To embed our video on your website copy and paste the code below:
<iframe src="https://www.youtube.com/embed/qLtWNBX81Sk?modestbranding=1&rel=0" width="970" height="546" frameborder="0" scrolling="auto" allowfullscreen></iframe>
Welcome everybody. As you can see, I'm looking frighteningly like Andy Burnham up on there — that's for those of you in the UK who are following the latest political shenanigans. But I'm here to get the ball rolling and I'm going to be requiring some audience participation shortly, so prepare yourselves for a quiz. Okay, first question for you in the room — who is familiar with the Roman poet Horace? Hands up. Okay, a couple of people with a classical education. But like most of you in the room, I was not familiar with Horace. According to the history books, he lived just over 2,000 years ago and was a friend of the Roman Emperor Augustus and the poet Virgil. Horace wrote about love, friendship, philosophy, and the fleeting nature of time. Okay, second question — who is familiar with the film Dead Poets Society? Okay, a lot more hands.
(01:31):
Me too. And one of the achievements of that particular movie is that it brought into the common modern consciousness the phrase carpe diem, which was first written by Horace and which translates as "pluck the day." But in the film and in popular culture, it's been interpreted as "seize the day" — to grab life by the horns, so to speak. And I think carpe diem is an apposite phrase right now because the telecom sector is, I think, at a crossroads that requires it to be proactive beyond its traditional borders, outside of its comfort zone, and quickly. That gamble's been taken before and the industry has the scars to prove it. But this time I think it's imperative. Telcos need to evolve in order to remain as relevant to society as they have been up to now, and that requires decisive action, guts, determination, skill, experimentation, and of course, investment.
(02:35):
Now, there's nothing in that list that isn't already in the day-to-day of this sector — it just needs to be ramped up a few notches and within shorter timescales. And in my view, we are starting to see evidence of that happening. Of course, the catalyst for all of this is the new era of AI. And as Guy mentioned, AI is going to get a lot of references during the next couple of days. It's also been at the heart of many of our free-to-download editorial reports, our DSP Leaders reports, during the past year, and you can see a few of the titles here. Now, one of our most recent reports was the 26-page Telcos and AI report, which was based on the results of an industry survey completed by almost 260 people who work in the telecom sector. So ahead of our first panel session, which is focused on scaling AI deployments, I thought I'd share one of the many results from that survey.
(03:37):
As you can see, we asked: should telcos be developing their own AI applications? And as you can see from the results, almost two-thirds of our respondents said yes. Now, I actually thought it would be more than that. We also asked our respondents to share any thoughts they had on the topic of telcos and AI, and here's an interesting one that was submitted — everything was anonymous. I'm just going to give you a couple of seconds to take a look at this and take it in, because there's an interesting phrase in there: proof of concept purgatory. And that's food for thought ahead of our first panel. But before we get there, and to go back to Horace, here's a couple more Latin phrases that I think are becoming more and more relevant in these AI-centric days. Caveat emptor — let the buyer beware. And qui bono — who benefits?
(04:40):
But enough about Bono, because later on we'll be talking about the edge.
(04:48):
I'm sorry to say I am here all week, but that's the only lined-up gag you'll be getting. So with that, it's now time for our first panel session. Could the panellists please join me on the stage, and please give them a big round of applause as they do so. So our first session is on scaling AI deployments. Now, telecom operators are overwhelmed with AI pilots, but they lack business transformation. And despite numerous AI-based network optimisation, customer service, and operations trials, measurable value remains far too low. The industry must transition from experimental AI to production-scale transformation to re-architect network operations and enable profitable new services. So to discuss all of this, let's meet our panellists. I'm going to ask them to briefly introduce themselves, starting at our far left.
Ohad Barnoy, Amdocs (06:13):
Thank you. My name is Ohad Barnoy. I lead the GenAI and Data Solutions for Amdocs, working with all our global customers.
Ray Le Maistre, TelecomTV (06:22):
Okay, excellent. Welcome.
Ahmed Hafez, Deutsche Telekom (06:24):
I'm Ahmed Hafez. I'm leading strategy in group technology and also leading AI networks at Deutsche Telekom.
Amol Phadke, Tech Mahindra (06:32):
Amol Phadke, Chief Transformation Officer at Tech Mahindra.
