Norges Bank Sovereign Wealth Fund Cuts Prep Time, Eyes 20% Efficiency Gain With AI

How We Moved 700 Employees from AI Experimentation to Measurable Value — Ai4

Ai4

Norges Bank Investment Management transformed 700 employees across four global offices into AI-enabled workers by pairing top-down executive drive with an ambassador network and multi-model agent infrastructure.

AI-assisted summary of How We Moved 700 Employees from AI Experimentation to Measurable Value — original source. Check the original for full context.

Source published: 2026-08-14

Detailed notes

## Background This video was published on YouTube by the channel Ai4 on August 14, 2026. Stian Kirkeberg, Head of AI & ML at Norges Bank Investment Management, presents how the fund scaled AI across its 700-person global workforce to move from experimentation to measurable ROI and production systems.

## TL;DR Stian Kirkeberg shares Norges Bank Investment Management's (NBIM) strategy for moving AI from pilot projects to production across a $2.2 trillion sovereign wealth fund. Built on a foundation of cloud migration and data modernization, NBIM achieved a 20%+ efficiency target through top-down executive leadership, bottom-up employee empowerment, and an active AI Ambassadors network. The fund deployed multi-model tooling, an Agent Platform, and mandatory upskilling to shift workflows toward autonomous three-person teams and implement production use cases like ethical mandate screening and executive meeting simulation.

## Atmosphere & style The presentation is professional, pragmatic, and highly structured, delivered with the clear, executive tone of a finance and technology leader. Kirkeberg builds a grounded narrative, transitioning smoothly from high-stakes corporate metrics to practical, human-centered operational changes. Watching the session feels like an insider masterclass in organizational change management, offering a calm, authoritative balance of technical depth and strategic transformation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) Norges Bank Investment Management (NBIM) manages roughly $2.22 trillion USD in assets with a team of only 700 employees across four global offices. - [01:38](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=98s) A podcast conversation between NBIM CEO Nicolai Tangen and OpenAI CEO Sam Altman established a strategic target of gaining 20% organizational efficiency using AI. - [03:21](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=201s) A decade of foundational technology modernizations—insourcing back-office operations (2013), migrating to AWS (2018), and moving data warehouses to Snowflake (2021)—enabled rapid AI deployment. - [05:42](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=342s) Organizational adoption was driven by combining top-down executive leadership with bottom-up employee empowerment, including floor walks by the CEO and 80,000+ internal experimentation projects. - [05:42](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=342s) NBIM expanded its dedicated internal AI team from 3 to 15 members and built a network of AI Ambassadors to drive departmental use cases. - [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) The organization utilizes a multi-model architecture with an Agent Platform for developers and a self-service Agent Factory for non-technical employees. - [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) Mandatory upskilling programs were rolled out to all 700 employees, starting with prompt engineering basics and evolving into agent evaluation and system integration. - [11:22](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=682s) Engineering development workflows shifted from traditional 8-developer Scrum teams to autonomous 3-person units consisting of 2 developers and 1 business domain expert. - [15:40](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=940s) Production AI use cases include an ethical and mandate screening agent swarm, as well as a meeting assistant that reduced executive meeting prep time from hours to minutes.

## Detailed walkthrough Here is a detailed, faithful walkthrough based on the transcript provided:

### [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) Managing a $2.2 Trillion Sovereign Wealth Fund Stian Kirkeberg, Head of AI & ML at Norges Bank Investment Management (NBIM), opens by introducing the organization. As the sovereign wealth fund of Norway—commonly referred to as the Norwegian "oil fund"—NBIM manages roughly $2.22 trillion USD in assets under management (AUM). The fund’s core mandate is to build and safeguard wealth for future generations in Norway, withdrawing a maximum of 3% per year.

NBIM’s portfolio spans more than 7,000 listed companies across 63 countries, owning between 1.5% and 2.5% of every listed company globally (representing roughly 70% of the total fund). About 25% is allocated to fixed income, with the remainder in real estate and renewable energy infrastructure. Remarkably, this massive global operation is managed by only 700 employees spread across four offices—Oslo, London, New York, and Singapore—trading continuously 23 hours a day.

- **Assets Under Management:** ~$2.22 trillion USD - **Global Footprint:** Invested in 7,000+ companies across 63 countries (1.5%–2.5% global equity ownership) - **Asset Allocation:** ~70% listed equity, 25% fixed income, remainder in real estate and infrastructure - **Organization Size:** 700 employees across 4 global offices (Oslo, London, New York, Singapore) operating 23 hours/day

### [01:38](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=98s) The Spark: A Podcast Conversation with Sam Altman The journey into generative AI was ignited by NBIM’s CEO, Nicolai Tangen. Tangen hosts a podcast where he invites CEOs and academics to discuss emerging trends. Following episodes featuring leaders like Dario Amodei and Sam Altman, a conversation between Tangen and Altman sparked the organizational push.

They debated how much efficiency a 700-person organization could gain using OpenAI tools. Tangen initially guessed 10%, but Altman proposed 20%, noting the impact on developer productivity. Tangen immediately adopted the 20% target and tasked Kirkeberg with scaling it across the entire organization—from traders and portfolio managers to back-office staff.

### [03:21](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=201s) Modernizing the Tech Stack: Cloud and Snowflake NBIM was positioned to move fast on AI because of foundational cloud and data infrastructure work carried out over the preceding decade: - **2013 (Insourcing):** NBIM insourced its back-office operations and official book of records, which had previously been handled by consultants in India, to take full ownership of internal processes. - **2018 (Cloud Migration):** The fund migrated from two physical data centers in Amsterdam (connected by dark fiber) to AWS. They adopted Infrastructure as Code (IaC), replacing servers instead of patching them. - **2021 (Data Modernization):** The relational SQL setup struggled with cross-server synchronization at scale. NBIM migrated its massive data warehouse to Snowflake, creating new schemas and executing a clean cutover.

This infrastructure foundation allowed NBIM to launch its full-scale AI initiatives smoothly in 2024.

