量身 AI← 案例库← Case library
案例 · AI 落地方案CASE · AI DEPLOYMENT PLAN · 公共养老金投资 · 多伦多Public pension investing · Toronto

加拿大养老金计划投资委员会 CPP InvestmentsCPP Investments

人人有工具,现在要管好怎么用Everyone has the tools. Now govern how they are used
基于公开资料的分析,与该企业没有合作关系。Based on public information; no affiliation with the company.

英文旁白、无字幕的 3–4 分钟管理层简报。3–4 minute executive brief, English narration, no captions. 在 YouTube 打开Open on YouTube ↗

一句话判断The verdict 工具已遍及全员;公开资料中未见对其使用的治理与审计层。The tools are everywhere; a governance and audit layer for their use is not described in public sources.
起步顺序Order of work 先盘点并分级所有 AI 使用,再把投委会分析正式化,最后搭智能体所需的数据底座。Inventory and tier every AI use first, then formalise committee analysis, then build the data base for agents.

公开事实Public facts 已核对Checked

  • ~2,000
    配备前沿 AI 模型的员工Employees given frontier AI models含 OpenAI CodexIncluding OpenAI Codex
  • ~500
    参加 Alpha Sprints 的员工,每年脱产 3–5 天Staff in Alpha Sprints, 3–5 days off-desk a year用 Codex、Claude Code 解题Solving problems with Codex and Claude Code
  • LLM
    用于评审投委会讨论Reviews investment committee discussions提问是否切中论点,回答是否回应问题?Do questions hit the thesis; do answers respond?
  • 2,125
    2025 财年员工数Employees in fiscal 2025第三方整理自年报Third-party tally of annual-report data

按层级治理 AI 的使用Govern AI use by tier 判断Judgment

  • 日常辅助Everyday assistance:允许使用清单与培训记录Allowed-use list, training record
  • 用 Codex、Claude Code 写的代码与分析Code and analysis built with Codex:进入生产前的复核规则Review rule before production
  • LLM 评审投委会讨论LLM review of committee talks:点名负责人,并书面限定其可判断范围Named owner, written limit
  • 处理定价与参考数据的智能体Agents on pricing and reference data:行动之前先有清单、数据血缘与审计Inventory, lineage and audit first

专家组裁决Expert panel 依据公开观点的推演Inference from public views

Foteini Agrafioti

技术与业务战略对齐Strategy alignment

技术、投资与风险三方的权责写成一页纸;每项 AI 使用都挂到投资或运营目标。One page of ownership across technology, investing and risk; tie each AI use to an investing or operating goal.

Erik Brynjolfsson

价值与度量Value & measurement

集训与工具是互补投入:培训时间算成本,衡量结果而非登录次数。Sprints and tools are complementary investment: count training time as cost, measure outcomes, not log-ins.

Ziad Obermeyer

算法在优化什么What it optimizes

LLM 给投委会讨论打分时,写清它真正测的是什么:问题质量,还是措辞自信。When an LLM scores committee discussion, name what it truly measures: sharp questions or confident phrasing.

Ethan Mollick

人怎么用How people use it

员工已在实验;给安全沙盒,并在真实投资工作中摸清模型的失败点。Staff already experiment; give a safe sandbox and map where the models fail on real investment work.

Andrew Ng

节奏与方法Pace & method

选一个工作流(如投委会分析),三到六个月内出数字。Pick one workflow, such as committee analysis, and show a number in three to six months.

Thomas Davenport

整家企业怎么变Whole-enterprise change

这改变的是投委会的工作方式而不只是工具;点名该流程的业务负责人。This changes how the committee works, not just a tool; name the business owner of that process.

保留的分歧Disagreement kept · 开放沙盒还是设关Open sandbox vs. gate

Ethan Mollick:保持集训开放、鼓励实验;边界就是这样摸出来的。Keep sprints open; that is how edges are found.

Ziad Obermeyer:凡评判人员贡献的模型,先建清单并设关。Gate any model that judges people first.

→ 沙盒保持开放;评判人员或影响资本决策的须过审查。Sandbox stays open; anything scoring people or informing capital decisions passes review.

六位专家是谁、怎样思考、有什么铁律 →Who the six experts are, how they think, their rules →

三个目标Three priorities 建议Recommendation

1

给每项 AI 使用分级Tier every AI use

AI 使用清单,含负责人、风险分级与培训记录。An inventory of AI uses with owners, risk tiers and training records.

2

让投委会分析成为受管工具Make committee analysis a governed tool

从试点升级为有负责人和书面边界的正式工具。Move the pilot to a formal tool with an owner and written limits.

3

建一个 AI 就绪的数据底座Build one AI-ready data base

定价、参考数据与基准,供智能体使用。Pricing, reference data and benchmarks, ready for agents.