Ray Le Maistre, TelecomTV (06:37):
Okay. And our co-host for this session is Philippe, and Philippe is going to introduce himself from the podium as he gives our opening DSP Leaders address.
Philippe Ensarguet, Orange (06:49):
Okay. So predictive maintenance, root cause analysis, next best action, run optimisation, network tuning. I think that we have a whole lot of POCs, exploration projects, and sometimes we also have real deployment cases. Perhaps what is missing the most is what we are calling the measurable outcomes. For me, the question today is not if the technology will make happen our AI transformation, because of course it will address the topic. My question is much more, basically, how we will really scale it. The question for me today is really about the scale. My name is Philippe Ensarguet. I'm an Orange Fellow and VP Cloud and Software Engineering at Orange. And with my team, we are basically working at the crossroads of networks, artificial intelligence, cloud, and automation. And my honest take — my first words are very simple. I think that today the most important production gap that we have is absolutely not technical.
(08:04):
It's absolutely, for me, organisational and structural. Today, basically, if we are doing the maths, we have the compute power, we have the models, we have the algorithms, and even sometimes we have clean data from time to time. What is really missing today is, I would say, the organisational wiring. It's about the operating models. It's about the ability to blend teams — to have people who are very knowledgeable on the business side alongside those who are able to manage all those deployments. According to me, if we want to target production-grade AI, we need three core components. The first one is clean data. The second one is programmable infrastructure. And the last one is AI output connected to the real decision and to the real owner. In most cases, when I'm talking with my peers, I think that at least one of those three ingredients is missing.
(09:16):
Today, I think that with this first session and this first panel, we are really talking about the hard stuff. We will explore and talk about the barriers, about the architecture patterns, we will talk about the operating model, and we will also talk about how we can communicate the value to our executives and to our partners. So basically, I'm really keen to sit back with my peers and start the panel.
Ray Le Maistre, TelecomTV (09:47):
Okay, round of applause for Philippe. Okay, fantastic. Great start to our opening panel. So let's open this up to the rest of the speakers to find out what specific organisational, technical, and cultural barriers still exist that prevent successful pilots from becoming operational. Ahmed, if we start with you.
Ahmed Hafez, Deutsche Telekom (10:19):
Right. So there are things that are common between AI and non-AI. If you start a proof of concept, every one of us knows that it's very easy to do. Much easier — no constraints, no integration, no demands, no security checks — all that stuff can be done much more easily. So by nature of what a proof of concept or a pilot is, it's very easy to be done. But we always recognise a huge gap in order to go into production. This gap is not only about getting things done properly — you have all the, let's say, hard gates you have to pass: security, works council approvals, regulatory requirements, a lot of stuff that you have to go through. So if you are not planning that the proof of concept is just a step in a wider process, you will fail, because the gate of accepting the result and going into production is a huge one.
(11:18):
So that is not AI-specific, to be honest — that's in general. But if we go specifically to AI, I would just pick two things. The first one is the return on investment of what you are going to put in. You come and say, what is the business value? And here you compare the delta — we are not all sitting here waiting for AI to arrive while everything was done manually. No, there is a lot of automation already in place. There is a lot of stuff that has happened, maybe with scripts, with other ways, and the fact is we are not starting from scratch. So you always ask: is there enough delta? Is the effort worth the value that I get back?
(12:01):
And you get stuck in those conversations sometimes. The other part is trust — that's the second one. And the trust issue is pretty significant, because even as we talk about making AI more reliable, we are actually transitioning from deterministic automation to probabilistic automation. That is a big transition. And many people — rightly so — are sceptical about it. Employees don't simply accept it unless they see real value. So these are not easy hurdles. Those are the two AI-specific ones, but there is also the general one as I was saying.
Ray Le Maistre, TelecomTV (12:38):
Okay. Ohad, do you come across this scepticism in your dealings with operators?