### [05:42](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=342s) Building Organizational Momentum: Top-Down and Bottom-Up To drive adoption across 700 employees, NBIM combined top-down executive leadership with bottom-up employee empowerment: - **Top-Down Drive:** CEO Nicolai Tangen actively champions AI, introducing it in every meeting and walking the floors to directly ask employees how they are applying AI in their daily roles. - **Bottom-Up Facilitation:** Employees are given dedicated time, tools, and training. At one point, employees created over 80,000 projects inside Claude Projects for experimentation and learning. - **Internal AI Team:** The dedicated AI team expanded from 3 to 15 members, overseeing upskilling and architectural deployment. - **Year-Wheel Integration:** AI was placed on the agenda of every internal event—including Tech Days, People Days, and leadership meetings. - **AI Ambassadors Network:** Starting with 20 champions trained alongside Anthropic, ambassadors were tasked with identifying, implementing, and showcasing concrete use cases in their departments.

``` TOP-DOWN (CEO & Executive Push) • Daily leadership focus & floor walks │ ▼ ┌─────────────────────────────────────────┐ │ NBIM AI TEAM (Grown from 3 to 15) │ │ • Platform • Upskilling • Governance │ └─────────────────────────────────────────┘ ▲ │ BOTTOM-UP (700 Empowered Employees) • 80,000+ Claude Projects • AI Champions ```

### [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) Tooling, Multi-Model Architecture, and Upskilling While Anthropic's Claude serves as the primary chat and co-working interface—with over half the workforce using it for code generation—NBIM maintains a multi-model approach using models from OpenAI, Meta, Mistral, and Cohere through Model Context Protocol (MCP) integrations with internal data sources.

To support different technical skills, NBIM built two complementary execution layers: 1. **Agent Platform:** A developer-focused execution environment for deploying autonomous cloud agents. 2. **Agent Factory:** A self-service tool enabling non-technical staff to configure and run task-specific agents.

All employees completed mandatory 30-minute upskilling modules covering AI fundamentals, Claude, prompting, and responsible AI. Six months in, NBIM updated the curriculum to focus on agent evaluation, system integration, and advanced tools.

### [11:22](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=682s) Measuring Value and Driving Behavioral Change NBIM’s goal extends beyond individual efficiency to shifting team-level workflows and culture. Beyond achieving a 20%+ efficiency gain across 700 employees, the fund focuses on alpha generation, enhanced risk reduction, and faster data synthesis across its 7,000 portfolio companies.

Development workflows were restructured around small autonomous units: - **Legacy Approach:** Traditional 8-developer Scrum teams where engineering throughput created bottlenecks. - **Modern Approach:** Autonomous 3-person units (2 developers and 1 business domain expert) equipped with AI coding tools to ship solutions directly.

### [15:40](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=940s) Governance, Compliance, and Production Use Cases NBIM maintains strict human-in-the-loop governance to ensure all AI-assisted decisions remain traceable and verifiable.

``` RAW DATA SOURCES PROCESSING AGENTS OUTPUT ┌────────────────┐ ┌──────────────────┐ ┌──────────────────┐ │ Vendor Feeds │ │ │ │ Active Portfolio│ │ News & Filings │ ───────────► │ Swarm of AI │ ────────► │ Compliance & │ │ Local Media │ │ Screening │ │ Risk Monitoring │ │ Meeting Notes │ │ Agents │ │ │ └────────────────┘ └──────────────────┘ └──────────────────┘ ```

Two key production systems highlight this framework: - **Mandate & Ethical Screening:** A swarm of agents continuously monitors news, financial reports, regulatory filings, and internal meeting logs against investment mandates (e.g., detecting if a portfolio company enters excluded sectors like weapons manufacturing). - **Company Meeting Assistant & Simulator:** Assisting in over 3,200 annual executive meetings, this tool aggregates past notes and filings to generate structured prep agendas with source-attributed citations. Prior to the meeting, analysts can run a simulation where the AI acts as the target company's executive team to roleplay potential Q&A scenarios.

- **Prep Time Reduction:** Pre-meeting preparation time for research analysts dropped from 3–4 hours down to minutes per session.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=0s) Most organizations struggle to move AI from pilot projects to production. — Stian Kirkeberg

> [00:00](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=0s) At NBIM, we transitioned 700+ employees across four global offices from experimentation to measurable impact—achieving 20%+ efficiency gains and deploying production systems. — Stian Kirkeberg

> [00:00](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=0s) I'll share the three-phase transformation strategy that worked: democratization, organizational adoption, and strategic execution. — Stian Kirkeberg

> [00:00](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=0s) You'll learn how we built 71 AI Ambassadors to create organizational muscle memory, moved from widespread experimentation to production systems, and maintained security standards required for a $1.7 trillion sovereign wealth fund. — Stian Kirkeberg

> [00:00](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=0s) If you're leading AI transformation at scale, this session will give you concrete strategies to move your organization from experimentation to value delivery. — Stian Kirkeberg

## Mentioned Stian Kirkeberg, Norges Bank Investment Management, NBIM, Norway, Nicolai Tangen, Dario Amodei, Sam Altman, OpenAI, AWS, Snowflake, Claude, Anthropic, Meta, Mistral, Cohere, Model Context Protocol, MCP, Agent Platform, Agent Factory, Company Meeting Assistant & Simulator

## Video description Stian Kirkeberg - Head of AI & ML, Norges Bank Investment Management

Most organizations struggle to move AI from pilot projects to production. At NBIM, we transitioned 700+ employees across four global offices from experimentation to measurable impact—achieving 20%+ efficiency gains and deploying production systems.

I'll share the three-phase transformation strategy that worked: democratization, organizational adoption, and strategic execution. You'll learn how we built 71 AI Ambassadors to create organizational muscle memory, moved from widespread experimentation to production systems, and maintained security standards required for a $1.7 trillion sovereign wealth fund.

Key takeaways: practical frameworks for scaling AI adoption, measuring real ROI, and building stakeholder trust in regulated environments.

If you're leading AI transformation at scale, this session will give you concrete strategies to move your organization from experimentation to value delivery.

