AI 风险管理：实施案例（MindForge）

系统被分类为从“不重要”到“集团关键”的风险等级，这决定了所需的文档和审查深度。

Systems are classified into risk tiers ranging from “Immaterial” to “Group Critical”, which determines the depth of documentation and review required.

LSRC可能要求在部署前进行回访会议，以确认执行符合批准的设计，从而加强整个生命周期的问责制和可追溯性。

The LSRC may require a return session before deployment to confirm execution against the approved design, reinforcing accountability and traceability across the lifecycle.

在初始设计批准后，项目团队在AI注册表中注册了AI用例，并通过AI用例重要性分类评估开始早期风险分类。

Upon initial design approval, the project team registered the AI use case in the AI Registry and began early risk classification via AI use case materiality categorisation assessments.

所有第三方AI系统在生产前需要进行第二次完整的AIWG风险评估和批准，随后根据解决方案的风险等级进行定期重新认证。

A second full AIWG risk assessment and approval is required for all third-party AI systems prior to production, followed by periodic recertification based on the solution’s risk tier.

初始设计概念由本地解决方案审查委员会（LSRC）审查，这是Prudential SDLC中的关键治理检查点，以确保新解决方案、增强功能和架构变更在实施前符合企业标准。

Initial design concepts are reviewed by the Local Solution Review Committee (LSRC), the key governance checkpoint in Prudential’s SDLC, to ensure that new solutions, enhancements, and architecture changes align with enterprise standards before implementation.

用例治理包括AIWG（作为多领域主导审查者，涵盖伦理、安全、性能、隐私、法律、模型开发报告等）、风险与合规（与内外部要求对齐）以及LSRC（架构、数据流和控制），并得到安全、工程和架构团队的输入。

Governance for the use case includes the AIWG, which served as a multi-domain lead reviewer (ethics, safety, performance, privacy, legal, Model Development Report, etc.), Risk & Compliance (alignment with internal and external requirements), and LSRC (architecture, data flows, and controls), with input from security, engineering, and architecture teams.

输入和上传在处理前通过编辑管道进行筛查；数据丢失防护应用于提示和输出；出口仅限于批准的渠道或目的地以最小化泄漏并确保合规；所有提示均记录以供调查和治理。

Inputs and uploads are screened with a redaction pipeline before processing; data loss prevention applies to prompts and outputs; egress is restricted to approved channels or destination to minimise leakage and ensure compliance; and all prompts are logged for investigations and governance.

集中式AI专用基础设施（如适配器层）在入口处应用身份验证和参数验证、个人身份信息检测/屏蔽、毒性/越狱筛查以及不安全格式过滤。

Centralised AI-specific infrastructure such as an adapter layer applies authentication and parameter validation, Personally Identifiable Information (PII) detection/masking at ingress, toxicity/jailbreak screening, and unsafe format filtering.

部署遵循分阶段推出，带有检查表和回滚计划；用户还将接受培训以确保有效和负责任的使用。

Deployment followed a phased rollout with a checklist and rollback plans; users would also undergo training to ensure effective and responsible usage.

变更咨询委员会在上游所有各方签署后给予最终批准。

The Change Advisory Board (CAB) granted final approval after all parties upstream had signed off.

所有必要工件（如测试结果、护栏结果和架构文档）均存档于治理存储库中。

All necessary artefacts – such as test results, guardrail outcomes, and architecture documents – were archived in the governance repository.

部署前，PRUShield 聊天机器人经过严格测试和治理，以确保在性能、安全和业务一致性方面准备就绪。

Before deployment, the PRUShield Chatbot underwent rigorous testing and governance to ensure readiness across performance, safety, and business alignment.

用户验收测试签署于2025年6月第一周获得。

UAT signed off was obtained in the first week of June 2025.

护栏仅强制使用公开可用的、经批准的产品来源和常见问题解答，不允许个性化建议/决策，并要求强制引用。

Guardrails enforce publicly available, approved product sources and FAQs only, with no personalised advice/decisioning permitted, and also require mandatory citations.

工程团队验证了拒绝处理，而业务用户和主题专家的用户验收测试确认了端到端行为，性能在仪表板上跟踪，护栏、越狱拒绝、毒性过滤和PII屏蔽等安全控制符合内部标准。

Engineering validated rejection handling, while UAT with business users and SMEs confirmed end to end behaviour, performance was tracked on dashboards, and security controls such as guardrails, jailbreak rejection, toxicity filtering, and PII masking met internal standards.

聊天机器人的回复必须始终正确且一致。

The chatbot's response must always be correct and consistent.

保诚通过配置选择退出提供商模型训练，转而采用仅推理、零/受控保留实例。

Prudential opts out of provider model training by configuration, opting instead for an inference only, zero/controlled retention instance.

