Moniepoint副总裁Wole Olorunleke基于处理超过100亿笔交易的实践经验，阐述了构建可信AI系统的核心要素。1他指出，数据治理而非AI模型本身才是决定AI系统可信性的关键因素。1通过完整的交易链条追踪、统一的数据定义和审计机制，AI系统可以从黑箱变为可解释、可追踪的生产系统。1
Olorunleke以MAU（月活跃用户）定义的歧义为例，说明了数据治理的重要性——这一指标在市场营销、产品和财务部门产生了三种不同的理解。1他强调建立了上传者与审批者分离的制度（maker-checker系统），使每笔交易都配有完整的链条记录，追踪决策管道、逻辑和参与人员。1
他还提出了三项评估AI公司数据治理水平的测试方法：三角验证测试（验证Revenue在各仪表板间的一致性）、语义测试（明确定义所有权）和事后分析测试（追踪错误能力）。1Olorunleke认为治理应被视为基础架构的必需组成部分而非合规成本，用他的话说，"治理是引擎，不是制动器"。1他还指出，"没有治理的对话AI界面只是更快写出坏SQL的聊天机器人"。1投资者的观点也印证了这一立场——能够清楚说明数据缺口的公司往往已有计划弥补这些缺陷。1
Wole Olorunleke, Vice President at Moniepoint, shared insights into building trust in artificial intelligence systems through the lens of processing over 100 billion transactions 1. Rather than focusing on AI models themselves, Olorunleke emphasized that data governance is the fundamental determinant of whether an AI system can be trusted in production environments 1. His experience demonstrates that rigorous data management practices—including complete transaction traceability, unified data definitions, and audit mechanisms—transform AI systems from opaque black boxes into explainable and accountable tools 1.
A concrete example illustrates the governance challenge: different departments at Moniepoint developed three distinct interpretations of a single metric, monthly active users (MAU), leading to conflicting understandings across marketing, product, and finance teams 1. To address such issues, Olorunleke implemented core principles including separation of data uploaders from approvers through a maker-checker system, ensuring every transaction maintains a complete audit trail documenting decision pipelines, underlying logic, and responsible personnel 1. He proposed three evaluation tests for assessing data governance in AI companies: a triangulation test verifying revenue consistency across dashboards, a semantic test establishing clear ownership of definitions, and a post-mortem analysis test measuring the ability to trace errors to their source 1.
Olorunleke reframed governance as foundational infrastructure rather than compliance overhead, invoking investor sentiment that "companies able to articulate their gaps have a plan to close them" 1. He concluded that governance serves as an engine driving AI capability, stating that "conversational AI interfaces without governance are merely chatbots that write bad SQL faster" 1.
评论
还没有评论，欢迎留下第一条。