[Source](https://www.youtube.com/watch?v=-yeTtuf5uy4)

中文简报

挪威主权财富基金借助AI削减筹备时间，瞄准20%效率提升挪威央行投资管理公司通过自上而下的高管推动、大使网络以及多模型智能体基础设施相结合，将其在全球四个办事处的700名员工转型为具备AI能力的员工。 Stian Kirkeberg分享了挪威央行投资管理公司（NBIM）在这家规模达2.2万亿美元的主权财富基金中，将AI从试点项目推向生产环境的战略。基于云迁移和数据现代化的基础，NBIM通过自上而下的高管领导、自下而上的员工赋能以及活跃的AI大使网络，实现了超20%的效率提升目标。该基金部署了多模型工具、智能体平台（Agent Platform）及强制性技能提升培训，推动工作流程转向自主的三人团队，并落地了道德准则筛查和高管会议模拟等生产应用场景。挪威央行投资管理公司（NBIM）在全球四个办事处仅有700名员工，管理着约2.22万亿美元的资产。 (0:04) NBIM首席执行官Nicolai Tangen与OpenAI首席执行官Sam Altman在一次播客对话中，确立了利用AI将组织效率提升20%的战略目标。 (1:38) 长达十年的底层技术现代化改造——后台运营内包（2013年）、迁移至AWS（2018年）以及将数据仓库迁至Snowflake（2021年）——为AI的快速部署提供了支持。 (3:21) 组织层面的普及应用由自上而下的高管领导与自下而上的员工赋能共同驱动，其中包括CEO现场走访以及超过8万个内部实验项目。 (5:42) NBIM将其内部专属AI团队从3人扩展至15人，并建立了AI大使网络以推动各部门的应用案例落地。 (5:42)

Reading summary

Norges Bank Investment Management (NBIM) transitioned its 700-person workforce from exploratory generative AI use into structured operations targeting a 20% efficiency gain across its $2.22 trillion portfolio. Stian Kirkeberg, Head of AI & ML, outlines how combining a modernized cloud architecture with top-down executive sponsorship enabled the sovereign wealth fund to deploy autonomous agents and reconfigure core workflows. Foundational Infrastructure Preceding Enterprise AI Deployment (3:21) NBIM's operational capacity—managing roughly $2.22 trillion in assets across 7,000 listed companies in 63 countries with only 700 employees—relies on a multi-stage modernization effort executed over the past decade. In 2013, the fund insourced its back-office operations and official book of records from third-party contractors to regain operational ownership. In 2018, NBIM replaced its dual on-premises data centers in Amsterdam with Amazon Web Services, adopting Infrastructure as Code to ensure servers were systematically replaced rather than patched. Data accessibility was further consolidated in 2021 by migrating NBIM's relational SQL data warehouse to Snowflake after legacy databases struggled with cross-server synchronization at scale. According to Kirkeberg, completing this clean data and cloud transition prior to 2024 created the stable foundation required to deploy enterprise AI models directly against internal records without major architectural bottlenecks. Top-Down Mandates Coupled with Bottom-Up Experimentation (5:42) The fund's generative AI push was initiated by CEO Nicolai Tangen following a podcast discussion with OpenAI's Sam Altman. While Tangen initially estimated potential workforce efficiency gains at 10%, Altman argued that 20% was attainable, prompting Tangen to formalize the 20% target across all operational roles, including traders, portfolio managers, and back-office staff. Adoption was reinforced through daily leadership engagement, with Tangen conducting floor walks to assess how staff integrated tools into active tasks. Concurrently, leadership provided open access to sandbox environments, resulting in staff generating more than 80,000 internal Claude Projects. To sustain adoption, the internal AI team expanded from 3 to 15 personnel, institutionalized AI topics within standard corporate offsites, and established a 20-person ambassador network trained alongside Anthropic to surface departmental use cases. Multi-Model Tooling and Dual-Tier Execution Platforms (8:46) NBIM adopted a multi-model architecture integrating Anthropic's Claude as its primary interface alongside models from OpenAI, Meta, Mistral, and Cohere. Internal systems connect these models to enterprise data sources through Model Context Protocol (MCP) integrations, with over half the workforce actively utilizing Claude for code generation. To accommodate varying technical proficiencies, NBIM deployed two software layers: an Agent Platform providing a developer-centric environment for autonomous cloud agents, and an Agent Factory allowing non-technical employees to configure task-oriented agents via self-service templates. Workforce readiness was supported by mandatory 30-minute training modules covering AI fundamentals, prompt construction, and responsible usage, supplemented six months later with instruction on agent evaluation and system integration. Restructured Engineering Units and Production Monitoring Agents (11:22) AI integration altered team structures and investment workflows. NBIM restructured traditional engineering throughput by shifting away from eight-developer Scrum teams toward autonomous three-person units comprising two developers and one business domain expert, who utilize AI coding tools to build and deploy solutions without conventional backlogs. In production, NBIM implemented a multi-agent screening swarm that continuously monitors global news, regulatory filings, and meeting records to flag potential breaches of ethical and investment mandates, such as unauthorized involvement in weapons manufacturing. Additionally, a meeting assistant aggregates historical notes and regulatory filings to prepare analysts for over 3,200 annual executive engagements. The system simulates target executive responses for interview rehearsal and has reduced pre-meeting analyst preparation time from 3–4 hours to several minutes, while maintaining human-in-the-loop verification. NBIM demonstrates that substantial productivity gains from enterprise AI depend on prior infrastructure modernization and structural workflow redesign rather than isolated tool adoption. The fund's governance model continues to rely on human oversight to audit model citations and verify autonomous agent outputs across its high-stakes investment mandates.