Ohad Barnoy, Amdocs (12:45):
So it's very interesting because we don't do POCs. We don't believe in POCs, and I'll explain what that means. A POC by premise is: I want to validate that this works. What we believe in is what we call pilot to production, where the concept is different both technology-wise and in what you're trying to evaluate. You are evaluating three things. One — does the solution actually work technically, does it work as expected? Two — does it work in my complex environment? Deutsche Telekom's environment is very different from Orange's technical and cultural environment. Does it work in my environment? And three — does it bring business value? Those are the three measurements we put into a pilot to production. The reason we believe in pilot to production more than a POC is because it takes the same time. It takes today four to six weeks at most to do a pilot where I am integrating with the underlying systems, actually showing that the solution can work, and measuring a KPI of success.
(13:40):
And just like Ahmed said, the discussion needs to be: let's assume for a second you're sceptical — excellent. Let's assume for a second this will work. Let's assume for a second there is an ROI behind it. Do you have the path to get this into production? You need to make sure that path is understood before you start theoretical POCs. Telecom companies are bombarded with POCs from every possible vendor. I'm sure they hear 20 times a day, "I have this amazing technology, let's POC it, let's try it." I'm saying: don't waste time on another POC that works in a lab environment and shows you something cool. Try it in your environment. It'll take you the same amount of time and you'll actually be able to make smart decisions about where to take it forward.
(14:38):
This will also drive a cultural change. If people think that way — that this isn't a POC where I don't believe it but I'm letting them try to prove it to me — but actually I have skin in the game, I'm going to connect it to my systems, I'm going to see that it works — that drives a cultural change where people feel ownership of showing that this is successful and can actually drive value.
Ray Le Maistre, TelecomTV (14:59):
Okay. Amol, you've been on both sides of the fence in the past couple of years, so you have lots of experience, and I know you're doing a lot of work right now on these kinds of things with operators.
Amol Phadke, Tech Mahindra (15:12):
I think it's quite interesting. If you look at pilots, as Ohad said, I can almost guarantee that the first set of pilots done by any operator are hugely successful. I have not seen anybody say the first pilot on AI failed miserably. And there is a reason for that — whenever you start doing a pilot, you want to create momentum in the organisation so you can then cascade success. But the big impediment in terms of your original question — how do we then move to production and scale — is that I still believe there is a huge talent gap. The first pilot always has your best people on it. That's how you do it to guarantee a higher probability of success. Your best data scientists, your best network engineers, your best IT engineers all work together on the first two or three pilots. Then when you go into the wider organisation and say, "Now I want to do a thousand of these across a hundred thousand people," you suddenly hit this issue of a step-down in terms of understanding what needs to be done and how.
(16:19):
So I do feel the industry hasn't focused on the massive talent upskilling programmes that would be needed to get us to the point where everybody can scale at the same time. Otherwise, it's going to be sequential — which I've observed — and that will just add more delays in getting things into production. That is one thing we just have to not gloss over.
Ray Le Maistre, TelecomTV (16:39):
Okay. Well, that's a good point on which to come to Philippe, because I know — having spoken to a number of executives from Orange in the past year — there's been a really big focus on upskilling within the organisation. Is that right?
Philippe Ensarguet, Orange (16:53):
Yes, yes, totally right. And Ray, if I remember your question, you talked about barriers, and from my standpoint, I think you highlighted three things: the business, the technology, and the culture.
(17:08):
From a business focus perspective, I think our telecom industry is heavily tech-pushed. We are really driven by technology and by vendors, and most of the time when you are walking the aisles at any kind of event, people are not selling solutions — they are selling technology. So I think if we want to get the AI thing right, the first priority on the business side is really to focus on the value we generate and the business case, before starting anything else. The second topic is technology — because the question is related to scale. When you are a multi-country telecom operator, it is very, very hard in the field and in the trenches to align everyone. So you really need to make a wise decision around what I'm calling the AI blueprint — where you identify the big blocks, the big patterns that are non-negotiable if you want to scale your business cases.
(18:12):
And the last point, which is very related to your remarks, Ray, about upskilling — from a culture standpoint, I think it's extremely important to train people and get their hands dirty with AI. Because honestly, when I'm talking with some of my peers, I often have the feeling that they are talking to me about AGI — something that can absolutely do everything and solve almost all our problems. But the reality is quite different. We need to understand when to use machine learning, traditional AI, generative AI, or agentic AI. We really need to train and educate our people, and that will be the opportunity to blend teams — business experts, technology experts, and transformation experts — that need to come together to succeed in the scaling exercise ahead of us.
