Transforming to a digital telco

We serve about 24 million customers across six markets, and understanding their pain points and what they want is what keeps us up at night. So we create insight along the journey, ask the customer directly, watch the relational NPS, and follow it up to the brand perception of the strongest brand in the region. Underneath that sits more than a petabyte of storage, terabytes of new data every day, 300 data sources and 6 billion data lake transactions a day. It is a massive amount of data in breadth, depth and velocity, and it has to be abstracted upward before it becomes insight and action.

Customer value management is the first action. We built it on AWS serverless with more than 300 attributes in the models, working out the next best action and the next best offer. Two things follow. Propose a relevant product at the right time in the right channel, and conversion goes up significantly. Give the customer useful information instead, such as why the home Wi-Fi underperforms and that a mesh would fix it, and satisfaction goes up too. We reuse the models across markets, so quality compounds rather than being rebuilt.

Advertising is our most technically demanding case. We run the TV4 business with a large advertising component, and we built what I believe is world-leading technology for real-time individual ad insertion into a live stream, with ad auctioning happening in the moment as well. It runs serverless on AWS, and we are probably one of the world's largest Lambda users when the capacity spins up at the break.

The networks serve those same 24 million customers, and we spend about 15 billion SEK every year on network and IT. The question is how to spend it optimally, so the decision model takes usage, geography, demography and build cost and tells us where capacity should go. Data traffic grows 30 to 35 percent every year, which is tenfold over a decade, and 5G is delivering 500 Mbps average downlink in a city like Helsinki with 70 percent population coverage in Finland.

On cyber security we protect about 100,000 endpoints, every one of them an attack surface, against a constant volume of attacks. The analytics job is identification and correlation: what is relevant, what is behind it, what the root cause is. Zero-day attacks have no signature to match, so we look at the behaviour of the systems and detect the anomaly instead, and machine learning lets us correlate patterns including events from outside.

External monetisation is Crowd Insights, built on a significant volume of anonymised, aggregated, fully GDPR-compliant data points, analysing around 50 million trips a day. During the pandemic we were asked whether people were adhering to travel guidelines, and we built a model with the health authorities to forecast the patient intensity arriving at hospitals. Public transport authorities use it to optimise networks against real travel, and brick-and-mortar retailers use it to decide where to open a store and what to offer when.

AI ops runs across the network, the IT and the customer touchpoints, processing millions of events per second. From that flow we extract about 5,000 events a day worth looking at, find the patterns, anomalies and correlations, tie them back to customer tickets, classify them, go to root cause and remove it. Thresholds tune themselves, because too low and you get an alarm storm, too high and you miss the real one. And predicting and preventing outages before they happen is the point of all of it.

Process mining sounds boring and is not. At the lowest level a company like ours has half a million process steps, across product, customer, technology and the enterprise functions. We partnered with Celonis, who invented the field, to apply it at large scale, and for one fibre ordering process alone we look at around 20 million process events, examining the variants, the delays and the inefficiencies. Intelligent automation then sits on top of those insights. We integrate through messaging and APIs where we can, and use bots where we must, around 500 of them, with AI applied on top for sentiment analysis and automatic classification of calls into the care centre. That saves hundreds of thousands of hours a year, and the bots get smarter as we go.

Underneath everything is a cloud transformation from bare metal to virtualised to containerised to serverless, on a fit-for-purpose cloud. BSS, OSS and network workloads run on premise in our own data centres, while a significant share runs public, and we recently completed a massive SAP migration from on-premise legacy into AWS in less than a year, moving 30 billion records, with a tenfold speed improvement from in-memory processing.