场景组合Scenarios 建议Recommendation

场景Scenario做法Approach状态Status
投委会评审成为正式工具Committee review as a formal tool
降风险Risk体验Experience
写明分析在测什么,设定边界,并加入人工复读环节。Name what the analysis measures, set limits, and add a human-read step.建议 · 基于已公开的使用Proposed · builds on public use
统一的 AI 就绪数据层Unified AI-ready data layer
提效Efficiency
定价、参考数据、公司行动与基准共用一层。One layer for pricing, reference data, corporate actions and benchmarks.建议Proposed
资产管理流程智能体Agents for asset-management workflows
提效Efficiency
自动化信用资产管理的例行步骤,由分析师复核。Automate routine credit asset-management steps; an analyst reviews.建议 · 基于公开招聘Proposed · builds on a public job posting
GenAI 审计与数据治理GenAI audit and data governance
降风险Risk
针对生成式 AI 使用与数据访问的审查机制。A review routine for generative A I use and data access.建议Proposed

12 个月路线12-month roadmap 建议Recommendation

奠基Foundation

第 1–3 月Months 1–3
  • AI 使用清单AI-use inventory
  • 风险分级与负责人Risk tiers and owners
  • 一页权责One page of ownership
◆ 清单与分级定稿Inventory and tiers final

试点Pilot

第 4–6 月Months 4–6
  • 点名投委会分析负责人Committee analysis owner named
  • 失败边界测试记录Failure-edge test log
  • 结果基线Baselines for outcomes
◆ 边界成文Limits written

上线Go live

第 7–9 月Months 7–9
  • 受管的投委会工具上线Governed committee tool live
  • 关键输出人工复读Human reads key outputs
  • 上线即监控Monitoring from day one
◆ 风险合规签字Risk sign-off

扩展Extend

第 10–12 月Months 10–12
  • 数据层智能体准入评审Data-layer agent gate review
  • GenAI 审计机制GenAI audit routine
  • 年度重排优先级Annual re-ranking
◆ 进入下一年Next year begins

治理与合规Governance 建议Recommendation

  • AI 使用清单;与法律顾问确认 OSFI E-23 适用性AI-use inventory; confirm OSFI E-23 applicability with counsel
  • 评判人员或影响资本决策的模型:完整审查Models that score people or inform capital: full review
  • 每项资本配置由人决策A person decides every capital allocation

三层指标Three-layer metrics 建议Recommendation

采用Adoption各团队周活跃使用Weekly active use by team · 已入清单的使用占比Share of uses in the inventory
过程Process复核通过率Review pass rate · 培训完成率Training completion · 每季度事件数Incidents per quarter
结果Outcome对照基线的分析师工时Analyst hours against baseline · 决策质量评审Decision-quality review

目标值由企业自定。Targets are the company's to set.

风险Risks 判断Judgment

衡量登录而非结果Measuring log-ins, not outcomes

广泛的使用权限可能掩盖价值单薄;设定结果基线。Wide access can hide thin value; set outcome baselines.

评判人员的模型A model that judges people

给讨论打分可能奖励风格而非实质;先定义目标。Scoring discussion can reward style over substance; define the target.

智能体用了未审计数据Agents on unaudited data

智能体按数据行事;数据血缘必须先行。Agents act on what data says; lineage must come first.

前 90 天First 90 days 建议Recommendation

第 1–2 周Weeks 1–2点名 AI 使用清单负责人Name the AI-use inventory owner
第 3–4 周Weeks 3–4调查各团队的 AI 使用Survey every team's AI uses
第 5–10 周Weeks 5–10风险分级初稿;沙盒规则成文Risk tiers drafted; sandbox rules written
第 13 周Week 13首次复盘First review
今天只需要一个决定:One decision today: 指定 AI 使用清单的负责人name the owner of the A I use inventory

我们可能错在:We could be wrong if: 集训文化本身已能自律,分级层只会拖慢人而不降低风险。the sprint culture already self-governs, and a tier layer slows people without lowering risk.

本页是基于公开资料的分析,与所述企业没有合作关系,不代表其官方立场或内部计划;企业数字多来自其自述或媒体转述;「专家组」是依据六位专家公开观点的推演,不是他们的原话,也不代表他们或其雇主的立场。资料整理日期:2026-11-14。This page is an analysis based on public information. It has no affiliation with the company and does not represent its official position or internal plans. Company figures come mostly from its own statements or media reports. The expert panel is an inference from six experts' public views, not their words, and does not represent them or their employers. Compiled 2026-11-14.

想给你的公司也做一份?Want one for your company?

同样的方法:公开事实、专属判断、12 个月路线和前 90 天。公司大小都可以聊,中文沟通。Same method: public facts, a company-specific judgment, a 12-month roadmap and the first 90 days. Any size of business; we can talk in Chinese or English.

Telegram @johnsunlei邮件 Email [email protected]