What about Autonomous Networks?
stein@opensky
I have not posted anything for a while now (have been busy), but recently the world has become very concerned about whether the most advanced AI models might eventually lead to the end of humanity – a concern that really took off after the “Mythos incident” earlier this year, where AI agents without authorization hacked into several external organizations on their own.
This is of course concerning – but I have no hard competence to comment on this – so I will not get into that topic – except for re-iterating that the battle for AI leadership is a geopolitical and industrial battle for dominance between the USA, China and others (refer e.g. my earlier commentaries on “Digital lead and geopolitics” and “Taking the digital lead” etc). The matter has reached a lot of attention – and even the CEOs of the most prominent American AI foundation models have all indicated that the AI development needs to slow down. We may wonder what their objectives really are, if it is for the good of the world or simply to hinder more stringent regulations etc – but Donald Trump’s ambition is clearly to win the arms race on AI.
This geopolitical battle also has implications for national and industrial “sovereignty” – as the USA and China are dominating and battling for leadership in the area (refer also my commentary on “Sovereignty - what about it?”).
On a more specific topic, however, I also see many of my old telecom colleagues working on an ambition for autonomous networks – something very much in focus during my attendance at TM Forum’s DTW Ignite in Copenhagen in June earlier this year. I will look a little bit more into this more concrete topic.
The data challenge
Quite a few years ago (2008), when I was chairing the GSMA’s Executive Management Committee (EMC), we had a “My Trail” project on capitalization of user data – simply because all operators have a lot of data about their customers and their networks – but they have been unable to make use of it (in competition with “Over The Top” players collecting data from consumers “over the top” of the networks). This project did not go anywhere at the time, but the world has moved on since then.
Some pre-requisites for autonomous networks, whether an operator will dare to fully go there or not, is 1) that data can be collected from the networks and customer information systems – and 2) that this data can be translated into useful information which can be analyzed and acted upon – more or less automatically. Traditionally, this has been done through various “Operations Support Systems (OSS)” or “Business Support Systems (BSS)” and handled manually.
The silo challenge
OSS or BSS systems are typically specific to the network vendor or to the specific use – and with manual actions taken by people in operations. There are quite a few of these systems and they are not interoperable, although there are Mediation Devices between them in some cases, and traditionally operators have had different people operating different OSS and BSS systems. In an Autonomous Network (AN) perspective, the “automatic network management entity“ will need to manage the whole network and manage the network holistically across silos - which requires a competence across silos that currently sits with a lot of different people.
The volume challenge
Networks have vast amounts of data. Traditionally OSS and BSS systems have helped to some extent, but it is generally too much for all to be collected. Typically, operators have thus had to make a choice: 1) collecting all the data but for a very short amount of time – or 2) sampling the data for a longer amount of time.
In the first case, the collected data is too limited to get a good enough picture of trends and developments that can be used to sensibly adjust or optimize the network parameters (like e.g. e.g. seasonal variations). In the second case, the collected data could miss important events so that the data and consequent actions could be wrong.
I both cases, the volume challenge could easily hamper the ambition of autonomous networks.
The AI model challenge
Another pre-requisite for autonomous networks is the use of Artificial Intelligence (AI). Referring to my comments above on geopolitics and the geopolitical and industrial battle for dominance between the USA, China and others, where we are today is that the world of AI is dominated by companies in these countries. Whenever most people in the Western world are referring to using AI, they are typically using foundation models like Claude, ChatGPT, Gemini etc from competing American companies. Whether they like it or not, these models are good at general reasoning and translating national language input into actionable intent.
The challenge in this area is that, in this case, an operator could be dependent on specific AI models. Operators are used to being dependent on telco vendors like Ericsson or Nokia, but in this case the power might, over time, shift towards such an AI model provider (possibly also a challenge over time for telco vendors?). In a sovereignty perspective, operators may therefore want to be able to use a model of their choice and to change model when relevant – or to use different models for different purposes. Bring your own AI model!