根据用例的风险重要性应用基线或增强要求。

Baseline or enhanced requirements are applied based on the use case's risk materiality.

我们还实施了额外的护栏和控制措施，以安全地探索、测试和采用这些新兴且快速发展的技术。

we also implemented additional guardrails and controls to safely explore, test, and adopt these emerging and rapidly evolving technologies.

所有AI用例和底层模型都在ALAN中注册，这有助于协调AI治理要求在AI生命周期中的应用。

All AI use cases and underlying models are registered in ALAN, which helps orchestrate the application of AI governance requirements across the AI lifecycle.

高级管理层和董事会的监督推动了对AI治理的清晰认识和问责制。

Oversight by Senior Management and the Board of Directors drives clear awareness and accountability for AI governance.

我们的基于风险的AI治理方法确保在整个AI用例生命周期中实施适当且相称的治理。

Our risk-based approach to AI governance ensures appropriate and proportionate governance across the end-to-end lifecycle of AI use cases.

每个AI用例的风险重要性根据RDU委员会定义和认可的重要性准则确定。

The risk materiality of each AI use case is determined against materiality rubrics defined and endorsed by the RDU Committee.

在每个单位和地点分配明确定义的角色和职责，以支持AI治理活动。

Clearly defined roles and responsibilities are assigned within each unit and location to support AI governance activities.

我们通过利用ADA支持模块化集成、可复用模型组件、模板和自动化来解决这一问题。

We address this by leveraging ADA to support modular integrations, reusable model components, templates, and automation.

我们还建立了一个跨职能的负责任AI工作组，由核心职能部门的资深主题专家组成，以彻底评估用例试点并指导风险缓解。

We also established a cross-functional RAI taskforce comprising senior and experienced subject matter experts from core functions to thoroughly evaluate use case pilots and guide risk mitigation.

持续学习的文化使组织能够跟上不断发展的AI技术、法规和社会规范。

A continuous learning culture enables the organisation to keep pace with evolving AI technology, regulations, and societal norms.

该工作组，加上通过RDU委员会提升的审批权限，确保对生成式AI用例有足够的高级管理层监督。

This taskforce, coupled with elevated clearance through the RDU Committee, ensures sufficient senior management oversight on Gen AI use cases.

AI的快速演进需要适应性治理方法，需要审查和增强现有治理实践，以确保它们随着时间的推移保持相关性和有效性。

The rapid evolution of AI necessitates an adaptive governance approach, with the need to review and enhance existing governance practices to ensure that they remain relevant and effective over time.

与监管机构和行业机构的持续接触（例如通过MindForge项目）进一步推动了金融服务行业在AI治理方面的集体进步。

Continuous engagement with regulators and industry bodies (such as through Project MindForge) further drives collective progress in AI governance within the financial services industry.

基于风险重要性的相称治理努力对于在不扼杀创新的情况下实现高效风险管理至关重要。

Proportionate governance efforts based on risk materiality are crucial for efficient risk management without stifling innovation.

我们最初生成式AI的采用范围有意设计为内部使用，具有高水平的人类监督和渐进式采用。

our initial scope of Gen AI adoption was intentionally designed for internal use with high levels of human oversight and incremental adoption.

星展银行的负责任AI方法建立在与业务、分析、风险、合规、技术和人力资源等内部利益相关者的广泛合作之上。

DBS’ approach to responsible AI is built upon extensive collaboration with internal stakeholders from the business, analytics, risk, compliance, technology, and human resource functions.

还实施了一项应急措施，允许用户禁用编码助手功能，以防止对开发活动造成干扰。

A contingency measure which allows users to disable the coding assistant feature was also implemented to prevent disruption to development activities.

该用例随后经过跨职能RDU委员会的全面审查和批准，以对潜在的生成式AI风险进行整体评估。

The use case then underwent a thorough review and approval by the cross-functional RDU Committee for a holistic assessment of the potential Gen AI risks.

还进行了培训，使用户具备熟练和负责任地使用编码助手所需的能力。

Trainings were also conducted to equip users with the necessary competencies for proficient and responsible use of the coding assistant.

部署前，为CodeBuddy制定了全面的监控计划，概述关键性能指标、报告频率、可接受的性能阈值以及沟通/反馈渠道，以确保持续适用。

Prior to deployment, a comprehensive monitoring plan was established for CodeBuddy, outlining key performance metrics, reporting frequency, acceptable performance thresholds, and communication/feedback channels to ensure continued fitness for use.

为进一步验证其性能并减轻不可预见的风险，实施了渐进式、分阶段的部署。

To further validate its performance and mitigate unforeseen risks, a progressive, phased rollout was implemented.

