直击WAIC 2025 | 专访德勤TMT行业主管合伙人程中：有效的AI治理范式应从被动向主动转变

Core Insights - The imbalance between value extraction and risk management in generative AI has become a critical gap for enterprises to bridge [1] - Deloitte emphasizes that generative AI governance is not an option to delay, and companies must act quickly to clarify responsibilities, enhance skills, and integrate risk management throughout the AI lifecycle [1] Group 1: AI Investment and ROI - The AI transformation process typically involves four stages: establishing an AI strategic vision, pilot exploration, deep integration into core business processes, and financial mapping [4] - In the initial stage, there is often a significant gap between management's ROI expectations and reality, with departments pursuing projects independently [4] - The final stage involves linking AI investments directly to financial metrics, although companies still face challenges in quantifying indirect benefits like customer satisfaction [4] Group 2: AI Architecture and Cost Management - Traditional enterprises face challenges such as complex legacy systems and limited budgets, which can be addressed through "light architecture, soft integration, and distributed evolution" [5] - Light architecture involves encapsulating AI capabilities as API services to reduce the need for core system overhauls [5] - Companies should maintain flexibility in technology selection and establish flexible contracts with technology vendors to mitigate cost risks associated with technology shifts [5] Group 3: Addressing AI Illusions and Black Box Issues - "Illusions" in AI outputs can mislead business decisions and compliance, necessitating a multi-layered defense strategy [6] - Structural illusions, which often appear in AI-generated tables and data analyses, should be prioritized for resolution due to their high risk of misleading decision-makers [6] - To quantify hidden costs from these illusions, companies can assess model output accuracy and operational data impacts [6] Group 4: Risk Mitigation in High-Stakes Scenarios - In high-stakes environments like healthcare and finance, a systematic approach to building illusion mitigation mechanisms is recommended [7] - A mixed architecture of small models and expert rules is suggested for better reliability in regulated fields [7] - Detailed logging capabilities are essential for traceability and accountability in AI outputs [7] Group 5: Strategic AI Governance - Effective AI governance should transition from passive to proactive, with clear strategic goals and dedicated governance teams [11] - Companies should adopt explainable AI technologies and data governance tools to ensure transparency and control [11] - Cultivating employee AI literacy is crucial for fostering a responsible AI usage culture [11] Group 6: AI Security and Revenue Impact - Companies should integrate AI into a unified architecture rather than treating it as an add-on to legacy systems [12] - A secure AI system can enhance customer satisfaction and loyalty, indirectly boosting revenue [12] - Real-world examples show that integrating AI into cybersecurity can significantly reduce response times and downtime, leading to revenue growth [12]