Ray Le Maistre, TelecomTV (19:16):
Yeah, I mean, it's always nice to look to the future, but we have to deal with the now. And as is mentioned a lot in conversations around AI, data is very important — you mentioned that as well. So let's shift a little bit to the technology that underpins what's going on at operators and look at what kind of critical infrastructure, operating models, and data architectures are needed to move from fragmented pilots to enterprise-wide AI deployments at real scale. Amol, let's start with you on what you're seeing out there at the moment.
Amol Phadke, Tech Mahindra (19:50):
I think the operating model point that you mentioned is quite critical, because we have seen a number of different data models being tried. You have the centralised data model, where you have a common data warehouse managed centrally at a group level and use cases are driven from it. The challenge with that model tends to be that individual data silos still exist — you made the point earlier about silos in the countries — and that creates the fragmentation we're trying to solve. The other approach, allowing a completely distributed data model, also creates issues because you then get so many duplications and variants of a data lake that it becomes almost impossible to achieve org-wide scaling. So I do feel a federated data model is the answer.
(21:00):
I had the fortune of working with hyperscalers, as you know, and they have a huge advantage because they always had the federated model in mind — a centralised data model with governance, with individual teams federating out to create use cases within the guardrails. So I do think a federated data model would be very critical in terms of achieving that scale. Otherwise, fragmentation will impede any real progress.
Ray Le Maistre, TelecomTV (21:31):
Okay. Ahmed, what's been the approach at Deutsche Telekom? This is something the industry has been talking about for decades — sorting out data management, cleaning up data, and so on.
Ahmed Hafez, Deutsche Telekom (21:31):
Right. So the approach we're taking — if you look at AI, the success of AI, and also scaling — scaling AI is different from scaling other things, and I'll explain that first. Basically, scaling AI has the challenge that things are changing so fast that technology you see today could be obsolete in a month. How do you deal with that? We are used to putting plans that span years, not months. So we are now transforming into being much more agile — doing things on the go. And that's why, in my opinion, there are three cornerstones for the success of AI, and this is what we're trying to do at Deutsche Telekom.
(22:37):
We have the data element — making sure we have the right, unified data. How do we do that? Clean up the data, make sure we have the right data products, contracts, and so on. We have the enablers, and the enablers are not to be underestimated. This is a huge part of the story and usually not really looked at properly. If you want to get into agentic AI automation, you need a context and semantic layer. This context and semantic layer is very hard to buy — you cannot buy it from a supplier.
(23:10):
You have to build it, and that makes things a lot more complex. And then you have the use cases — proving the value — because organisations won't wait for a few years while you build everything. You'll never get the data 100% right anyway. You will always be improving, adjusting — you can even change architecture on the move to adapt. So we are working on all three in parallel, not trying to get stuck waiting for each dependency before moving to the next. That's not going to work.
(23:41):
What we did is: you can do some use cases. You need specific data. We get the data with the best possible architecture available to make sure the use cases are proven and running. And then in parallel we are building the proper architecture, so that once it's available, we migrate the data and the use case still runs. And then for solving the problem within the use case — instead of making agents talk to each other directly, we make them talk through an A2A server. Instead of connecting to certain tools and data directly, we connect with MCP. So you start adjusting your architecture and scaling it as you go.
(24:12):
If we wait until we've built everything and then build use cases, we will not get there — because no one will have the appetite to wait, and there is no direct return on investment from data and enablers alone. The return only comes when you implement a use case. So if you want to kill the use cases and kill the story, you wait. Waiting is not helping. You need to work on all three in parallel. And we have proven this works across multiple use cases that are up and running in the network, including agentic AI.
Ray Le Maistre, TelecomTV (24:46):
Okay. Ohad, are you seeing lots of different approaches to models and data management at different companies? Because as you mentioned, everyone is different, so you would imagine there would be lots of different approaches.
Ohad Barnoy, Amdocs (24:46):
So I think what is really interesting is the elephant in the room around data. Before AI, the issue was that network never wanted to share data with IT. IT did not want to share data with the network — they said, "Tell me the insight you want, and I will give it to you. You want to predict churn? I will tell you what the network score is. I am not going to give you the raw customer experience data from the network." This made sense in the pre-AI world. Today, that has to change. And it is going to change. Because now when I am building a semantic layer, a context layer, an ontology, and even exposing MCP servers to the data, I am going to have to be able to expose that.
