优化治理以加速生成式AI部署

Optimizing governanceto accelerate GenAIdeployment Author:Dawn Bushaus, Contributing AnalystEditor:Ian Kemp, Managing Editor Contents 3The big picture6Key findings7Section 1:Are CSPs deploying GenAI in production, at scale?11Section 2:Data and AI model governance are inextricably linked17Section 3:Explainable AI provides critical guardrails21Section 4:AI centers of excellence are a steppingstone to becoming AI native27Section 5:How ODA is evolving to optimize AI governance33Additional sponsor feature:Accelerate L4 autonomy: Scale agentic AI solutions into production37Additional resources We hope you enjoythe report and, mostimportantly, find ways touse the ideas, concepts andrecommendations detailedwithin. You can send yourfeedback to the editorialteam at TM Forum viaeditor@tmforum.org The bigpicture Governance is critical in telecoms to ensure ethical, responsible and legaluse of AI technologies. For communications service providers (CSPs),implementing a comprehensive AI governance strategy and frameworkis fundamental to progressing from proofs of concept to realizing truebusiness value from widespread deployment of the technology. “If you look at projects that do not have governance,you will see a lot of failures, because they won’t getpast legal; they won’t get past a risk assessment; andthey won’t go into production,” says Eoin Coughlan, CTOand Industry Lead for Telecommunications, Media andEntertainment at IBM. of concept – provided they can address four mainchallenges, which Willetts describes as “data chaos,fragmented tech, security and governance, and talentand culture”. Data and AI governance are inextricably linked.Indeed, a recent TM Forum Insight survey aboutdata architecture found that lack of a modern dataarchitecture and mindset threatens to keep CSPs fromfully exploiting the benefits of AI. In this report, weanalyze survey results about data governance practicesand explainable AI in more depth. We also provide asnapshot of AI regulation, which is increasing globally. Indeed, two recent studies illustrate how difficultit is to scale GenAI in production. A February 2025report commissioned by Lenovo and authored by IDCfound that 88% of generative AI (GenAI) pilots neverreach production, while a more recent study fromMIT, published in July, highlighted that a staggering95% of GenAI tools fail to deliver expected return oninvestment (ROI). We explain the role of AI centers of excellence inmanaging risk and compliance as operators increasedeployment of both generative and agentic AI usecases. These centers act as an important guardrail forAI, but there is a risk that if governance remains toocentralized, innovation will be stifled. As a result, manyoperators are looking at AI centers of excellence as asteppingstone on the journey to becoming AI-nativeorganizations. “And here in the telco industry, we know that over halfof telco AI initiatives never really get off the startingblocks,” said TM Forum CEO Nik Willetts during hiskeynote at Innovate Americas earlier this month.“That’s not yet scale – that is just frustration.” Watch Nik Willetts' keynote: Governance is imperativeTelco executives are optimistic that their companies can move beyond generative and agentic AI proofs Finally, we look at how the TM Forum Open DigitalArchitecture (ODA) is evolving to help CSPsaddress data and AI governance challenges, andwhy collaboration is key to getting AI agents tocommunicate with each other in a secure andexplainable way. Read this report to understand: •CSPs’ progress in deploying generative and agentic AI•Why data and AI model governance are critical towidespread deployment•Which countries have adopted AI regulation•The role for AI centers of excellence•Why explainable AI is important•How companies like AT&T, Reliance Jio, Telenor, Telusand Vodafone Group are scaling their AI deployments•How ODA can help with data and AI governance. “AI governance is a broad and multifaceted topic. Youcan start at a high level with overarching principles andguidelines and then drill down into how governanceis operationalized through tools, platforms andprocesses… like AIOps, MLOps [machine learningoperations] – and now LLMOps [large language modeloperations],” says Jawad Saleemi, Director of AI &Cloud at Telenor and Co-Chair of TM Forum’sData Architecture Project. Gauging CSPs’ progress in scaling AI The key challenge lies in bringing governance closerto where the execution happens by embedding it intothe workflows where decisions are made and risks areassessed in real time. TM Forum’s Insight team has been surveying CSPs about their progress in deployingGenAI for almost two years. This report draws on the results of three surveys, whichare analyzed in the research shown below. Key report findings AI governance is an extensionof data governance. Poor dataquality can lead to flawed AIoutputs, which are then amplifiedby automation. CSPs must adoptmodern data architectures thattreat data as a produc