精炼阅读摘要

挪威主权财富基金（NBIM）人工智能与机器学习负责人 Stian Kirkeberg 阐述了该机构如何推动 700 名员工跨越概念验证阶段，将生成式 AI 深度融入管理规模达 2.22 万亿美元的投资运营流程中。该基金依托此前十年积累的云原生与现代数据底座，配合双向驱动的组织动员及新型人机协同架构，在推进全员 20% 效率提升的同时实现了关键投研流程的重塑。十年基建奠基与效率目标确立 (1:38) 作为管理约 2.22 万亿美元资产的主权财富基金，NBIM 持有全球 63 个国家超过 7000 家上市公司的权益（约占全球上市股权的 1.5% 至 2.5%），但全机构仅有 700 名员工分布在奥斯陆、伦敦、纽约和新加坡四个办公室，维持着全天 23 小时的连续交易。首席执行官 Nicolai Tangen 在播客中与 OpenAI 首席执行官 Sam Altman 探讨后，采纳了 Altman 提出的全组织效率提升 20% 的目标，并责成 Kirkeberg 在交易员、投资组合经理及后勤支持等全员范围内推广落地。这一快速推进的基础源于过去十年开展的三次核心技术升级：2013 年，NBIM 收回此前由印度咨询公司承担的中后台运营与官方账本，实现核心流程完全自主可控；2018 年，机构将阿姆斯特丹的双物理数据中心迁移至 AWS 云端，全面引入基础设施即代码（IaC）并确立以系统置换替代传统补丁的原则；2021 年，为解决关系型 SQL 架构在跨服务器同步时的扩展瓶颈，机构将核心数据仓库整体割接至 Snowflake。这一底层架构确保了 2024 年大规模 AI 应用能够顺畅对接底层数据。自上而下与自下而上的全员转型 (5:42) 为推动 700 名员工真正采纳 AI，NBIM 建立了高层驱动与基层赋能深度绑定的运作机制。在自上而下的推动中，首席执行官 Tangen 持续巡视办公区，直接向员工询问 AI 工具的具体使用情况，并将 AI 固定列入技术日、全员日与管理层例会等年轮日程。在自下而上的赋能中，机构向员工提供实验时间、工具支持与环境，员工在 Claude Projects 中建立的探索项目累计超过 8 万个。为支撑组织转型，专职 AI 团队编制从 3 人迅速扩展至 15 人，统筹平台架构、技能重塑与合规治理。机构联合 Anthropic 针对首批 20 名核心骨干开展深度培训，建立起覆盖各业务部门的 AI 大使网络，负责发掘并展示本地化用例。此外，全员均完成了涵盖基础原理、Claude 操作、提示词工程与负责任 AI 的 30 分钟必修课程；半年后，培训体系进一步转向 Agent 评估和系统集成等高阶应用。多模型架构与敏捷研发重塑 (8:46) 在技术架构上，NBIM 采用了多模型策略。尽管 Claude 作为主要的日常交互工具且有超过半数员工使用其辅助编写代码，机构仍通过模型上下文协议（MCP）将 OpenAI、Meta、Mistral 和 Cohere 等模型与内部数据资产连通。在执行层面，团队打造了差异化的双层运行环境：面向开发者的“Agent 平台”用于运行自主云端 Agent，面向非技术人员的自服务“Agent 工厂”则允许业务人员自行配置并执行特定任务的自动化流程。技术能力的普及促使研发组织形态发生根本改变。传统的 8 人 Scrum 敏捷开发团队因工程交付瓶颈被打破，重构为由 2 名开发人员与 1 名业务领域专家组成的 3 人自主单元。在 AI 编码工具的辅助下，这种小型单元可以直接承接业务诉求并独立完成端到端开发与交付，消除了跨部门沟通带来的交付阻滞。人机协同治理与核心投研落地 (15:40) NBIM 建立了严格的“人在回路”（Human-in-the-loop）治理机制，确保所有 AI 辅助决策过程具备审计追踪与合规验证能力。该框架下落地了两个代表性生产系统：一是投资授权与道德合规审查群，由一组协同运行的 Agent 自动化监测新闻、财报、监管披露及内部会议记录，全天候对比投资授权基准，在持仓企业涉足武器制造等受排斥行业时触发合规预警。二是服务于每年 3200 多场上市公司高管对话的会议助手与模拟系统。该系统自动汇聚历史纪要与监管披露，生成带有原始信源引注的结构化会议议程，使分析师的单场调研准备时间从原先的 3 至 4 小时大幅缩减至数分钟。在正式会议前，分析师还可调用模拟器让 AI 扮演目标企业高管团队进行问答推演，从而在会前检验论点并优化质询策略。 Kirkeberg 阐明了高度精简的人员编制与庞大管理资产规模如何促使机构将 AI 转化为生产力杠杆。然而，从单点流程的工时缩减到实现全员 20% 的综合生产力提升，以及其对长期超额收益获取与风险控制的实际贡献度，仍依赖于后续工作流演化与更长周期的持续检验。