(25:52):
Now, Amdocs builds ontology, context, and semantic layers on top of BSS and OSS, because that is mainly where we play. Some of our customers are asking, can you expand this to a telco-grade ontology? And we work with them to see whether it makes sense and there is value in doing so. But I think the most important thing is that people are realising that as you start to talk to your data — whether through an agent or through a human — the ability to drive business outcomes is going to require that layer that Ahmed was talking about, which has to be built in. Some customers are very advanced. Vodafone Group, for example, has already built an entire data lake.
(26:27):
It already collects from the different OpCos and allows them to make real advances. On the other hand, I was in Singapore last week — with Singtel — and also in Manila with other customers. They are taking very different approaches. In some cases, a more siloed approach, because they feel that is the right way to approach AI in their domains. But they are all still looking for a semantic layer, a context layer, and the ability to have an agent access the data through agent-to-agent or agent-to-MCP. And in my view, that is going to bring a huge cultural change. If I'm head of a network for a telco, I'm going to have to share a lot more than just the insight — the AI transformation is going to drive a cultural transformation.
(27:35):
And I think that is going to help evolve and change the operating model.
Ray Le Maistre, TelecomTV (27:35):
Okay. I mean, we hear a lot in this industry about collaboration. Philippe, I saw you nodding there about the bringing together of IT and network teams. Are we seeing internal collaboration now, as well as external collaboration?
Philippe Ensarguet, Orange (27:35):
I think that collaboration is getting more and more real, and for multiple reasons — because the question always comes back to the topic of scale.
(28:05):
For me, there are at least three topics that are very, very important. You can't scale AI faster than the weight of your legacy. You can't be successful on AI if you are not successful on automation. You can't be successful on automation if you're not successful on data. And you can't be successful on data if you're not successful on infrastructure. One thing that could turn very bad is thinking that AI is a new feature. It's not a new feature.
(28:36):
It's a new layer, and it's not something that can be built apart from what exists. It's a new layer that needs to live inside what is existing. So being an AI-native telco company is impossible if you are not already successful on the cloud telco transformation, because that is the substrate we need to make the transformation. And I'm strongly involved in the SYLVA project — we released a very comprehensive white paper a few weeks back, where we advocate how much we think the SYLVA project could be a perfect substrate to bring the AI layer onto the path where our most sensitive network functions are living.
(29:07):
The second point is that with AI we won't be in a greenfield environment. The real life is brownfield, and we have all the layers of our history and vendor relationships. So you can't have an operating model that works for the brownfield which is completely different from one that works for the greenfield. If you want to keep coherency across the parts of your infrastructure and network where you have legacy running alongside new AI capabilities, you can't have a disconnected operating model for both.
(29:43):
And the last point — and Ahmed, you mentioned things that are very, very important to me — on the question of scaling, there are at least corporate-wide topics that need to be tackled, otherwise scaling is impossible. The first is security. If you're not able to bring authentication, authorisation, and accountability, you can't scale. You need end-to-end observability — of the infrastructure, the network workload, and also the AI — across all layers. You also need to address guardrails and policy: what will be possible in my network with AI, and what do I want to govern.
(30:49):
So yes — security, observability, inference isolation, guardrails, policy, and certainly many more. But those ones are absolutely critical and must be thought of as enterprise-wide assets. And it's one more time a reason for me to say: we can't be successful building an AI stack as something separate. It necessarily has to be built on top of something existing. Otherwise, we will create a new silo — and I think we have enough silos in the telco ecosystem.
Ray Le Maistre, TelecomTV (31:55):
Okay, fantastic. Now, we're going to come on to the business angle, because obviously this is important in everything we talk about. But as Guy mentioned at the beginning, we are going to try and get some audience interaction here. So if you have a question, get ready to put your hand up for our roving mic in just a minute. But first, I want to come to the topic of how we measure and communicate AI's impact on revenue enablement, service velocity, and customer experience — because there have to be proof points, ways to measure this. Ahmed, is there any kind of guide or blueprint the industry can look at to see how well they are doing in terms of AI implementation?