系统被分类为从“无关紧要”到“集团关键”的风险层级，这决定了所需文档和审查的深度。

Systems are classified into risk tiers ranging from "Immaterial" to "Group Critical", which determines the depth of documentation and review required.

AIWG将PRUShield聊天机器人评为中等风险（因其被财务代表用于支持客户互动），并要求每年重新认证以确保持续适用。

The AIWG assigned a Moderate risk rating to the PRUShield Chatbot (given that it is used by Financial Representatives to support customer engagement) and mandated annual recertification to ensure ongoing fit for purpose.

高风险系统需接受正式审查，包括技术深度评估和AIWG投票。

Higher-risk systems undergo formal review, including technical deep dives and AIWG voting.

根据道德原则评估绩效指标，持续监控确保与内部政策和全球标准持续合规。

Performance metrics are assessed against ethical principles, and ongoing monitoring ensures continued compliance and alignment with internal policies and global standards.

该过程以AI风险评估问卷（AIRAQ）为核心，评估系统在透明度、公平性、隐私和问责制等方面的表现。

The process centres on the AI Risk Assessment Questionnaire (AIRAQ), which evaluates systems across dimensions like transparency, fairness, privacy, and accountability.

AI 项目团队使用 Faithfulness 和 Context Precision/Recall 进行抽样评估；当阈值被突破时，回归门触发治理审查。

The AI project team runs sampled evaluations using Faithfulness and Context Precision/Recall; regression gates trigger governance review when thresholds are breached.

重新认证的节奏由风险重要性等级设定（高风险需要更频繁的重新认证），所有审查/结果记录在 AI 注册表中。

The cadence of recertification is set by risk materiality tier (higher-risk requiring more frequent recertification), and all reviews/outcomes are recorded in the AI Registry.

用户反馈循环捕获标记的响应和满意度评分，用于分类和再训练。

A user feedback loop captures flagged responses and satisfaction scores for triage and retraining.

所有 AI 系统，无论是内部还是第三方，都需要定期进行 AIWG 重新认证，涵盖运营/质量指标、护栏重新验证、风险等级重新评估以及持续伦理合规/批准状态的确认。

All AI systems, whether in house or third party, require periodic AIWG recertification, covering operational/quality metrics, guardrail revalidation, risk tier reassessment, and confirmation of continued ethical compliance/approval status.

PRUShield Chatbot 持续监控并正式重新认证，以保持性能、安全性和合规性符合内外部治理期望。

the PRUShield Chatbot is continuously monitored and formally recertified to keep performance, safety, and compliance aligned with internal and external governance expectations.

事件管理遵循企业协议；如果运营受到重大影响，应急计划将移除或替换受影响的 AI 组件，特别是高风险用例，以维护业务连续性和治理完整性。

Incident management follows enterprise protocols; if operations are materially impacted, contingency plans are in place to remove or replace the affected AI component, especially for high risk use cases to preserve business continuity and governance integrity.

运行时监控跟踪故障/事件、安全过滤器激活和检索健康（过时/缺失的产品内容），所有指标集中记录以便追溯和审查。

Runtime monitoring tracks failures/incidents, safety filter activations, and retrieval health (stale/missing product content), with all indicators centrally logged for traceability and review.

将 AI 引入现有非 AI 解决方案会触发新的 AIWG 评估，以维护 Prudential 的 AI 伦理原则。

Introducing AI into existing non AI solutions triggers a fresh AIWG assessment to uphold Prudential’s AI ethics principles.

高风险用例满足以下任一标准：处理敏感数据、数据保留用于训练/微调、自动化操作（无人类参与）。

High-risk use cases are those that meet any of the following criteria: Processing sensitive data. Data is preserved for training/fine tuning. Automated actions (no human in the loop).

所有SaaS AI产品需经过AI风险评估、技术风险评估和第三方风险评估。

All SaaS AI products go through the AI Risk Assessment, Technology Risk Assessment, and Third-Party Risk Assessment.

用例提交至AI委员会，开发和实施方法获得批准。

The use case was presented at the AI Council and the approach to development and implementation was approved.

第三方解决方案需经过AI网络安全审查、法律合规审查和工程实施审查。

A third-party solution was identified for the search and answer, and went through: AI specific cybersecurity review to ensure that both AI- and technology-related risks were identified and mitigated. Legal and regulatory compliance to ensure that the AI solution and output were in line with regulatory requirements. Engineering implementation reviews to ensure that data flow to AI LLMs had the appropriate guardrails and security hardening.

