of reinventing the wheel every morning.

Guy Daniels, TelecomTV (30:22):
Great. Thanks for the question, Francis. Great answer. Philippe and George, thank you both very much indeed. Run post for our guests and we'll move on to our panel.