Ahmed Hafez, Deutsche Telekom (32:31):
Right. So of course there are lots of KPIs that could be used. Let's go into categories first. There are three categories of value you can get from AI. Of course you can categorise further, but at least the three big ones are: customer impact — whatever the customer impact is, you have KPIs like a customer experience index you can define, and you can get from direct network KPIs and understand on an individual customer basis what type of experience they're getting. You have the NPS score, the churn numbers, the uptake in certain services, if your use case is addressing or expanding those. That's the customer category. That's number one. The second one is efficiencies. This is very easy and usually the one we look at first — although, honestly, I think it should be the last one we look at, because we rush into trying to find out whether AI can reduce our costs — OpEx, CapEx, FTEs, and all these elements —
(33:35):
but we rush so fast that we even cut before we have the AI in place. That's where the danger comes. The third one, which is in my opinion also one of the most important, is technology excellence. What do I mean by that? Meantime to detect a failure, meantime to find the root cause, meantime to resolve a problem. Number of instances you get through — you can actually, with AI, significantly reduce the number of incidents. How many night shifts do you have to do? These kinds of KPIs give you excellence, but that excellence results in a more resilient, more reliable network — and of course that's a better customer experience.
(34:09):
The reason it doesn't just sit under the customer category is that there is an excellence element where you also reduce the impact on your people. How you run your team, how you run your networks, how it becomes more reliable — we've realised in some use cases that we could do roughly 40 to 100 times more in certain aspects. These aren't structurally possible without AI. So you find that with AI, it's not about the margin you create — it's about the pie you enlarge. You are enlarging the pie of abilities that you have. Maybe you need in some cases to hire more people to run this AI, to run the network, and to run things. But definitely your pie is growing — your capabilities, maybe your revenues, what you can deliver to the market.
(35:16):
So time to market is also a crucial element. Sometimes we need to do something that takes six months — now with AI it can take six hours. When you look at that kind of analogy, you understand that can have a strong business impact. All of this is measurable. And what we do is have every use case address some of these. We need to highlight that from the very beginning. Assume success — what is the outcome? We put that down, our aspiration is to achieve that outcome, we work towards it, and we measure. Did we get that outcome? Better or worse? Is it passing the threshold of value? If so, it's kept. If not — what's the problem? We go back and find the better solution.
Ray Le Maistre, TelecomTV (35:59):
Okay. Amol, are you seeing any quantifiable measurements in the industry now that can help to assuage fears that this might be money going down the drain — which I'm sure some people might think?
Amol Phadke, Tech Mahindra (36:14):
I think Ahmed made a great point about the fact that the industry is obsessed with efficiency, and that's a phenomenon we see across the industry — we look at AI as giving us a huge TCO benefit when in reality it is actually service velocity that would really drive AI scaling. The ability to launch products faster, to launch features faster, has been the biggest issue for the industry for many decades — taking months as opposed to days. And Ahmed also mentioned the MTTR issue. If you can really move to seconds, minutes, and maximum hours for service velocity, I do believe that putting that steel thread of using agentic AI to drive better outcomes from a product development, service development, and feature velocity perspective — that is what we should focus on. The rest are consequences of doing that. But we sort of are going about it the other way around at the moment, and we need to pivot.
Ray Le Maistre, TelecomTV (37:24):
Okay. Ohad, are you seeing timescales change? Are you seeing strategies change as a result of early implementations and early experiences?
Ohad Barnoy, Amdocs (37:36):
So I think, from an Amdocs standpoint, a year ago when we realised we needed to do a cultural change, we actually established a new division for data and AI — collected everybody in the organisation that works with data and AI under one roof, product delivery, R&D, operations, and so on — to drive that cultural change that can actually lead to those KPIs. And I remember our first KPI, because we needed to find a KPI — it was the simplest one: adoption. It's very easy to measure how many people are actually using AI tools and so on. Very, very quickly we started adding more and more KPIs. And to both my colleagues' point about operational efficiency — we weren't looking at that to begin with. We were actually looking at how we improve velocity, how we make sure that as a software company, our software reaches our customers faster, we're able to deliver more story points per sprint, a lot fewer defects, and so on.
