How Orange is realizing value from AI

In the six years since Steve Jarrett, Chief AI Officer, Orange, joined the company as Group Senior Vice President, Data and AI, there have been huge shifts in how telcos approach AI.

“The job is very different today than it was six years ago [when] my team was entirely research PhDs. Today there is also an enormous software engineering team,” says Jarrett.

His teams serve the group’s 26 operating CEOs and enterprise and wholesale businesses, across four domains: customer experience, smarter networks and operational efficiency and new revenue streams.

“We are more and more focused on the core business challenges and the core business opportunities, working with the CEOs to identify where there's a big value opportunity and cost savings,” explains Jarrett.

The AI teams collaborate with their peers in networks and IT to help Orange’s operating companies hit their targets across the four domains. And a company-wide dashboard displays progress.

“It’s created a nice, open marketplace for ideas, where, if I'm the CEO of [an operator in] a mid-sized country, it's very easy for me to see where the value is coming from,” according to Jarrett.

One of his KPIs is value generation. “This year [2025] we are going to [achieve] 300 million EUROS in value generation across the business across those four domains,” he says.

Jarrett stresses that his AI teams do not deliver value in isolation and that their success depends on much more than tracking operational results.

“The reason why it works is because we've got good coordination and prioritization.” For example, Jarrett works closely with Laurent Leboucher, Orange’s Group CTO, on how to use AI to improve network operations for customers and the business.

Picking partners

When it comes to technology deployment, Jarrett’s teams’ first port of call is to license software or integrate open-source software. And Orange has opted to partner on language models, including with France’s Mistral, rather than build them: Jarrett says he is careful to avoid being “stuck in a particular AI model family, because it's hard to pick winners.”

On occasion, however, market solutions are unavailable or too costly to deploy. This was the case for Live intelligence, which Jarrett’s teams developed internally to avoid the mushrooming use of shadow GenAI applications. It proved to be so popular with over 90,000 of Orange’s employees using it that it became a B2B product.

“Where the investment makes clear sense for us internally usually it also is a very, very compelling B2B offer for Orange Business,” says Jarrett.

Responsible AI

Amid change, one constant is a focus on the environmental impact of technology, stresses Jarrett, with Orange positioning itself with customers as a provider of ‘responsible AI’.

In practice, this means “using the right tool for the right problem. If you don't need to use machine learning for a problem, don't,” he says.

This helps Orange meet its sustainability targets in its home market of France, where “we're really careful about only using large language [and] complex AI agents when we have to.”

This approach also makes sound business sense in low ARPU markets.

“It works very, very well in emerging markets where we're cost constrained, and where power provision is nothing like France where there is so much green power. So, it's both out of philosophy and of necessity,” explains Jarrett.

In the Ivory Coast for example, Orange has developed a contact center chatbot that carefully calibrates its use of natural language query (NLQ) and large language models (LLMs).

For example, this could be explaining to a customer who is ready to buy a phone which device best suits their budget and requirements. In addition, the platform can link to the customer’s Orange Money account to verify funds and allow the customer to validate their purchase within a single chat.

Jarrett’s teams are also helping customers across Africa make use of services in multiple languages that are not otherwise served by AI models.

“The skills of my core language research and AI team ... fine tune models to understand these languages that other AI models don't understand.”

Keeping it simple

Jarrett’s teams may have the skills to use the latest bleeding-edge technology, but they keep AI deployments simple if it is more efficient and effective for the business.

“We are really impressed with how much value we’re generating with simple forms of AI in serving customers the right offer and being much more sophisticated about predictive models to determine which offer to give” to whom, he says.

Predictive AI tools “prompt the human customer support person with recommendations of offers we know will be compelling for the customer,” says Jarrett, adding that: “Treating each support call as a revenue generating opportunity is proving to be very, very useful.” As a result, the company is extending its predictive insight capabilities to employees in retail stores.

Other examples of where Jarret’s teams are helping to simultaneously improve customer experience, internal operational efficiency and the bottom line include using AI to improve network root-cause analysis.

“It not only drives improvement in the resiliency of the network and efficiency at the NOC [network operations center] level, but also it'll drive enormous benefits in field service, because of our ability to better understand and predict what expertise to send.”

Bringing data to the people

Orange is able to use AI widely across the business in large part because Jarrett’s teams have invested heavily in data democratization. In recent years they have put in place the processes and rules that give employees “the freedom and the ability to be agile,” explains Jarrett.

“It's a multi year journey [and] we're nowhere near done,” he adds. “But we've made huge progress. High quality data and clear rules will be crucial for any future AI agent deployment.”

Orange now has its own version of a data mesh, for example. “The central team developed tools and infrastructure, and the operating teams are responsible for managing and ensuring the quality and accessibility and freshness of their data,” Jarrett explains.

This has given business lines access to higher quality data. As a result, “a marketing person can run campaigns that relate to network quality,” he says.

Limiting agents’ agency

High quality data will be fundamental as Orange looks towards agentic AI. Currently, software developers are the early adopters of AI agents within Orange “because programmers are going to be sophisticated at knowing how to manage these systems and monitor them,” according to Jarrett.

Indeed, Jarrett uses agents to analyze his teams’ work.

“I take their source code and I have an AI model tell me what [it] is. I can appreciate what's going on there, and it's just transformative,” he says.

But as much as he personally enjoys using agents in his work, he does not pretend widespread deployment will be easy. “As those tools mature, I will expect it to extend to many other job lines, but there's lots of complexity there,” Jarrett acknowledges.

Nor does he foresee people going away altogether anytime soon.

“You want to have people in the decision-making loop. You need to be able to manage these agents at scale and monitor them, so they don't overspend, and they don't make mistakes,” Jarrett explains.

“We're at the crawling stage of managing these systems at massive scale,” he adds. “It is another multiyear transformation.”