The cost and energy challenge
With extensive use of general AI foundation models, the cost side might become quite serious. According to Bain, AI compute and token costs “will make up to 20% to 30% of the operating model”. Refer also this good article from Patrick Lopez.
I have not double-checked these numbers, and it is not clear to me what the 100% reference is, but with a data platform collecting vast amounts of data and analyzing it, there will be challenges in storage and processing – which again requires racks and GPUs, needing floor space and use of energy. Also, the cost model for AI is typically based on the amount of tokens used (often unpredictable). Autonomous networks might therefore incur high costs.
Efficient use of AI will be necessary – and a strong business case must be made for it. Potential savings with autonomous networks will most likely be in terms of people / headcount in operations and related legacy processes – but this requires that these legacy costs are actually removed. Knowing that people tend to keep on doing what they have always been doing, generally resisting change, and that legacy costs with existing systems and processes are not easily removed, there might be clear challenges in reaping the benefits. There is therefore a risk that operators may get 20-30% additional costs with AI and AN but not getting rid of the legacy costs – and, if it will succeed, it may take a long time. This also rests on the assumptions that AN will make the right decisions and with adequate security.
Degrees of Autonomy
Without elaborating further from my side, the TM Forum categorizes network autonomy into six distinct steps:
“Level 0 (Manual Operations): All operations and maintenance require human intervention, with little to no automation.
Level 1 (Assisted Operations): The system provides descriptive analytics or basic automation, but humans are still required to execute or approve tasks.
Level 2 (Partial Autonomy): Some automated closed loops exist for specific tasks, but significant manual oversight and intervention are needed for decision-making.
Level 3 (Conditional Autonomy): The network can operate autonomously within specific, predefined conditions. It executes automated tasks, escalating to human operators only when it encounters unknown events.
Level 4 (High Autonomy / Intent-Driven): The network is driven by intent. Leveraging AI and Machine Learning, it can predict demands, prevent faults, and self-heal across domains with minimal human input.
Level 5 (Full Autonomy): Complete end-to-end autonomous operations across all network and business domains with seamless, self-optimizing capabilities and no human involvement.”
Most operators today consider themselves in the Level 2 category. Some go further but getting to level 4 or 5 is a long way away.
Summary
Without getting into longer elaborations, here is simply a summary with some thoughts and suggestions for operators in approaching their Autonomous Networks (AN) ambitions:
Whether a full AN is planned or not, operators need to start with a data platform collecting as much relevant data as possible – ideally holistically across silos.
With this collected data the operator can manually assess the information coming out for network optimization or targeted customer relationship management and upselling. Some of this could even be automated – but with people in the loop, quality can be assured and, over time, potentially more can be automated.
To ensure good decisions, it is important that the data platform collects most of the data and that the data is analyzed and structured into useful information.
Data storage and processing are expensive, and so is using AI foundation models, so a good business case for AN will be needed (whether level 4, 5 or lower). It should be noted that the benefits / savings may take a long time to materialize. Also, the revenue side should be considered.
AI is a pre-requisite for AN but operators must ensure that they are not locked into a specific AI foundation model. Operators should be able to use their model of choice – or even use their own SLMs for specific purposes (or a more telco-grade AI model if it comes). Bring your own AI model !
Quite a few countries and companies today are concerned about sovereignty (refer above) – and sovereignty is basically about where is the data, processing and how can it be controlled. Sovereign / on-prem solutions may be required.
Operators are generally afraid of lock-in and have been working on interoperability for decades. Whatever platform they choose, it is critical that it is based on recognized standards, APIs etc.
Finally, to ensure trust in actions taken when using AI, it is important that there are strong mechanisms for governance and auditability.
Finally, AN may happen or not (may be not to the full level?) but there is no rush. Thus:
(Full) AN can wait ! Start slowly !
On the risks for humanity side of AI, although somewhat scary, this will be an interesting area to watch in the coming future.
Comments