Orange hedges its GenAI bets, avoids LLM lock-in

• AI is now influencing all of Orange’s tech and network strategic thinking
• The giant telco has a two-fold approach to its AI thinking – how to use AI and how to cope with customer use of AI
• Chief technology and information officer (CTIO) Bruno Zerbib says GenAI is delivering amazing productivity gains
• The telco is developing applications using multiple large language models (LLMs) via an in-house-developed interface to avoid LLM lock-in 
• The Orange team also has a watching brief on the Global Telco AI Alliance

LONDON – The potential application of generative AI (GenAI) has forced a major rethink of the technology and networks strategy at international telco group Orange. It is embracing the use of GenAI tools internally and working closely with multiple large language model (LLM) developers on specific applications, whilst also pondering how best to design and run its networks in order to cope with the anticipated demands that AI will place on its infrastructure and IT systems.

That was the big picture AI vision presented to journalists and analysts in London this week by Orange’s chief technology and information officer (CTIO) Bruno Zerbib and the telco’s group CTO and senior VP of Orange Innovation Networks, Laurent Leboucher. They also placed their AI update and views in the context of Orange’s broader technology and networks transformation towards an API-enabled platform “where our infrastructure looks more like a cloud…. [with] services on top of a virtualised, containerised infrastructure” coupled with a DevOps methodology.   

Zerbib noted that some parts of Orange’s technology strategy have remained consistent over the past year or so but, due mainly to the impact of AI, “a few other things have changed. We are now using artificial intelligence to make Orange a better company across the board. So we’ve been training [staff]... We have 40,000 people using our AI capabilities as tools within the company. And the way we think about AI is not to replace humans but to augment humans, to give them more capabilities and more tools, from the way they’re interacting with customers to the way we’re designing our network and network capacity, making sure we optimise the capacity of our network… be smarter about the way we deal with outages [and]  become much more predictive in the way we design our network. So that has become very efficient,” noted the CTIO. 

He said Orange is “using AI in the development of all of our tools… When projects are simple and fairly isolated [such as] one developer working on a mini app, then what you get from GenAI is incredible” in terms of productivity gains. “It’s north of 50%” in those instances. But “when the projects become more complex and it’s much more distributed and people have to work together,” then coordination, alignment and project complexity means “improvement you get from AI is more like, I would say, is between 20% to 30%,” he explained. “Eventually AI will become more powerful, and I’m sure we’ll handle more extremely distributed projects. But right now, the sweet spot for AI in terms of productivity gain, for developers, it’s more on projects where you have a bit less complexity in terms of coordination,” noted the CTIO.

These are significant advances and Zerbib is convinced of AI’s power and potential. “I’ve been a big fan of the hype curve my entire professional life – it’s broken my heart multiple times… [but] I think AI breaks the hype curve. The kind of progress we’re making right now, in AI, is exceeding expectations. Many people don’t understand what we’re talking about – they think ChatGPT is still a gadget, a gimmick,” noted Zerbib.

But the Orange team is very conscious of how it is using GenAI (including its associated power consumption) and ensuring that, as a company, it doesn’t get overly dependent on one LLM. So it has developed its own interface to multiple LLMs, including OpenAI (ChatGPT), Gemini (Google), Mistral and others “so our employees can have an understanding of the cost impact, energy impact” of the tasks they are performing “and shuffle across multiple LLM providers. “We have to be very careful – we don’t [want] a monopolistic situation emerging with LLMs, so we’re using this technology all the time.” 

And will that use of multiple LLMs extend to the use of telco-specific LLMs, like the ones being jointly developed by the five telco members of the Global Telco AI Alliance? See Telco giants form JV to push telco-specific LLM plan.

“We’ve exchanged a lot with this initiative,” noted Leboucher, but “we have not joined yet. We are experimenting with the use of generic LLMs – Gemini, OpenAI – for some use cases on the network side but we are not yet convinced that we need a specific LLM, even for network operations. It may be the case, but we are not yet convinced.” 

Orange is using a GenAI application for summarising trouble tickets, which is extremely useful as “network engineers get all the information in a way that is easily understandable in a fraction of the time it takes to read all the information” – but that tool was developed using a generic LLM, noted Leboucher. Ultimately there “might be some use cases where a telco-specific LLM could be useful, so that’s why we are watching [the alliance] very carefully, and we might join, but it’s not decided yet,” he added.

Zerbib pointed out that advances in LLMs are happening at an incredible pace and that what generic LLMs are capable of right now was not even conceivable two years ago. This is changing the perception of whether vertical industry-focused or use case-specific models are going to be needed – hence the watching brief on the Global Telco AI Alliance.  

All of this experience and research is also affecting the way that Orange thinks about the potential of process automation. Along with all other telcos, Orange has been exploring automation for years, but “the use of AI, and the use of generative AI in the last couple of years… is having a massive impact on the way we’re thinking about automation [and how to] become much more effective and further increase the quality of service,” added Zerbib, noting that the only way to make 5G-enabled network slicing useful and attractive to potential customers is to deploy “some very advanced orchestration powered by AI.”  

So there’s a lot going on inside Orange in terms of GenAI usage, interaction and R&D. That, in turn, has the telco’s tech team thinking about how it needs to engineer its networks to cope with the potential impact that the use of AI applications will have on them and how Orange needs to think about its edge computing strategy. TelecomTV will report on that topic in a subsequent article. 

- Ray Le Maistre, Editorial Director, TelecomTV