Data and analytics: eliminating silos, enabling scale

The foundation came first, and within less than a year we had a complete data pipeline and stack on AWS, with all our operating companies on it and analytics enabled from the start. It carries 2.5 petabytes in a common data lake fed by 263 source systems, running more than 5 billion data lake transactions and 3 million queries a day, and supporting over 100 analytical products across our markets. Six local roadmaps and architectures became one, and the analytics community behind it numbers around 350 people. The part people skip is that platforms are only a third of it. We launched the region's largest in-house training programme for cloud and analytics fluency, taking ten percent of a twenty-thousand-person workforce through it, and deliberately not only the technical people: the sales, service and product people too. And we ran a data democratisation initiative, so about a quarter of our people now work from live dashboards for daily decisions rather than from PowerPoint and Excel. Without that foundation, going further is very hard.

Then the basic use cases, and customer insight is the right place to start. With the data in a single lake we understand behaviour, and we put customer satisfaction insight across every channel, so an order-to-activate, a bill-to-cash or a trouble-to-resolve experience comes back in real time, closing the loop on the individual issue and feeding the outer loop that improves the products. Customer value management is the second: machine learning models deciding the next best action and the next best offer for cross-sell, up-sell and churn reduction, where we see around a third more conversion and, unsurprisingly, a significant lift in NPS, because a relevant offer makes for a more satisfied customer. On the network side we spend a great deal every year on infrastructure, so we use analytics to understand usage, cost and revenue by geography and optimise the rollout for coverage and capacity at the right time. Cyber security is a big field: every endpoint is secured, and now every signal goes into the big data platform, so we can find the relevant signature in the noise and prevent an attack rather than watch humans try to hold all of it in their heads. And intelligent automation, several hundred bots with analytics and AI increasingly infused into them, particularly for call centre efficiency.

The advanced use cases go into monetisation and deeper operations. We own a broadcaster with a substantial advertising business, so we insert individual advertising into a live stream at the moment the break starts, a workload that spikes enormously and needs elastic infrastructure and deep targeting behind it. We monetise data as crowd insights, using anonymised mobility data so enterprises can understand who passes a store and public transport can optimise routing. We run process mining with Celonis across the processes of the company, an x-ray of the variants and delays in something like a fibre order, so we can see the inefficiencies and then simplify and automate. And we run AI ops across all our networks from a single operations centre, where billions of events arrive and the job is to extract the signal that indicates a real outage early enough to prevent it. Most recently we correlated customer complaints on one side with system signals on the other, so we can tell an affected customer which product is affected and when we expect the fix, and close the loop when it lands.

The sequence is the lesson. Build the foundation of people, platform and process. Start with the easy use cases, because customer value management pays back immediately and buys credibility that it actually works. Only then experiment with the advanced ones, where the velocity of data is high and the analytics get harder. And get everyone onto live data instead of PowerPoint and Excel.

On hybrid, cloud native matters more than public cloud. What we want is applications in containers that we can shift to whatever environment is optimal, with management across a hybrid estate, so we are never stuck in one. Regulation decides some of it: the further south you go in the stack, the more stays on premise, so the core network runs in our own data centres while analytics and application workloads run in public cloud, and traffic data never leaves. Cost decides the rest, because a contiguous workload is cheap to run ourselves while a fast CI/CD pipeline is worth an AWS.

Asked who should build the use cases, I do not believe in the central team model. We define roles: editors, publishers and consumers, graded by how much data work someone does, and the more you do the more you are trained for. The deep modelling, the training and the inference for something like customer value management, sits with a central data science and engineering team. Everything else goes to the edge of the organisation, right out to the end user, because no company becomes data driven by having one analytics department and hoping.