AI Journey at Ericsson started over 15 years ago when a group of enthusiasts at Ericsson Research started to look into big data analytics methods in attempt to simplify the lives of our colleagues who worked in operations and sat on numerous trouble ticket and alarm data around the world. At the same time, lots of knowledge and information about the telecom domain was quite structured in manually-readable text such as product documentation. Nowadays we know that text written in a digital form can easily be transferred to a machine-readable format, so we made use of all this beautiful documentation and created the first version of our telecom knowledge-base that was successfully used in trouble ticket resolutions. Interestingly, back then, the sizing of our first and second line support teams were several thousands of people because Ericsson’s Managed Services that even today serve over one billion of subscribers used to be a highly labour-intensive business. Things have changed since then. Many manual tasks have been automated, and even privacy aspects became improved when switching from manual handling of operations belonging to different customers, where knowledge sharing across the competition boarders was not allowed. 

After solving the most obvious parts urgently requiring automation, and what i used to call “low-hanging fruits”, we switched to optimisations that were closer to the heart of telecom system, namely radio access network and core. The specifics of usage of AI in these parts of telecom system include real-time aspects. Here we’re talking milliseconds. In order to be able to proactively adress any unwanted situation in the network (such as service degradation, which can lead to fatalities if safety-critical enterprises are relying on it), one needs to receive the timely input, meaning that the data sampling needs to be of a high frequency, and to be able to process this input in “real-time”. In this context by “real-time” we actually mean “as fast as possible” unlikely the formal definition from real-time systems research when the result is supposed to be delivered “just in time”. 

The world of telecoms has seen some fundamental changes thanks to AI. The most important paradigm shifts include the following aspects:

• Firstly, thanks to knowledge sharing techniques, we can see more reuse within telecom industry, and across adjacent industries. This change has been made possible by global knowledge sharing techniques including ontologies, semantic interoperability and large language models.

• Secondly, AI is gradually replacing dangerous tasks of our field support operators, such as tower climbing in order to perform an inspection. Telecom sites are nowadays capable of preventively notifying decision makers about any failures or predicted service degradations. This can be done through analysing the data coming from the base stations or external evaluations via drones in combination with computer vision. In many cases, this predicted service degradation can be resolved remotely, without exposing humans to dangerous tower climbs.

• Thirdly, AI has proven to be successful in control loops, including the fastest, close to real-time control loops. These control loops assume execution of an AI algorithm on a dedicated hardware, as close as possible to where the decision is taking place. These fast control loops are found in the heart of radio access network.

• In addition, the latest developments obviously include various applications of large language models, where we see clear benefits from using these models in customer support, R&D efficiency and even innovation.

All these AI developments have been very impressive so far. Let’s discuss what’s missing to be fully capable of tapping into the potential of AI in our society.

• First of all, the AI market is fragmented. Even in the scope of one company (even though we’re talking of a pretty big one), after a new technology has been detected, you can find over 20 places where it has been developed and applied. This is simply a human nature, and is much closer to behavioural science than computer science. However, all these numerous deployments will still facilitate the overall progress, and eventually we will converge on one (or a couple of) winning solutions, just like the hyperscalers of the world today.
  
• Secondly, this technology is not being developed according to standards (which we, telecom people, normally love so much because in telecoms standards mean scale). In AI, however, the pace of development is much higher than in telecom industry, and forcing any AI developer to comply to standards will simply hinder innovation. 
  
• Another gap – and after this you’re free to call me Captain Obvious – is Observability. We’ve been living with this gap way too long and, to be honest, it caused me some gray hair. Here we’re facing a digital divide, but not talking about consumers but rather industries. See, a “youngster company” on the market would be a digital native, and would hopefully architect all its processes and products accordingly from scratch. This means that it would probably be able to read out any data from the products and services in the field where matters to be able to improve. When you instead look at a 100+ years old company, then the processes, mindset and the culture will need to be redesigned according to the digital native principles, putting AI in the middle and not “on top”. 5G was the system partially observable to take advantage of AI techniques. 6G-baby is being created as an AI-native baby.
  
• Further, AI trustworthiness needs to be taken seriously, and algorithms ensuring adherence to non-bias, non-discrimination and explainability principles need to be in place. 
  
• Lastly, let’s not shy away from the fact that AI is a software, and without a proper actuation its precise outputs are useless. Precise actuation is just as important for the whole system as precise observability.
  
  Take-aways:

1. AI is not an icing on top of the cake. It has to be embedded in a product or a service to offer the full advantage of its capabilities to either take proactive measures or be able to react and compensate the network’s imperfections in rea-time.

2. AI in telecoms needs to be able to observe what matters. We have tons of data, and most of it is not interesting. It’s when something is about to happen, we need to be able to observe it with high frequency.

3. We need to give AI time to prove itself. In industries that have existed for over 100 years, there are some fine-tuned algorithms that are difficult to challenge. Even though the challenge is important and necessary. Still, replacing a conventional rule-based system with an AI-bases algorithm may lead to performance degradation at first. It’s like bringing a talented trainee to a workplace full of experts that have been there for a while. There is a slim change for that trainee to make an impact unless we give her time to prove herself, since AI algorithms are adaptive by nature. 

4. AI-to-AI communication is important. Every industrial domain has its own enterprise data, and its own AI crunching this data. When we connect or stack different domains, such as, for example, intelligent transport systems domain relying on connectivity domain relying on compute infrastructure domain, we will face different AI algorithms operating within these three domains. Each of them will have its own strategy, so we better ensure collaboration between them. 

5. Trustworthiness of AI. Trustworthiness of AI is something that needs to be there from the start so that we ensure that the algoritms, even following their best intentions, won’t lead us into unwanted situations.