Detailed notes — zh

## Background 该内容由 Ai4 于 2026 年 8 月 14 日发布在 YouTube 平台上。根据页面简介，本次演讲由 Norges Bank Investment Management（NBIM，挪威石油基金）的 AI 与机器学习 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) 挪威主权财富基金（NBIM）管理着约 2.2 万亿美元资产，全球 700 名员工分布于四个办公室，全天候运营并持有全球大量上市公司股份。 - [01:38](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=98s) 转型始于 CEO Nikolai 与 OpenAI CEO Sam Altman 的交流，最终定下在全公司范围内利用 AI 实现 20% 效率提升的明确目标。 - [02:59](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=179s) NBIM 过去在技术基础设施上的长期投入（内包中后台、全面上云及迁移至 Snowflake）为 2024 年的大规模 AI 应用奠定了坚实基础。 - [05:47](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=347s) 采用自上而下的 CEO 强力推动与自下而上的全员工具赋能相结合的策略，极大地促进了全公司的 AI 实验与实践。 - [05:47](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=347s) 通过建立由 71 名“AI 大使”构成的内部网络和全景日程，在各部门内部寻找实际用例并向全公司推广展示。 - [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) 以 Claude 为核心构建工具栈，普及 Claude Chat 和 Claude Code，并建立内部智能体平台与智能体工厂来承载独立运行的智能体。 - [10:30](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=630s) 推出面向全体员工的强制性 30 分钟技能提升课程，并随技术迭代将重点转向 Claude Code、智能体构建与评估。 - [10:30](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=630s) 转型最核心也最具挑战的环节是改变员工与团队的日常工作习惯，跨越个人使用与团队协作流程之间的“文化冲突”。 - [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 不仅实现了 20% 以上的运营效率提升，AI 更通过深入挖掘海量数据辅助投资决策，创造 Alpha 收益并降低风险。 - [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 开发模式从传统的 8 人两披萨团队转变为由 2 名开发人员与 1 名业务人员组成的精干自主团队，极大提升了交付效率。 - [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 在治理上坚持人类问责制，要求人类对 AI 的建议和决策保持全流程监控、可追溯性与质疑能力。 - [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) 利用“智能体群”对全球 7,000 家被投公司进行全天候自动化合规筛查，确保符合挪威政府的投资政策规范。 - [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) 针对每年 3,200 多场公司高层会议，系统可自动调取历史数据生成可追溯的草拟议程，极大缩短分析师的前期准备时间。 - [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) 引入 AI 角色扮演模拟器，让 AI 扮演目标公司回答问题，帮助研究分析师在实际会议前进行高度逼真的模拟演练。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) Norges Bank Investment Management 与主权财富基金背景演讲者 Stian Kirkeberg 首先介绍了 Norges Bank Investment Management（NBIM）的基本情况。作为挪威的主权财富基金，其资产管理规模（AUM）大约达到 2.2 万亿美元。基金的核心目标是为挪威的未来几代人保障、安全并建立财富。由于资金主要来源于挪威政府及石油相关活动，因此该基金在民间通常被称为“挪威石油基金”。基金每年最多使用总额的 3% 作为开支上限。在投资组合方面，NBIM 的资产分布广泛： - 投资于全球 63 个国家的 7,000 多家公司。 - 持有全球每家上市公司的股份比例大约在 1.5% 到 2.5% 之间，这部分权益类投资占了基金总规模的大约 70%。 - 固定收益资产（Fixed income）占了 25%。 - 剩余部分则投资于房地产和基础设施。让人惊叹的是，管理这庞大资产规模的员工总数仅有 700 人，分布在新加坡、奥斯陆、伦敦和纽约四个全球办公室，并且基本上保持着每天 23 小时的全天候交易。 - **资产管理规模**：约 2.2 万亿美元。 - **办公室分布**：新加坡、奥斯陆、伦敦、纽约。 - **资产配置**：约 70% 股票（投资全球 6,000-7,000+ 家公司）、25% 固定收益、其余为房地产与基础设施。 - **员工人数**：700 人。 --- ### [01:38](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=98s) AI 旅程的起点：CEO 的播客与目标设定关于 AI 旅程是如何开始的，Stian 提到公司的 CEO Nikolai Tangen 拥有一个播客节目，经常邀请其他公司的 CEO、教授以及能够带来灵感的人士进行对话。Nikolai 曾邀请过 Anthropic 的 Dario Amodei 以及 OpenAI 的 Sam Altman 上播客节目。 AI 转型真正拉开序幕，源于 Nikolai 与 Sam Altman 的一次讨论。当时两人在节目中互相抛出数字，试图估算仅仅通过使用 AI 工具，这 700 名员工的效率究竟能提升多少。当时的焦点主要集中在 OpenAI 的工具上： - Nikolai 提出提升 10% 的效率。 - Sam 在考虑到开发人员及其工作重点后，将数字提升到了 20%。然而，Nikolai 回头就把 20% 这个数字交给了 Stian，并要求在全公司范围内落实。挑战在于：CEO、交易员、投资经理、研究分析师或中后台人员的工作内容截然不同，如何让 AI 在公司的每一个人手中都真正飞起来，成了摆在技术团队面前的重大课题。 --- ### [02:59](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=179s) 为 AI 时代的到来奠定坚实基础为了支撑后来的 AI 大爆发，NBIM 其实在过去多年里默默完成了多项重大的基础设施准备工作。Stian 庆幸他们提前完成了这些蜕变： - **2013年（内包 back-office）**：将所有的中后台业务及账务记录（book of records）收回自营。此前曾由印度的顾问团队提供出色的服务，但公司希望自主掌握流程。 - **2018年（全面拥抱云端）**：选择 AWS 作为云服务商。此前他们在阿姆斯特丹拥有两个数据中心，并通过黑光纤（dark fiber）连接，面临着许多组织都经历过的容量瓶颈。迁移初期采取了“原样搬迁”（as is）策略，唯一的改变是引入了基础设施即代码（Infrastructure as Code）——当需要修补服务器时，不是直接修补，而是拉起一个新服务器并替换掉旧服务器。 - **大型数据库瓶颈**：随着巨型数据仓库迁移上云，AWS 给出了他们能提供的最大实例来运行 Microsoft SQL Server，并采用了链接服务器（linked servers）。但即便是最大实例，也无法承受数据同步的高强度负荷。 - **2021年（迁移至 Snowflake）**：果断将系统迁移至 Snowflake 现代数据仓库平台，为此进行了重建架构、并行运行等大量痛苦的工作，最终成功转型，为 2024 年大爆发铺平了道路。 --- ### [05:47](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=347s) 推动 700 名员工采用 AI 的策略与组织架构为了让 700 名员工全面接受 AI，NBIM 采取了自上而下与自下而上相结合的立体战略： - **顶层推动（Nikolai 的“金霸王电池”效应）**：CEO Nikolai 在日常每个会议中都会谈论 AI，甚至会亲自在办公室走廊巡视，走到员工工位旁点名并坐下来，要求对方展示实际用 AI 在做什么。这种既温和又带压力的督导方式让员工不敢懈怠。 - **底层赋能**：公司为员工提供工具、时间和培训，鼓励他们在日常工作中寻找 AI 的应用场景。在引入 Claude 之后，仅 Claude Projects 就曾一度衍生出超过 80,000 个不同的项目（其中许多是概念验证尝试，但这带来了极佳的学习效应）。 - **AI 团队与矩阵式日程**：AI 团队从最初的 3 人扩大到了当前的 15 人。公司设立了“年度全景日程”（year wheel），在全公司所有的活动中——无论是技术日（tech days）、全员日（people days）还是领导小组会议——AI 都是不可或缺的核心议题。此外，各办公室还会定期举办内部技术日，让员工向组织展示成果，形成内部推介（evangelism）。 - **AI 大使计划（AI Ambassadors）**：最初精选了来自不同部门的 20 名大使（后来扩展到 71 人），由 Anthropic 团队为其提供额外培训。他们的核心任务很简单：在自己的团队中找到一个好的用例，在 AI 团队和 Anthropic 的协助下将其实现，然后向全公司展示，以带动整体士气与内部势头。 --- ### [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) 内部 AI 工具栈与模型生态在工具应用方面，公司内部主要采用 Claude 作为核心 AI 工具： - **全员普及**：所有员工都在使用 Claude Chat 和 Claude Cowork。 - **数据表明**：有一半的员工在使用 Claude Code——这意味着即使是投资经理或交易员，日常也会编写一些代码，这使得他们与其他基金有着明显的不同。 - **平台建设**： - 构建了**内部智能体平台（internal agent platform）**：用于运行需要脱离个人电脑、在云端独立运行的智能体，这需要一定的编程或协助技能。 - 构建了**智能体工厂（agent factory）**：为普通员工提供自助服务解决方案，满足日常快速生成和运行智能体的需求。 - **MCP 与多模型策略**：通过 MCP（Model Context Protocol）将几乎所有的内部数据源打通并连接到 Claude 和其他智能体中。除了 Anthropic（Claude）之外，他们还会根据具体用例的需要，灵活采用 OpenAI、Meta、Mistral、Cohere 等不同厂家的模型。 --- ### [10:30](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=630s) 强制性技能提升与文化挑战为了改变员工习惯，公司推出了面向全体员工的强制性 30 分钟快速 upskill（技能提升）课程。课程内容涵盖： - AI 与 NBIM 简介 - 认识 Claude - 提示词工程（Prompting） - 批判性思维、治理与负责任的 AI（Responsible AI）然而在仅仅实行了 6 个月后，由于工具和概念迭代速度太快，关于提示词的教学已不再是最重要的内容，课程重心迅速转向了 Claude Code 的使用、智能体（Agents）的构建以及智能体评估。 Stian 指出，整个转型的终极目标是**改变人们的习惯**，让 AI 融入日常工作的方方面面，而不仅仅是用来总结一封邮件——“如果那（总结邮件）就是你做的全部，说明你用得还不够。”改变个人习惯极其困难，但更具挑战性的是**文化冲突（cultural crash）**：当个人频繁使用 AI 时，并不意味着团队协作层面的工作流程也同步发生了转变。目前，他们正在极力跨越这道文化关卡。 AI 团队目前的四个平行发展轨道： 1. **技能提升**（Upskilling） 2. **平台建设**：让所有人都能极其简便地部署智能体。 3. **高层访谈与价值对齐**：与公司各部门负责人进行访谈和工作坊，找出最具实施价值的项目。 4. **外部伙伴关系**：与 AWS、Anthropic 及其他同行保持紧密合作，走在技术前沿。 --- ### [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 业务价值：效率提升、投资回报与去风险化谈到切实的业务回报，NBIM 取得了显著的成效： - **效率增益**：年复一年实现了 20% 甚至更高的效率提升。对于 700 名员工的规模而言，其释放出的价值在具体项目中表现得极为巨大。 - **Alpha 创造与投资回报**：除了行政和运营效率外，基金价值的另一大来源是提高投资回报。通过 AI 辅助做出更明智的决策，进行 Alpha 创造，自动化流程和降低风险带来的效益远远超出了单纯的运营效率提升。 - **深度数据挖掘**：面对需要紧密跟踪的数千家公司，人工进行全量数据审查非常耗时。AI 帮助团队更深地挖掘海量数据、财报、新闻和公司会议记录，从而做出更具知情权的决策。在开发模式上，传统的 Scrum 开发方法遭遇了“两披萨团队（two-pizza team）”的瓶颈——过去由 8 名开发人员和 1 名产品经理组成的团队，由于开发人员跟不上代码交付速度，常常导致项目拖延。现在，他们转变为**由 2 名开发人员和 1 名业务人员组成的小型自主团队**，信任他们端到端完成整个项目，效率呈指数级上升。不过，组织内部滋生官僚主义和层级扩大的趋势是一大挑战，公司必须极力维护这种小型自主团队的敏捷性。在**治理层面**，虽然技术上可以在几小时内跑完整个 AI 投资组合，但公司始终坚持**问责制**：要求人类对所有发生的事情进行监控、保持可追溯性，并能够随时向 AI 的建议发起挑战和迭代。 --- ### [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) 落地实例：公司筛查、3200+ 场会议与智能模拟 Stian 分享了具体的落地案例，展示 AI 如何重塑日常核心业务流程： - **7,000 家公司的合规筛查**：挪威政府对基金有严格的投资政策与政策委任（mandates），例如严禁投资武器制造等特定领域。为了实时掌握这 7,000 家被投公司的动态，AI 系统会将供应商数据、新闻报道、财报、契约申报、本地媒体以及公司每年的内部会议纪要全部输入给一组协同工作的“智能体群（swarm of agents）”，进行全天候的自动化 SCHEDULE 监控和筛查。 - **公司会议助手与模拟器**：结合上述能力，该系统支撑了每年超过 3,200 场公司高层会议（针对重点跟进的公司）。 - **背景准备**：过去研究分析师准备一场会议需要 3 到 4 个小时。现在以 Meta 为例，分析师输入公司名称后，系统会调取历史会议记录和相关资源，自动生成一份**草拟议程（draft agenda）**。议程旁边标有绿黄相间的标记，点击即可追溯到 AI 提取该信息的原始出处。 - **侧边栏交互**：分析师可以在侧边栏通过聊天实时迭代、修改这份工作议程。 - **模拟（Simulation）阶段**：一旦议程敲定，系统中的模拟功能便会启动——**AI 可以扮演被会见的公司的角色**，尝试按照该公司的口吻和逻辑回答分析师的问题，从而让研究分析师在走进真正的会议室之前做好充分的准备。 Stian 最后表示，他们仍在不断将更多数据源引入这一系统，并在公司内部实现更广泛的普及。 [https://www.youtube.com/watch?v=-yeTtuf5uy4](https://www.youtube.com/watch?v=-yeTtuf5uy4) ## Notable quotes > [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) 因为资金主要来源于挪威政府及石油相关活动，因此该基金在民间通常被称为“挪威石油基金”。 — Stian Kirkeberg > [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) 挪威的主权财富基金，其资产管理规模（AUM）大约达到 2.2 万亿美元。 — Stian Kirkeberg > [00:04](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=4s) 管理这庞大资产规模的员工总数仅有 700 人，分布在新加坡、奥斯陆、伦敦和纽约四个全球办公室，并且基本上保持着每天 23 小时的全天候交易。 — Stian Kirkeberg > [05:47](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=347s) 为了让 700 名员工全面接受 AI，NBIM 采取了自上而下与自下而上相结合的立体战略。 — Stian Kirkeberg > [05:47](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=347s) 仅 Claude Projects 就曾一度衍生出超过 80,000 个不同的项目。 — Stian Kirkeberg > [08:46](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=526s) 有一半的员工在使用 Claude Code——这意味着即使是投资经理或交易员，日常也会编写一些代码，这使得他们与其他基金有着明显的不同。 — Stian Kirkeberg > [10:30](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=630s) 如果那（总结邮件）就是你做的全部，说明你用得还不够。 — Stian Kirkeberg > [10:30](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=630s) 整个转型的终极目标是改变人们的习惯，让 AI 融入日常工作的方方面面，而不仅仅是用来总结一封邮件。 — Stian Kirkeberg > [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 年复一年实现了 20% 甚至更高的效率提升。 — Stian Kirkeberg > [13:17](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=797s) 从过去由 8 名开发人员和 1 名产品经理组成的团队，转变为由 2 名开发人员和 1 名业务人员组成的小型自主团队。 — Stian Kirkeberg > [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) 挪威政府对基金有严格的投资政策与政策委任（mandates），例如严禁投资武器制造等特定领域。 — Stian Kirkeberg > [16:16](https://www.youtube.com/watch?v=-yeTtuf5uy4&t=976s) AI 可以扮演被会见的公司的角色，尝试按照该公司的口吻和逻辑回答分析师的问题，从而让研究分析师在走进真正的会议室之前做好充分的准备。 — Stian Kirkeberg ## Mentioned Stian Kirkeberg, Norges Bank Investment Management, NBIM, Nikolai Tangen, Sam Altman, Dario Amodei, OpenAI, Anthropic, Claude, Claude Projects, Claude Chat, Claude Cowork, Claude Code, AWS, Microsoft SQL Server, Snowflake, Meta, Mistral, Cohere, MCP, Model Context Protocol ## Video description Stian Kirkeberg - Head of AI & ML, Norges Bank Investment Management Most organizations struggle to move AI from pilot projects to production. At NBIM, we transitioned 700+ employees across four global offices from experimentation to measurable impact—achieving 20%+ efficiency gains and deploying production systems. I'll share the three-phase transformation strategy that worked: democratization, organizational adoption, and strategic execution. You'll learn how we built 71 AI Ambassadors to create organizational muscle memory, moved from widespread experimentation to production systems, and maintained security standards required for a $1.7 trillion sovereign wealth fund. Key takeaways: practical frameworks for scaling AI adoption, measuring real ROI, and building stakeholder trust in regulated environments. If you're leading AI transformation at scale, this session will give you concrete strategies to move your organization from experimentation to value delivery. [Source](https://www.youtube.com/watch?v=-yeTtuf5uy4)

