Challenges and enablers3 Faced with these opportunities, organizations require a cohesive strategy, with senior buy-in and long-term investment plans. Fundamentally, CSPs must establish their AI ambition and vision. Four example approaches are outlined, each requiring increasing levels of investment:34 –Non-differentiated adoption: Consume third- party AI services, with progression measured through an AI maturity model. –AI-enabled differentiation: Train existing AI models with proprietary data and assets. –Bespoke foundations: Build bespoke LLMs/ small language models (SLMs) which are specialized for CSP needs and services. –AI-service provision: Build AI use cases and applications for end customers. This articulation helps define success criteria, which in turn informs a clear value-based initiative prioritization model and resulting asset, capability and investment requirements. Based on these requirements, CSPs are equipped to adopt a comprehensive approach to key challenges and enablers, including technical assets, workforce capabilities and responsible AI principles, frameworks and practices. The level of ambition will also determine their ecosystem partnership strategy, which is the subject of Section 4.A successful AI strategy hinges on legacy system modernization, supporting and empowering workers, and ensuring a trustworthy deployment. 3.1 Data, infrastructure and architecture The success of AI models hinges on a strong data foundation that can ingest, correlate and analyse data from multiple sources while enabling integrated, decentralized access for diverse use cases. This approach not only avoids disconnected automation of existing processes but also reimagines processes from the ground up using agentic architecture. However, legacy telecommunications systems have evolved organically with hyper-customized applications, leading to siloed data pools. A lack of “clean, quality, usable data” is perceived as the single largest challenge to implementing AI at scale.35 Further exacerbated by unreliable data quality, accessibility and validity, this contributes to the invisible twin of tech debt – data debt, the inability to unlock value from the vast pools of data with unreliable quality, accessibility and validity. This creates a reliance on manual interventions for both insight generation and data housekeeping – a classic case of looking for a needle in a haystack. The lack of a clean data core makes it difficult for the organization to implement AI. Moreover, the demand for integrated, unstructured data sources to support genAI exacerbates this problem. While 92% of senior CSP executives recognize the role of new data architectures, only 28% have implemented data mesh architectures that enable decentralized access.36 Modernization, therefore, is imperative.37 A value-based approach to prioritize critical data and identify unused “dark data” that is thought to account for 65% of the total data held by organizations38 can streamline modernization efforts. Beyond data, technical infrastructure must support AI with assets such as compute (GPU), model orchestration and agentic frameworks. Complementary core capabilities – advanced analytics, natural language processing and automation – complete the stack. CSPs face strategic choices: develop capabilities in-house, co-develop solutions with partners or buy solutions off the shelf. These decisions will profoundly influence required skill sets, operating models and processes across the organization. A value-based approach to prioritize critical data and identify unused “dark data” can streamline modernization efforts. Artificial Intelligence in Telecommunications 19

Artificial Intelligence in Telecommunications 2025