Orange: Three unexpected lessons about AI in business

Sometimes it seems like there are two forms of Artificial Intelligence. There is the AI you see in movies, which is hyper intelligent, all-powerful, and probably out to take your job. We could call it sci-fi AI. Then there is the AI that companies like Orange are putting into our day-to-day businesses, often with impressive results in customer relations, business operations, and financial outcomes.

For the past year Orange has worked with Google Cloud to deploy our data in new and effective ways, using real AI to solve concrete business challenges. I’d like to share three lessons we’ve learned, in the hope that they help you work with AI quicker and more effectively.

Focus on two experiences: Arriving and expanding

Our team, which is responsible for Orange’s AI use cases from ideation until their effective deployment by AI factories, had a number of ambitions. We wanted to deliver to Orange France more revenue, cost savings in both operations and capital expenditure, and improvements in our customer experience, as defined by Net Promoter Score. That is a lot of different areas, and a lot of data from many different parts of the business. And all this involves data analytics or AI.
We had that, in the form of more than 60 data lakes and data warehouses, spread across a number of on-prem sites. By far the largest of these was called HubData, a shared Hadoop infrastructure that was used for many types of standard Business Intelligence analytics. As we will see, BI functions differently from AI, but the raw materials of data are very similar. This was the initial data we wanted to have in the cloud, stored in a more flexible state so it could be used in a lot of different AI use cases.

We examined several cloud providers, and chose Google Cloud both for the ease of moving data over, and the best in class tools we’d be able to use once we were in the cloud. In particular, Google Cloud’s BigQuery had the capacity, ease of use, and flexibility we needed to start our work in AI with minimal disruptions. Several ready-to-use ML algorithms are directly available from within BigQuery and once trained, they can be instantly versioned and deployed in production through Vertex AI, in a way that can be customized and offers optimized scaling, lower latency, and higher quality of service. 

Your choice of cloud provider may be different, but focus on both the ease of arriving at the right place to begin work, and the tools you’ll work with as you expand your AI practice. The sooner you are in the cloud, with the least hassle, the more you can focus on adding value.

When it comes to AI projects, bigger may not be better 

Because it is a powerful tool, we initially thought AI would be useful to Orange for a handful of big blockbuster projects. In fact, in the first year we worked both on big bets and on 30 to 40 smaller projects, all of them successful, and we have many more awaiting production. We’ve been able to do that with a very limited growth of our teams, which is a credit both to our people and the tools we are using. 

Once we struck the right balance between real business needs and challenging, but achievable, technology capabilities, project teams could start to imagine lots of areas of potential value. And, since we’re not aiming for moonshots, frequently both staff and customers don’t have to go through big changes in how they operate, they just see improvements.

These projects have been across a number of areas. In one case, AI is being used to personalize product recommendations for our customers through multiple channels: digital, direct marketing, or it could be the recommendation that a person in an Orange retail outlet makes to a customer. This might be as radical a change as recommending improved connectivity instead of, say, subscribing to a new sports channel, but a successful sale can convince the sales associate of the value of the AI. In another example, a field technician may take a picture of a complex wiring job on a local telecoms substation, and in a few seconds visual AI can say whether the job looks right. These kinds of simple, measurable improvements power our teams’ imaginations to find more applications.

Planning AI projects has not been like using BI software, since BI can be very complex and rules-based. AI, on the other hand, is both rigorous in its design and sometimes surprising in its outcomes. Which leads me to my final point.

There’s lots of value where machines meet humans

I mentioned two AI use cases, in retail and field operations. In both cases, AI is working with a human, in a subordinate position. Final decisions are still up to the human, whose imagination, creativity, and experience continue to hold immense value.

This is in keeping with our “responsible AI roles,” in which humans always make final decisions, with particular care in customer-facing activities like fraud detection or product experience. It also supports the idea that Intelligence Augmentation of humans, or IA, is a powerful source of value (and possibly easier to find) than stand alone AI projects. 

No different from using any new tool, there is a learning curve with AI, along with a few surprises. In our case, these were mostly pleasant, like the speed we’ve been able to get up, and the number of successful projects we’ve launched. Like all new tools, as you begin to use them well you gain new ways of seeing both them and the problems they solve, and even understanding things in new ways. 

The usual term for this is “change management,” or bringing teams to appreciate new ways of working. Several factors made the process smoother, including an easy data transition to the cloud, keeping projects inside existing work, focusing on positive outcomes and an improved customer experience, and enjoying rapid feedback to our success. The more people see a place for themselves in the future, the easier it is for them to adapt.

To learn more about Cloud Workstations and try it today, visit its webpage here.