Source quotes

[0:04] Stian Kirkeberg: 因为资金主要来源于挪威政府及石油相关活动，因此该基金在民间通常被称为“挪威石油基金”。 [0:04] Stian Kirkeberg: 挪威的主权财富基金，其资产管理规模（AUM）大约达到 2.2 万亿美元。 [0:04] Stian Kirkeberg: 管理这庞大资产规模的员工总数仅有 700 人，分布在新加坡、奥斯陆、伦敦和纽约四个全球办公室，并且基本上保持着每天 23 小时的全天候交易。 [5:47] Stian Kirkeberg: 为了让 700 名员工全面接受 AI，NBIM 采取了自上而下与自下而上相结合的立体战略。 [5:47] Stian Kirkeberg: 仅 Claude Projects 就曾一度衍生出超过 80,000 个不同的项目。 [8:46] Stian Kirkeberg: 有一半的员工在使用 Claude Code——这意味着即使是投资经理或交易员，日常也会编写一些代码，这使得他们与其他基金有着明显的不同。 [10:30] Stian Kirkeberg: 如果那（总结邮件）就是你做的全部，说明你用得还不够。 [10:30] Stian Kirkeberg: 整个转型的终极目标是改变人们的习惯，让 AI 融入日常工作的方方面面，而不仅仅是用来总结一封邮件。 [13:17] Stian Kirkeberg: 年复一年实现了 20% 甚至更高的效率提升。 [13:17] Stian Kirkeberg: 从过去由 8 名开发人员和 1 名产品经理组成的团队，转变为由 2 名开发人员和 1 名业务人员组成的小型自主团队。 [16:16] Stian Kirkeberg: 挪威政府对基金有严格的投资政策与政策委任（mandates），例如严禁投资武器制造等特定领域。 [16:16] Stian Kirkeberg: AI 可以扮演被会见的公司的角色，尝试按照该公司的口吻和逻辑回答分析师的问题，从而让研究分析师在走进真正的会议室之前做好充分的准备。

Source quotes

[0:00] Stian Kirkeberg: Most organizations struggle to move AI from pilot projects to production. [0:00] Stian Kirkeberg: At NBIM, we transitioned 700+ employees across four global offices from experimentation to measurable impact—achieving 20%+ efficiency gains and deploying production systems. [0:00] Stian Kirkeberg: I'll share the three-phase transformation strategy that worked: democratization, organizational adoption, and strategic execution. [0:00] Stian Kirkeberg: You'll learn how we built 71 AI Ambassadors to create organizational muscle memory, moved from widespread experimentation to production systems, and maintained security standards required for a $1.7 trillion sovereign wealth fund. [0:00] Stian Kirkeberg: If you're leading AI transformation at scale, this session will give you concrete strategies to move your organization from experimentation to value delivery.