(38:30):
I think a lot of our customers are also software houses in their own right and should look at that aspect too. I agree that the KPI of "can I predict an outage 24 hours before it happens at 95% probability" is a game changer with AI — even with ML we weren't really able to reach that, and those are very important. But I think we also need to look at the very basic things. How am I actually measuring the KPIs of my people leveraging AI in their day-to-day? Because at the end of the day, I'm hoping that AI doesn't end up replacing everybody, but works side by side with us — to make us a lot more efficient towards the outcomes we're talking about: better customer experience, increased revenue, reduced costs overall, while working together with us.
(39:29):
So I think the focus for us is about how do we make sure that our employee experience with AI is good — that they don't have to look for tools, we find the tools for them before they even ask for it. So I don't have people running and saying, "Can I have a licence for Cursor?" We make all those available because we want to make sure that people can be efficient with AI and not see AI as a threat. And for us, those are the KPIs. And I think it applies also to a lot of the telcos that are also techcos — that have big software development shops — that should look at those aspects too.
Ray Le Maistre, TelecomTV (40:02):
Okay. Philippe, what kind of targets — what is the Orange CFO expecting, or the board — in terms of what do they want to see?
Philippe Ensarguet, Orange (40:13):
I think it's a great question. I may answer with one word — it's a question of maturity. Because we did not even scratch the surface on the maturity topic. We are here on an AI panel, and gen AI is three years old, agentic is almost one year old, and we are here talking about potential outcomes and benefits like it's a technology we deployed decades back. So I think we must be very cautious about the maturity topic. And that's the reason why I'm perfectly aligned with what you mentioned, Ahmed, about the families or the areas where we may have benefits. We have basically two pockets: the pocket of what we have — efficiency — and the pocket of what we could have — customer value and revenue generation. On the efficiency part, we know we got a threshold, but at the end of the day, it's where we could have the first major tangible result.
(41:17):
And that's why when you talk about mean time to detect, mean time to repair, time to value — those are indicators and KPIs that are much more rooted in what we used to do. And perhaps it's also a question of accepting the need to dissociate the team working on the efficiency topic from the one working on value generation and customer focus, because the angle for me is absolutely not the same — even if they can share part of the technology stack, they are different.
Ray Le Maistre, TelecomTV (42:02):
Okay, excellent. So I did promise we would come to the room. I'm going to have to rely on my colleagues to see the hands — Tony's got the mic. Tony, if you can find a hand that's up, and whoever's got the microphone, if you could say your name and a short question, and who you're asking the question of.
Neil McRae, CityFibre (42:29):
Neil McRae from CityFibre. So it's actually more of a general question for the whole panel. One of the things you've missed out is the actual journeys in a telco and how inefficient they are. And actually, you don't need to build any AI systems. You can use AI to do task and process investigations, and you'll win there more than you'll win on any other systems. I just wonder — you mentioned using AI to rewire your business, and I kind of wonder where the thinking is on that. It's all very much "hey, we need to build stuff, we need to develop stuff." Actually, I predict that in every single company there are about 50 apps where there's an AI button you could press and immediately start seeing benefits.
Ray Le Maistre, TelecomTV (43:15):
Okay. Is AI rewiring your business, Ahmed?
Ahmed Hafez, Deutsche Telekom (43:20):
Of course. It's a very good point, Neil. This wasn't on the question list so we didn't discuss it. But yes, indeed. It's the difference between effectiveness and efficiency. Sometimes we have to look into the process and redo the process entirely. We saw that in fibre — in the fibre factories we're building in Germany, we had to put a lot of AI in. But when you come to put AI in, you have to look at the process: what's going on? Is this the right approach? Is AI transforming that process? Maybe bypassing steps that we don't need anymore? And the fact is, yes, in many cases it's transforming the way you approach things — not only removing humans from certain aspects, but completely eliminating some steps. So indeed, yes — the short answer is yes, it is.
(44:06):
And Amol?
Amol Phadke, Tech Mahindra (44:09):
Low risk, yes.
Ahmed Hafez, Deutsche Telekom (44:10):
Sorry, the lights in my eyes — thanks, Amol.
