Speaking at FutureNet World back in April, Miguel Alvarez, AI chief at Orange Business, discussed his firm’s dual-mode AI platform strategy in the context of its own transformation, its sovereignty pitch to enterprises, and the broader sector’s efforts to establish tech-co influence in the AI ecosystem.
In sum – what to know:
Platform – Orange Business is building Live Intelligence as a layered AI platform combining open-source tooling, multiple frontier models, and sovereign cloud infrastructure across its telecom footprint.
Strategy – The firm is prioritizing control, compliance and reliability, keeping humans in the loop while pushing toward “carrier-grade” autonomy and enterprise-scale agent deployment.
Position – It argues that telcos are uniquely positioned to orchestrate AI safely at scale, even as inference shifts toward distributed edge infrastructure over the longer term.
We ran a version of this story a couple of weeks ago, essentially cribbed from recent news items; but here’s the real commentary, plus some philosophical meanderings about the state of telecoms, including how the unruly promise of AI has made a virtue of its protagonists’ regulatory discipline and parochial scale. “We have some muscle there, useful all across the world,” says Miguel Alvarez, chief data and AI officer at Orange Business, speaking at FutureNet World in London back in April (see coverage here also). The premise is the firm’s own “duality”, selling digital transformation to enterprises based on its own road-tested AI change-programme – as pursued across the whole Orange group.
“It’s funny how the problems you think are operator or telco problems are the same for everyone,” he says.
The medicine here, which Orange is sucking down for itself and standing-up for others, is an AI platform called Live Intelligence, discussed previously. “It is a product we sell, and a capability we also use internally,” says Alvarez. It is based on the open-source LangChain and LangGraph frameworks from US AI firm LangChain, plus open source LLM gateways, tooling connectors, and retrieval engines, and presented as a vendor-neutral overlay on top of its own data lake and third-party business intelligence layers, which hooks into sundry frontier models from OpenAI, Anthropic, Google, and Mistral, with a “toggle button” to switch between, and control access and capability.
The platform has built-in retrieval mechanisms to connect AI models to enterprise knowledge sources. While ostensibly independent, it is hosted in Orange’s sovereign private cloud, Cloud Avenue, in Grenoble in France, and supports on-prem open models, and access to hyperscaler infrastructure – so customers can switch between latest AI models without being locked into a single provider. “We partner with everyone,” he says. “And it is available to everyone.” Which means basic licences are affordable, to democratize and standardize usage across enterprises and reduce shadow AI; AI performance can be turned up and down within the portal for power users.
In March, Orange Business launched a new agentic capability in the platform, called Live Intelligence Studio, to go beyond generative AI bots and “simple agent creation”. Orange is embedding Live Intelligence into its own products, such as its network-as-a-service Evolution Platform – where agents ‘get’ the context and requirements, and propose configurations and other support; same with its Intelligent Together solution, with agents on hand to help develop “voice bots” for service departments. Internally, it is deploying agents to automate incident qualification in complex SD-WAN environments, reducing manual investigations that might take an expert 30 minutes to just a few minutes.
Alvarez says: “We’re different because this platform is used internally across the group and externally by customers. We have over 100,000 internal users; we have over 50,000 active [external] users. We have agents running tens of thousands of times a month.” The external figure is spread across at least 100 private and public sector customers. Last November, the firm introduced a PowerPoint agent just to create slide shows; it created 20,000 decks in its first two months, he says. “It was a huge take-up – because it’s available to everyone. We know what works and what doesn’t. Our product iteration cycle works very well.”
Which sounds like Orange is all-guns-blazing with AI; it is and it isn’t – and that’s the point.
Riding the clutch
Orange Business is taking a cautious approach to autonomy, including with these agentic solutions, keeping humans in the loop for configuration changes until AI can meet carrier-grade reliability requirements. “Our network runs at five-nines (99.999 percent of the time). Always. AI cannot be allowed to impact that. The way we manage that today is with humans in the loop – the AI does some heavy lifting to give an intermediate result, and humans do the rest. The discussion here (at FutureNet World in London in April; also at DTW in Copenhagen this week, and at every telco show) – is how to achieve Level 4 autonomy. For that, the AI has to operate with telco-grade reliability.”
Equally, it has to do so “in a cost effective way”, says Alvarez. Hence the platform strategy with Live Intelligence. “Otherwise, your system is unreliable, or else over-engineered and expensive,” he reasons. He references a meeting with an unnamed French industrial outfit, which “made an early bet on a cloud-based tool” and saw good results with “basic office work”, but has since come unstuck with “more structural use cases”. Its solution “doesn’t scale downwards”, he explains; the firm came knocking to ask about “good platform strategy”. Orange Business has avoided the trap – too simple (won’t scale) or too complex (can’t sustain) – by creating a layered and open system.
