Uses for Gen AI in the Telco Network and OSS
By Charlotte Patrick.
A telco network remains the most financially beneficial place to implement intelligence – offering capex and opex savings, as well as supporting new revenue streams.
However, deployment of Gen AI looks likely to be limited to a small number of use cases in the short term because:
- Any deployment of AI/ML into the network present unique challenges, with the potential to harm customer experience
- The complex nature of telco networks necessitates specialised models – which have less applicability than some popular Gen AI areas such as chatbots
- The current general decline in Gen AI enthusiasm as its limitations become clear.
In this environment, it is important to discern what is viable and in what timeframe to make the best use of investments and manage risk. The diagram below describes a range of opportunities being discussed currently.
Gen AI use cases in a telco
The diagram is arranged around seven ways in which Gen AI will be used in telcos:
- Content creation – this area currently generates most headlines with models such as ChatGPT. Includes creating original content (e.g. blog posts, product descriptions for marketing) and audio, image and video content creation. Also, a range of other related capabilities, such as translation and proofreading of text
- Human-machine – centred around improvements in natural language understanding and generation (NLU/NLG), providing better human-machine communications within chatbots, IVRs and digital assistants*. Also, improved accessibility and several Gen AI improvements to sentiment and emotion analysis
- Human-human – the improvements in digital assistants will be most important in the contact centre, helping agents reduce time and improve customer satisfaction when engaged in complex interactions.
- Knowledge management – a group of Gen AI use cases relating to telco catalogues and knowledge bases – for example, in contact centres or field services. Gen AI capabilities will bring improvements from NLU/NLG (similar to chatbot functionality) and be involved in improving the knowledge base itself (e.g. undertaking gap analysis or providing the ability to summarise material).
- Process improvements – A catch-all category that includes Gen AI’s use in the improvement of processes. Including the creation of code, its use in testing and around the management of processes (e.g. API call creation and the improvement of process documentation).
- Data management – This category includes managing underlying data and activities such as governance. Today’s primary use cases appear to be around the generation of synthetic data and the augmentation of existing data sets to improve quality. However, likely that more uses will occur over time.
- Intelligence improvements – Lastly, the use of Gen AI to improve the intelligence within a telco. It is at its strongest, currently, in the area of anomaly detection – but it can also enhance other models for prediction and personalisation.
Observations from the diagram
Creation and updates of topology and architecture When looking at some of the most popular use cases for Gen AI today (providing generation of innovative images and configurations) a question arises about its potential use for generating optimum network topologies, simulating network environments or suggesting cost-saving configuration changes on the network. Vendor products are not yet seen – but the discussion focuses on analysis of various geographical and demographic data with image data of the location, to recommend optimal locations for new cell towers or base stations. Also, creation and update of coverage maps potentially from sparse or incomplete data, where Gen AI is augmented with additional external data, providing information for network planning and also, potentially, in real time for troubleshooting and capacity management.
Other discussions are around predictive models to create network topologies by suggesting the arrangement of nodes, links, and connections. Also, models trained on network topology and configuration data to suggest the configuration of network elements or improvements in energy savings. Questions remain here on whether Gen AI is the most appropriate model type for prediction – and it may become one model amongst others used.
Digital twin management There is discussion of the ability of LLUs to train digital twins on the behaviour of their physical counterparts – creating simplified twins that accurately represent the counterparts while reducing running costs. If LLUs could be utilised in this way, it may make the use of digital twins in the network more viable – as there are many potential use cases for twins in the network but questions about the business case for deployment of many of them.
Digital assistants The use of domain-specific digital assistants to support operational and field services teams. These could provide summaries or answer questions on large bodies of vendor or standards documentation to the NOC/SOC or support field services team when on site. Summarization is the most implemented form of Gen AI currently – making this a good area for investigation.
Creation and maintenance of documentation A related area is the knowledge management of documentation and catalogues in the network and OSS. There are opportunities to improve the quality of documentation provided to a variety of operational teams or SLA documentation and network diagrams provided directly to the customer in regard to their new service.
Code or script creation LLMs offer to decrease coding time across the network/OSS. Examples seen discussed recently include customer-facing tasks, such as a customer’s on-demand service request triggering an LLM to create the necessary scripts and commands for auto-provisioning. Also, network-facing tasks such as the development of network functions, where a relatively unskilled person could be offered recommendations, auto-completing of code and checks against best practise. LLMs can also translate older coding languages into more modern languages to cut support costs.
Declarative instruction The ingestion of written requirements from customers or internal staff in simple language and translation into instructions which can be executed by a variety of automations on the network/OSS.
Multi-agent systems This system is composed of multiple interacting intelligent agents using generative AI and reinforcement learning (RL) techniques. An LLM can already break large tasks into sub-tasks and could break down a task among agents. Multi-agent RL allows these agents to coordinate and act simultaneously on the environment and collaborate towards achieving a joint goal and/or individual targets. Inter-agent communication and shared learning allow the system to adapt to unforeseen challenges and evolving scenarios by collaborative learning.
Validation and testing The strength of Gen AI in anomaly detection and the creation of synthetic data provides opportunities in this area. Real-world testing is expensive and time-consuming; risks develop when test data sets are not kept up to date, leaving the network or other equipment under test vulnerable. Gen AI could generate test data samples more quickly and efficiently and can be used for adversarial testing to test the network, and other models deployed in the network, against attack.
Tasks around network data management: As it becomes more important to manage data sets coming from the network/OSS, there are Gen AI use cases that become relevant in the data management space. Examples include the creation of synthetic data to add synthetic samples for a minority data class, improving a model’s ability to predict rare events accurately and the augmentation of datasets by creating additional samples with slight variations. LLMs could also support the real-time training of ML models used in the network.
Support for predictive models A number of suggestions around the ability of Gen AI to model and predict outcomes were seen. It is currently understood that Gen AI is best used to prevent overfitting, estimate uncertainty, and data augmentation – supporting other models which are better suited to prediction.
Anomaly detection looking for unknown-unknowns Anomaly detection has been a popular use for ML in the network for some time; and is already an important part of security and troubleshooting processes across the network. A generative model doesn’t require a labelled example of every possible anomaly and can detect deviations previously unknown. But, has the potential to be less accurate due to its ability to make biased assumptions. Also, an LLM’s “black box” nature which makes it more difficult to understand why the model has flagged an anomaly. It is unclear what use cases it might be most suited for, but it is possible that it will be used to support other anomaly detection models.
About the Author
Charlotte is an independent industry analyst covering the use of artificial intelligence, automation, and analytics by telecom companies. Her areas of interest are the uptake and efficacy of these technologies and the resulting financial benefit. She brings 13 years of experience as an analyst at Gartner, 5 years as an independent analyst, and 10 years of practical experience in AT&T, COLT and Telefonica O2 – with a mix of strategic, marketing and financial skill sets.