Data translation a massive hurdle for AI-native telcos, says DT’s Hafez

• DT’s Hafez addressed a major data translation issue during his opening keynote address at the AI-Native Telco Forum 
• Telcos have vast volumes of different types of data to work with in order to make sense of their network operations and businesses, he noted
• This data can hardly be understood by humans, let alone AI, because there is no common language or context 
• The DT exec said the telco is not going to work with suppliers that can’t deliver clear, comprehensible, verifiable data

DÜSSELDORF – The AI-Native Telco Forum (ANTF) – Deutsche Telekom (DT)’s senior VP of network strategy and data & AI in networks, Ahmed Hafez, opened this year’s ANTF with a stark reminder of the data management challenge faced by network operators as they seek to use AI at scale to improve the way they run their networks and businesses. 

The vast volumes of data that telcos have at their disposal, and which acts as the foundation for AI-native processes, “is not ready. It’s not understandable by machines. The data is hardly ready for us as humans to consume,” noted Hafez. 

That’s a big issue and one that the giant telco is trying to address as it seeks to take further steps with its use of agentic AI tools, an area where it made impressive strides during the past two years with its automated network management tools, MINDR (Multi-Agentic Intelligent Network Diagnostics & Remediation) and RAN Guardian.

Hafez noted that, when planning how to make use of AI, telcos need to consider their entire estates and not take the approach of “a use case here, a use case there… [you cannot] scale with this approach. It’s impossible. You need to scale with an approach which is much bigger, strategic.” 

And telco estates are huge, Hafez reminded the audience and those watching the sessions via the free-to-access live stream broadcast. 

“The network’s dimensions are massive – vast dimensions… however you splice it, it's more than six domains. Then you have a huge amount of functions, at least 100 unique functions [and] every function generates at least five to 10 different types of data. That is a huge amount – between 500 and 1,000 different types of data. And then you have the dimension of suppliers, because we don’t have one supplier – even in a single domain, we have multiple suppliers,” noted Hafez, who pointed out that in DT’s core platform there are systems from about 10 to 15 different suppliers and that when a service is enabled, “you stitch a lot of functions together. That becomes a service chain, [which] can be up to 20 [or] 30 different functions that have interdependencies, and you have to understand this in order to do anything intelligent.” 

The DT man also explained there is the added challenge of having to deal with multiple generations of mobile technology. “The lucky ones have 4G and 5G only. The unlucky ones, many of them in Europe, have 2G as well, and some have even 3G – and 6G is coming.” And for major telcos, such as DT, there is also the issue of trying to harmonise operations across multiple countries or regions. 

“This is not an excuse not to do things but just to understand that we need to focus… we have to be very clear what we want to do, when and where,” added Hafez.  

According to the DT executive, the vendor community isn’t making it easy for the telcos to focus. 

As a result of the vast dimensions of the network, there are hundreds of different types of data that need to be considered and it comes in different formats that make it unfit for optimum processing and introducing AI-native telco processes at scale. 

He told the vendor community, “you’re making it so difficult for us to consume the data,” and that’s “assuming the data is of high quality. It’s difficult to consume because the data doesn’t have a clear meaning. Even sometimes the names are misleading, and then there is no metadata that describes this data. The data is not ready for agentic AI or for AI to consume. It’s only ready for domain experts who have been 10 years in the field,” noted Hafez, adding that DT has to also spend a lot of time and effort cleaning up data to get it to the required level of quality. 

But that isn’t something that telcos should be having to do, he stated. Ensuring that data is clean and of high quality is “not hard in this era” and there are lots of mistakes being made that are “really avoidable”. 

As a result, “going forward, we are not accepting [poor quality data] from suppliers anymore. Data has to be clear, metadata has to be clear, and we will [verify] the data at the source because this is the cheapest way to do it. So don’t get surprised if you see these demands coming up in RFQs and so on,” noted the DT exec.  

But that’s not the end of the data management challenges.

“Everybody knows the problems about data… but we often forget the semantic layer… the context you give to the AI agents in order to increase their accuracy. If you don’t do that, agents take the data and they don’t understand it” because AI agents don’t know about the relationship between the radio access network and the core, what comprises a mobile service or the dependencies between different elements and domains. 

“If something fails in a load balancer on a lower layer of the stack, it can impact six services. This context needs to be developed and put to the agents to consume [so they can] understand what’s going on.”

And these are just the technical challenges – the ones that can be solved, stated Hafez. The non-technical hurdles take longer, and are harder, to address because the people factor is still important in AI-native telco processes and the main problem is there’s a lack of knowledge between network operations and data scientists about their respective areas of expertise, which makes it hard for them to work together constructively, especially as they are also used to working in completely different ways, with different organisational structures and using different terminology. 

Hafez went on to discuss the challenges associated with running on-premises AI infrastructure – some related to GPUs, some to the datacentres themselves, some to energy – and explained why he is no longer interested in engaging with suppliers for standalone proofs of concept (PoCs) – any PoC now has to be a stepping stone towards the development of a minimum viable product (MVP), noted the DT executive. 

There was much more to the opening session of the event, including a brief discussion with speakers from Rakuten Mobile, Totogi and Wind River about some of the challenges raised by Hafez – you’ll be able to watch an on-demand recording of the session in the coming days on TelecomTV. 

In the meantime, the AI-Native Telco Forum continues until the end of the day on 9 September, so check out the agenda and tune in via the live stream! 

- Ray Le Maistre, Editorial Director, TelecomTV