Which means an easy interface on top, and an any-cloud (including sovereign-ready) infrastructure on the bottom, and all the open connectors, retrieval engines, and agentic tooling to run workloads in between – as required, according to private performance, compliance, and security preferences, plus any commercial biases in play. “What I like about our platform, and I’m not saying it’s the only one that works, is that the same top layer that serves end users, which is very simple, can also be fed by the more sophisticated and powerful agentic layer.” So the interface is not limited by its simplicity, in other words; complexity and intelligence are abstracted, but easily available.
Its human-in-the-loop five-nines caution is not just about higher standards – for customers with the same. It is also because there is plenty of work to do, still, behind the scenes. Enterprises, like telcos generally, are not ready for AI. “I can tell you, I am in charge of a few hundred people, and, even though 80 percent of conversations are with people that want the latest AI, more than half my team and more than half my time is spent working on data – data quality, data preparation, data governance, data security, data compliance. That’s the baseline. If the data is not right, the AI won’t work. Because your data governs how these agentic systems interact with your business reality.”
Alvarez adds: “The limiting factor today is not the capability of the models. Most companies are not coming even close to what these models can do. Because nothing else is as advanced as them. Which is good, because it makes everyone aware of the potential. But even the best model in the world needs the right data, in the right way – so it doesn’t introduce security risks into your organization. You need to know it’s compliant – with your data policies, security policies, sovereignty policies. You need to be sure you can run it in an effective and reliable way. Because nine times out of 10 is not good enough; it’s not good enough for a telco. Ninety-nine out of 100 is not good enough.”
Hitting the gas
Some practical throttling of runaway corporate AI is a good thing, he suggests. “The power of these agentic models is huge. There is pressure in the market to go very fast, which introduces systemic risk. I tell you, I see the difference between what we are testing, very early, in a very limited way, and what we’re rolling out at scale. Those are not the same things, and we are very serious about the transition between the two. But there is no guarantee everyone will make it. There is an implicit risk, and people know it – but the incentive to move fast is very strong.” Coding is the clearest example, he suggests: 15 months ago, he asked the team who was using third-party AI for in-house coding.
Hands went up, by all accounts. “It is something we have worked very hard to stop,” he says. “It presents a security risk, where intellectual property gets leaked out, and so on.” The entire team at Orange Business – not just Alvarez’s cohort of a couple hundred, but all 4,000 “experts on data and AI” in Europe – have been instructed to use the upgraded Live Intelligence platform for coding work. “Because the risk, even when they think that they have created enough of an air gap, is just too big. So this pressure to adopt AI is huge, even in industries that are conservative.” The conversation skirts around the perceived AI backlash from software engineers – very visible on LinkedIn, say.
But does it not take as long to unpick the AI errors? There is lots of talk about AI coding slop, from people that appear to know. Alvarez’s response seems like a good one: “The biggest software developers in the world have adopted AI at massive scale with very few regrets. This is psychoanalyzing the market, but two things happen. First, and yes, these tools will produce a lot of slop in the wrong hands – more code, but less value. People who were bad at coding, or just lazy, whose productivity was close to zero, will suddenly produce lots of code, but the code is shit, and needs to be unpicked. So you need to eat your veg; you need good coding practices and good quality control.
“Same with any profession. Good developers have a good quality bar; it is just that they can now iterate almost as quickly as they think – rather than only as fast as they type. Which is a big gain. Enterprises need to build [control] processes in; it is not just about code completion. Second: change management matters. A lot are comfortable with coding, but not very comfortable with AI; and maybe they’ve had disappointing experiences – like it was supposed to be like magic. So there is a backlash around that, too. Developers that are good at this are good at both; developers in Orange are good at both, and have seen productivity not just increase by 10 or 20 percent, but by a factor of X.”
But there is a sense to release the gas, as well. Alvarez has been with Orange for 17 years, involved variously with solutions, pricing, transformation, partners – per his resume. He was put in charge of ‘data and AI’ in April last year. His first three months were an “awakening”, he says. “There are so many things we want to do, but we have to be very thoughtful about how we do them. It was frustrating in the early days. The tech can do so much, but you have to think about how to go beyond just demos – to create something you work with every day.” Reality bites, as ever; AI tools often fall apart in real work environments, full of proprietary systems, and security and regulatory hang-ups.