聊天机器人使用检索增强生成（RAG），仅从公开可用、经批准的产品来源和常见问题解答中回答，并拒绝超出范围或特定账户的查询。

The chatbot uses retrieval augmented generation (RAG), answering only from publicly available, approved product sources and FAQs and refusing out of scope or account specific queries.

缓解措施包括实施护栏，如免责声明和提醒，在分享给客户前始终验证聊天机器人的回复。

Mitigation steps included implementing guardrails such as disclaimers and reminders to always validate the chatbot's response before sharing with customers.

检索设置和强制引用与索引一起版本化，以实现透明、可审计、安全的设计响应和基于事实的评估。

Retrieval settings and mandatory citations are versioned with the index for transparent, auditable, secure by design responses and grounding aware evaluation.

与业务用户和主题专家进行的用户验收测试（UAT）验证了端到端行为（代表性问题、歧义处理、拒绝），系统跟踪忠实度和上下文精确度/召回率作为上线门禁和回归防护，在UAT期间和通过定期抽样报告。

User Acceptance Testing (UAT) with business users and SMEs validated end to end behaviour (representative questions, ambiguity handling, refusals) and the system tracks Faithfulness and Context Precision/Recall as go live gates and regression guards, reported during UAT and via scheduled sampling.

强大的数据管理控制确保了符合DBS核心价值观的道德和合法数据处理。

Robust data management controls ensured ethical and lawful data handling that is aligned with DBS’ core values.

CodeBuddy的开发从一开始就明确了用例和模型所有者，以确保AI生命周期中的端到端问责。

The development of CodeBuddy began with the clear identification of use case and model owners from the outset to ensure end-to-end accountability across the AI lifecycle.

CodeBuddy还采用人在回路中的方法，要求用户主动审查和批准生成的输出，以减轻AI不准确或意外行为的风险。

CodeBuddy also employs a human-in-the-loop approach, requiring active user review and approval of generated outputs to mitigate risks from AI inaccuracies or unintended behaviour.

这包括详细的PURE评估和在ADA中提升安全控制，例如实施无状态处理，使LLM能够安全地与DBS数据交互并防止未经授权的第三方访问。

This includes a detailed PURE assessment and uplifting security controls in ADA, such as the implementation of stateless processing, for LLMs to safely interact with DBS data and prevent unauthorised third-party access.

尽管CodeBuddy作为编码助手，用户（DBS员工）对验证生成代码的正确性和适用性承担全部责任。

Although CodeBuddy functions as a coding assistant, users (DBS employees) bear full accountability for verifying the correctness and suitability of the generated code.

为确保整个生命周期的透明度和有效治理，用例和模型细节从早期阶段就记录在ALAN中。

To ensure transparency and effective governance across the lifecycle, use case and model details were documented in ALAN from the early stages.

部署前由高级管理层批准的初步风险重要性评估，评估了解决方案的潜在不利影响和自主性，以确定符合组织AI治理标准的适当治理。

A preliminary risk materiality assessment, approved by senior management prior to deployment, evaluated the solution’s potential adverse impact and autonomy to identify appropriate governance in line with organisational standards on AI governance.

使用检索增强生成，将预训练的LLM与内部知识库集成，将代码建议基于经过验证的来源，从而减轻模型预训练数据中错误或潜在偏差的风险。

The use of retrieval augmented generation, which integrates the pre-trained LLM with our internal knowledge base, grounds code suggestions in verified sources so that the risk of errors or potential biases in the model’s pretrained data are mitigated.

所有用户必须完成涵盖负责任AI使用和AI风险的强制性电子学习模块。

All users must complete a mandatory e-learning module covering responsible AI use and AI risks.

一个可见的免责声明提醒用户注意模型的知识截止日期、生成错误信息的倾向以及在使用前验证所有输出的必要性。

A visible disclaimer cautions users about the model’s knowledge cutoff date, tendency to generate incorrect information, and the need to verify all outputs before use.

此类变更还需记录、审查和批准，以确保可追溯性和问责制。

Such changes are also documented, reviewed, and approved to ensure traceability and accountability.

部署后，CodeBuddy接受持续监控，以跟踪与定义指标相关的性能、评估采用率并收集用户反馈。

Post-deployment, CodeBuddy underwent continuous monitoring to track performance against defined metrics, assess the adoption rates and collect user feedback.

根据我们的AI变更管理流程，变更（例如LLM升级）在生产部署前经过彻底测试，以减轻此类变更可能带来的风险。

In line with our AI change management process, changes (e.g. LLM upgrades) are thoroughly tested before production deployment to mitigate potential risks arising from such changes.

此外，重大变更需经过同行评审流程，并向下游用户进行必要的沟通。

Additionally, major changes are subject to peer review processes, as well as any necessary communication to downstream users.