Ray Le Maistre, TelecomTV (44:14):
Any more quick responses to Neil's question?
Ohad Barnoy, Amdocs (44:18):
Yeah, I think this is one of the unique capabilities that AI brings. Ahmed mentioned before the deterministic automation — what we had until now, which a lot of us are still using — where we would take a process and automate it with RPA or whatever, and usually it would be the same process we would automate. This is exactly where with AI we can come and rewrite and rewire our business processes and our technical processes, without having to determine the deterministic branch ahead of time — the step-by-step process — both for us and our customers.
Philippe Ensarguet, Orange (44:56):
I want to build on this one, because outside of AI, in a very predictive world, when you're taking a process and you have multiple partners who are part of implementing it — even in a very predictive way — honestly, today it is still a mess. It's a question of who is responsible and accountable for what. And most of the time, it's hell doing validation and qualification, and it's totally misaligned with the speed at which we need to move. That is what's in our legacy. If we turn this into agentic AI, honestly, I'm very, very curious to see how we will do cross-vendor agentic AI integration if we don't change the mindset and the way we implement guardrails and policy. So I think it's very much linked back to what we mentioned about the reality of the operating and operational model.
(45:58):
I think we've only scratched the surface of this. Honestly, it already takes a huge amount of time today to do validation of network functions and the processes we need to automate. I'm very curious to see what it will be in five years from now, when we turn more agentic, and who will accept to take the risk of something that is not predictive.
Ray Le Maistre, TelecomTV (46:18):
Okay. Now we have 90 seconds left, basically. So I just want to very quickly ask Ahmed and yourself, Philippe — can you tell us about any key AI initiatives that are in play right now at your companies that you believe could transform your network operations in the next six to 12 months?
Ahmed Hafez, Deutsche Telekom (46:45):
We have an initiative we call MINDER — Multi-Agentic Intelligent Network Detection and Remediation. The idea is a multi-agentic layer that would actually take the network to a completely different level of autonomy using agentic AI. This is a very big initiative and it's under development — we are working very actively on it to bring it to life. It has a lot of challenges, but we're learning and we're developing. If this succeeds and we get to level four-plus automation — meaning that observability and incident management are all managed by AI in certain domains, whichever domains we implement it in, whether that's transport and other areas — that's the idea.
Philippe Ensarguet, Orange (47:34):
We can't be successful in our AI transformation without being successful on automation, because otherwise we won't be able to trigger back or manage the closed loop. So I strongly agree with your point, Ahmed, about the relationship between the AI transformation and automation. On our side at Orange, we are driving an AI for network transformation programme that is, I would say, multi-track. We have the business layer — starting with the value. We have the technology stack. We have the training, and we have the people empowerment. Today we have between 10 to 15 different business cases that we are working on — I would say the hard way, in the trenches — with our countries and affiliates. And I think it's extremely important to spend as much time as we can in the field, in the trenches, listening to the real pain and the real need, and bringing real solutions. And perhaps that is where we should measure our first benefits and outcomes — to make the deployment and envisage the scale.
Ray Le Maistre, TelecomTV (48:33):
Okay, excellent. Well — the fleeting nature of time, as mentioned earlier — how quickly did that go? That was amazing. So let's give a round of applause to our first panel. Thank you so much, gentlemen.
Please note that video transcripts are provided for reference only – content may vary from the published video or contain inaccuracies.
Panel Discussion
Industry leaders from Orange, Deutsche Telekom, Amdocs and Tech Mahindra examine the critical barriers preventing operators from scaling AI deployments beyond pilot projects. They discuss the organisational and structural challenges that outweigh technical obstacles, and share insights on data architecture, talent gaps and the transition from deterministic to probabilistic automation. The session addresses the industry’s struggle with what one survey respondent termed ‘proof-of-concept purgatory’ and explores the requirements for enterprise-wide AI transformation.
Broadcast live 19 May 2026
Participants
CO-HOST
Philippe Ensarguet
VP Cloud & Software Engineering, Orange
Ahmed Hafez
SVP Network Strategy and Data & AI in Networks, Deutsche Telekom
Amol Phadke
Chief Transformation Officer, Tech Mahindra
Ohad Barnoy
Head of GenAI and Data Solutions, Services and Consulting, Amdocs