Fuel in the tank
Which is what telcos already know, of course – or so their “carrier-grade” sovereignty pitch goes. Alvarez comments: “We have had to deal with an environment in our home market in France, which is very heavy on sovereignty, trust, regulation. And we have developed some muscle there, which is now very useful all across the world – not just in France and Europe. We are not perfect by any means. But our value proposition… is fairly unique.” For value proposition, read both its platform play with Live Intelligence, and its inherent trust credentials as a telco. Is Orange better at this than the other telcos in this room – in London, at FutureNet World? “Hard to say,” he responds.
But if you had to characterize Orange Business – versus all the rest – are you doing the same thing, broadly, or is there something different in the old OBS DNA? Because RCR has been writing about your enterprise integration and cyberdefense activities for years, and it feels like all the new talk about critical infrastructure and digital sovereignty plays into your hands – and into the whole industry’s to an extent. He says: “Well, first, I want everyone in this room to win every time – except when they are against us. But no one here wants to be another hyperscaler; this industry has outgrown all of that.” The opportunity, just the concept, has gone anyway; telcos have been eclipsed for scale.
He says as much: “When I joined Orange, Google was the smaller company. Today, it is not a smaller company; it is not even roughly the same size.” Later, he says the same about frontier AI builders – just in case it needs saying. “Our job, our calling if you will, is not to be the frontier lab – to bring the revolutionary breakthrough. Others will do that.” There is some kind of awe-struck thanks in there, almost. “We are lucky as telcos because we get to work closely with the biggest companies in the world.” But there is steelier opportunism, too, borne from the sector’s own structure and heritage – which are now being rendered as virtues amid the unruly prospectors in AI boomtown.
He explains, also getting to the original question about Orange’s position among its peers: “We know there is lots to do with AI, and that it is a turbulent environment. People are betting billions today, and not waiting for tomorrow. I mean, you asked earlier what keeps me up at night, and you think what happens if inference becomes 10 or 100 times more expensive… But we need to follow our path – everyone here thinks that – which is not to say everyone’s path is the same. We have more stakes in [play] – in data, AI, cloud, cybersecurity – because of the acquisitions we’ve made. So it is a bigger part of what we do. Not versus everyone, necessarily, but compared to many others.”
So are telcos happier in their own skin? “It is a business that always feels like it should be doing better, and is always doing a little better than it thinks it is… But two things have changed, which we’re happy about. One: good plumbing is [perceived again as] super important. These models are fantastic when they work; like magic. But the plumbing is hard work. And telcos are good at it. It’s not sexy, but we do it well. Two: trust, sovereignty, choice have been important for telcos for years, and are now important to everyone. These trends put us in a central place in the AI ecosystem. Maybe we won’t make the same margins as some others, but we can offer what enterprises want.”
Reading the map
So does this new telco role, ostensibly about responsible orchestration and stewardship of AI in the mass market, become deeper in the ecosystem as inference workloads are increasingly attached to edge nodes, and scattered within national and metro (and even campus) networks? “Well, transit time is not the biggest factor right now to avoid undue delays to manage real time flows. Compute time and inference time are an order of magnitude bigger. It could be that, as specialized mid-sized open-weight models come available, as we are starting to see, which can do inference at a reasonable rate on limited infrastructure, that you’re going to see AI migrate towards the edge.”
He goes on: “Like everything, rare commodities tend to be centralized, and become more diffuse as they are more available. That’s going to happen with inference as well. But I don’t think that’s for today or tomorrow morning; it’s going to happen in the mid-term. And telcos will have a very interesting role to play – because we have very granular access to and distribution of compute, at country and metro level. Orange operates in more than 20 countries, and we have already built-in compute and connectivity in each of them. Which gives us a very good starting point. But like I say, that’s not going to make our business plan in 2026 or even 2027.”
Meanwhile, the framing is about sovereignty and trust, and the strategy is to shepherd enterprises along a jet-heeled AI transformation path – without being rear-ended by risk or outrun by ambition, or otherwise run into the ditch of industrial revolution. For Orange, the vehicle is Live Intelligence – always road-tested for steering and velocity. Alvarez closes: “Telcos have brought new tech to the mass market in a safe way for 50 years – the telegraph, fixed telephony, mobile telephony, the internet; accessible to everyone… We are on the right path, roughly. We are here today and we will be here in 10 years. And for companies that don’t want to be exposed [to risk], that is reassuring.”
