Specialised AI models could help telecoms operators reduce AI costs, improve performance on industry-specific tasks and retain greater control over their technology infrastructure. As AI adoption expands across billing, network operations and customer management, models trained specifically for telecoms are emerging as an alternative to relying solely on general-purpose systems.
General-purpose models such as GPT and Claude offer broad capabilities, but telecoms operators often need detailed knowledge of network architecture, charging systems, technical standards and business support systems. Supplying this context repeatedly can increase processing time and inference costs, particularly when AI is used across high-volume operational workflows.
AT&T’s OTel 2.0 model represents a step towards more specialised AI for the sector. Built on Google’s Gemma 4 31B-IT, it has been post-trained on more than 400 billion telecoms training tokens, drawing on industry standards, technical specifications and documentation from organisations including the GSMA, 3GPP, ETSI, the O-RAN Alliance, TM Forum and ITU.
The model is designed to handle telecoms-specific tasks with greater efficiency, while an accompanying AI Gateway routes requests to models according to their cost, speed and quality. This approach allows operators to use more capable models where deeper reasoning is required and lower-cost alternatives for routine, repeatable tasks.
The potential financial benefits are significant as AI moves from occasional experimentation into everyday operations. Retrieving documentation and supplying extensive context for each request consumes tokens and can increase latency. A model that already understands common telecoms terminology and standards could reduce this overhead, helping operators manage inference costs as usage scales.
Specialised models are also gaining recognition in industry-specific performance benchmarks. AT&T’s OTel-LLM-8.3B-QnA and Chunghwa Telecom’s CHT-TeleX-8B have ranked strongly on the Open Telco AI leaderboard, demonstrating the potential for smaller, domain-adapted models to compete with larger general-purpose systems on telecoms tasks.
The approach does not require operators to abandon frontier AI models. Instead, a multi-model strategy can allocate tasks according to complexity, cost, accuracy and security requirements. General-purpose models remain valuable for broad reasoning, while specialised systems can support more predictable workloads such as analysing billing anomalies, interpreting charging rules, troubleshooting mediation issues and working with telecoms standards.
Cerillion plc (LON:CER) is a leading provider of billing, charging and customer management systems with more than 20 years’ experience delivering its solutions across a broad range of industries including the telecommunications, finance, utilities and transportation sectors.
