- Orange Business shifted sensitive AI work to open-weight models it runs itself, with sovereignty outweighing cost as the main driver
- AI is transforming customer support, with an agent cutting incident triage from as long as 40 minutes to about three minutes, forcing a rethink of what level one support does
- Orange Business's Live Intelligence platform, built for 100,000 Orange employees, now sells to 150 enterprise customers as a multimodel platform
Orange Business is moving its most sensitive AI workloads onto open-weight models running on infrastructure it controls, driven primarily by sovereignty concerns.
"The importance of being able to do AI in a trusted environment has increased a lot," said Miguel Alvarez, Orange Business chief data and AI officer, in an interview. Agentic systems, whose autonomy introduces risks requiring greater scrutiny, is a significant reason for the increased focus on trust.
Orange Business is using open-weight models on-prem for applications where the company does not want source code or data to go to the public cloud, Alvarez said. These include agentic systems, whose autonomy introduces risks that require greater scrutiny than previous AI implementations.
An open-weight model is one whose trained parameters are published for download, letting any organization run it on servers it owns without routing data to a vendor's cloud.
Open-weight models are increasingly suited to advanced AI work. Work that required frontier models four to six months ago now runs acceptably on the cheapest tiers from OpenAI and Anthropic, or on open-weight models, Alvarez said.
Some of the open-weight models Orange Business uses for coding are Chinese. Alvarez named Alibaba's Qwen and the models from MiniMax, and said his teams are testing Moonshot AI's Kimi. The latest batch of Chinese open-weight models is good at coding, and those are the ones Orange Business is using more, he said. (However, Chinese models are controversial.)
Security is now one of the first conversations Orange Business has with any customer, Alvarez said, as AI's own failures have become high-profile. Hugging Face disclosed on July 16 that an autonomous agent system had broken into part of its production infrastructure, and five days later OpenAI identified the attacker as its own models. Researchers then found that a swarm of OpenAI agents had used a dormant German wiki as a coordination board. And on Sept. 9, Anthropic published an alignment assessment disclosing a fourth incident in which one of its models reached real third-party systems during a cybersecurity evaluation.
Trusting vendors to watch on your behalf is no longer sufficient to assure security, Alvarez said. Vulnerability management now involves more thorough, more frequent assessments of known vulnerabilities in the software the telco runs. Proactive scanning for suspicious activity and the ability to quarantine sensitive services and cut their connectivity will become increasingly important, he said, adding that the industry is still working out what that entails.
Orange Business is among the telcos furthest along in putting AI into production, both for itself and as a service for enterprise customers. The shift toward sovereignty, open-weight models and security is part of continuous change in Orange Business's AI strategy.
"On AI, if your expectations are not being reset at least twice a month, you are maybe not paying attention to everything that's going on," Alvarez said.
Orange Business has seen a shift in what it believes AI can do. "At the beginning we had the impression that AI was going to allow you to do things as well as an expert, but without necessarily being an expert, and I think that for many tasks that is less true," Alvarez said.
Building beyond chatbots
In another example of Orange Business's changing priorities, the company is moving beyond building chatbots. Two years ago, AI was delivered as a chat window running beside the applications people actually worked in. Now, AI is being pushed into those applications directly, behind buttons, with agents fitting inside existing systems of record and governance.
Orange Business's own sales tool began as a chat interface for asking questions about an account, gaining the ability to edit documents, synthesize real-time customer news and connect to the CRM, and then a library of saved prompts. But users wanted more. They didn't want to switch context to a dedicated chatbot and type actions manually. They wanted more autonomy from the tool. In the new generation of AI, the user states a goal and the agent supplies the steps.
AI is also changing incident management. Early work there was summarization, ticket correlation and root cause analysis. "You still have the same roles doing more or less the same work in more or less the same way," Alvarez said.
As a next step, Orange Business gave a mixed team of AI and operations process experts a three-month mandate to attack the front third of the incident management chain: qualification, triage and routing. That segment ran 30 to 40 minutes per ticket, with a technician working down a checklist of connectivity and configuration tests. An agent can do the same work in about three minutes, with 95% of the work automated, Alvarez said.
The new way of working may have staffing consequences; the Level 1 support team may not survive in its current form, Alvarez said. Instead, these teams will manage triage and customer communication while supervising agents.
Similar exercises are running at Orange Group level across HR, finance, legal, software development and B2B sales, aimed at whole job lines rather than individual use cases, Alvarez said.
Selling trust rather than scale
Orange Business offers its internal AI platform as a service for enterprise customers. Live Intelligence has drawn 100,000 Orange Group users, roughly 60,000 of them monthly, and Orange Business has sold the service to 150 enterprise customers.
The architecture runs on Cloud Avenue, Orange's private cloud, with Live Intelligence Studio for developers, a portal for end users and a management console for administrators.
Using Live Intelligence, customers pick their models. Options include models served from Orange's sovereign platform in France, which Alvarez said is very important to the company's French customers. The stack is built on open standards, including LangChain and LangGraph, OpenTelemetry for observability, Model Context Protocol for interconnection, and open-source components for the LLM gateway and MCP registry.
Sovereignty is an option rather than a mandate. The platform serves Gemini, ChatGPT and Claude alongside Mistral and the open-weight models Orange hosts itself, letting a customer route ordinary work to a hyperscaler model and keep sensitive workloads inside a secure enclosure. That optionality became a practical priority in June, when the U.S. government ordered Anthropic to suspend all access to its Fable 5 and Mythos 5 models by any foreign national.
Where the rest of the industry stands
Orange is not alone among telcos turning to open models. SK Telecom is building its own open models to support sovereign AI, putting model ownership at the center of its AI strategy. AT&T is leaning on open models from American companies to hold down token costs. T-Mobile is routing workloads to open models where they are the best fit rather than the cheapest option. Nvidia’s recent telco AI survey found 89% of operators said open-source models and software are important to their AI strategy.
As for the larger telco AI picture: Deutsche Telekom is building an Industrial AI Cloud with Nvidia under a €1 billion partnership announced in November, a sovereign German stack that went live in the first quarter of this year. SK Telecom is scaling up AI data center capacity through a new subsidiary, SK Hyper, targeting 15 GW by 2035. MTN Group has proposed "data embassies" — regional hubs where other countries' servers stay under their home country's law — as Africa's answer. Last year, Telus commercialized Fuel iX, the internal platform more than 50,000 of its employees use to reach some 40 models, on a path similar to Orange's.
Read more about artificial intelligence
Orange creates internal AI cloud for employees, sells to businesses, too
U.S. Anthropic order throws a megaton of fuel on sovereign AI fire
Orange Group taps insider Usman Javaid as chief AI officer
What is sovereign AI? Only 13% of enterprises know.
Data embassies are MTN's answer to the digital sovereignty crisis