Original transcript

[0:04] So who are Noris Bank investment management? We are the sovereign wealth fund of Norway. [0:13] We have roughly $2.2 2 trillion US dollar under uh asset under management [0:20] and our goal is to safeguard and uh secure and build the wealth for future generation in Norway. [0:30] So we are popularly called the Norwegian oil fund because all the money from [0:36] our uh government and oil activities goes into this fund for uh [0:44] for the next generations. We only use maximum of 3% of it per year. So we are invested in [0:53] more than 7,000 companies throughout the 63 countries. We own about 1.5 to 2.5 [1:02] of every company uh listed company in the world and this accounts for roughly 70% of the fund. [1:11] Uh 25% is fixed income and the rest is invested in real estate and infrastructure. [1:21] Managing all of this is only 700 employees. We have office four offices Singapore, [1:27] Oslo, London and York. And we basically trade 23 hours uh every day. [1:38] So I'm not going to talk about all investment, but I'm going to focus on our AI journey [1:44] and how did all of this start. Our CEO Nikolai Tongen has a podcast. Invites [1:53] other CEOs, professors, people he wants us to get inspiration from to his podcast [2:01] and they discuss whatever is the topic of today. He has had both u [2:07] Dario and Sam on the podcast and the AI journey really started with a discussion between [2:14] Nikolai and Sam. They were tossing numbers back and forth trying to uh decide on a number [2:22] of how much more efficient can we become as as these 700 employees by just using uh AI and the [2:34] the focus then was on on um open AI tools only. [2:40] Nikolai said 10. Sam uh reverted with 20 after thinking about developers and the focus they do. [2:48] But uh as it turns out, Nikolai took that number, turned around and asked me or told me [2:54] to implement this for the whole of the organization. So how can we do this? [2:59] Uh what Nikolai does is not the same as I do nor a trader, portfolio manager, a research analyst [3:05] or anyone in the back office. So we needed to make this fly for absolutely everyone in the company [3:13] and to get into the details there I will take you on a little bit of a backst step [3:21] from what we've done to prepare us for this for this journey and I think we're are actually a [3:27] bit lucky that we implemented all of this before AI become uh the thing so in 2013 we [3:36] insourced all of our back office, our book of records. We had a lot of consultants in India. [3:42] They were doing a great job, but we wanted to own the process ourself. So, we took it all home. [3:49] In 2018, we went all in for the cloud. We selected AWS as our uh cloud vendor. [3:57] uh we had had two data centers in Amsterdam, black fiber between them [4:03] and we had all of this issue with capacity which I'm sure you also experienced. [4:08] When we first got to the cloud we moved everything as is. The only change we did was to [4:14] create infrastructure as code. Meaning that if we needed to patch a server, we didn't uh [4:21] patch it. We patched a new one, brought it up, and then replaced the one we needed to take out. [4:29] One of the challenge coming with our gigantic data warehouse into the cloud was also capacity. AWS [4:35] gave us the biggest instance they got to run our Microsoft SQL servers. We used linked servers. Um, [4:44] that was no longer a good idea. even their largest uh instance was struggling to keep [4:50] up with the synchronization between all of the the the data traffic. Uh so the next step for us [4:58] was to move this to a new modern data varos platform and we moved to snowflake in 2021 [5:05] and that was a incredible amount of work for us preparing us for where we are today. Uh we built [5:12] new schemas for more or less everything. We had a long process with things running in parallel [5:18] but we finally just decided we need to cut this off and stop just focus on a new one. It was [5:24] rather painful but that was that worked for us. So in 2024 when the podcast happened we were finally [5:33] able to focus on AI but we hadn't been able to do so unless we had had this journey already then. So [5:42] how have we made this work for us throughout the 700 employees? [5:47] First of all, we have our Durac cell battery from the top. That's Nikolai. And in every day, [5:55] uh he is trying to focus on AI. He's bringing that a topic in every meeting he has. [6:00] But he also works around walks around the floors asking people what they're actually doing with AI. [6:06] And that's not just his uh calling a meeting. He's actually walking around pointing fingers [6:12] at people, sitting down with them and asking them to actually explain what they're doing. [6:17] And if you're not doing something with AI, you probably feel really really bad. But it's [6:23] doing this in a gentle way as well. So from the bottom up, people have been giving the tools, [6:28] the time, and the training to focus on their project. what can be what can AI be used for [6:35] in their daily job to make them more efficient and after we started using claude [6:42] we had claude project and at uh and at some stage we had more than 80,000 different project living [6:49] inside of claude many of them were of course just you know proof of concept people trying [6:54] out things but that was fine because it was a learning effect here as well [6:59] and from the side we created AI team. We have grown from uh three people to 15 at the moment [7:09] and we have also created a mandatory uh upskilling program for absolutely everyone in the company. [7:18] On the other side, [7:22] when we started this year uh this we created a year wheel where every event in the company [7:31] had AI on the topic. So it's a crowded slide but it just to show you that [7:38] if we were running an AI upscale, a tech day, we had AI. If you had a people day, [7:44] if you had a leader group meeting, everything uh uh consisted of something with AI on the agenda [7:52] every single time. But we also have this internal tech days in each office where people showcase [7:59] what they've been doing to the rest of the organization to have internal evangelism. [8:06] And as you also can see in the start here, we have AI ambassadors or AI champions. [8:12] We started out with 20 ambassadors from different parts of the company. We gave them [8:19] extra training together with uh or from Antropic and they were giving them one task. [8:24] Find a good use case in your team, implement it with the help of the other ambassadors, [8:31] the AI team and Entropic and then turn around, showcase it to the rest of the organization. [8:37] So we get some momentum and more internal evangelism here. [8:46] So internally our main AI tool is cloud. Everyone in the company is using [8:55] cloud uh chat and cloud co-work. We are measuring this and we have numbers for it. Everything half [9:03] of the employees are using cloud code. So even if you are a portfolio manager or a trader you are [9:09] likely doing some coding as well. And that keeps us a bit makes us a bit different from other uh [9:19] sovereign funds or other funds. We have also built an internal agent platform. [9:26] So if you need to have something running, you know, in the cloud separate from your computer, [9:32] you deploy it there. You need some skill to do that and you need to be able to code or get help [9:36] to do that. that separates the agent platform from the agent factory which is a self-service solution [9:43] for the other people or if you need an agent just running doing something for you then you [9:49] can do it there. We have created MCPs connecting I would say all of our internal data sources through [9:57] uh through the cloud here through the through to cloud uh claude and other uh agents as well [10:05] and we are not only using entropic but we're using models from is beneath the subtext here u openai [10:15] um meta mistral coher we are using the model which we think makes the most sense to the use case [10:22] even though it in our presentations it often looks like we are favoring a lot to cla so the upskills [10:30] upskill session they were mandatory for all but they were quick 30 minute sessions [10:36] and they are not that complicated it started out with AI and NBIM [10:41] meet claude prompting critical thinking uh governance and uh responsible AI [10:49] we kick that off in spring but after only six months we had to revisit the whole idea [10:57] because the tools the concept has changed so much it wasn't that important to thinking uh to [11:04] learning about uh prompting uh uh after 6 months so we focused more on the new tools like load code [11:12] how we have built agents and also evolutions evaluations of the agents with all of this focus [11:22] had one goal goal altogether. We wanted people to change their habits. We want wanted people to use [11:30] AI in all aspect of their daily work. We didn't want them just to use AI to summarize an email. [11:39] If that's if that was all you did, you didn't use it enough. So changing habit for people is really [11:48] hard and that's why we had this year the the reminders that's why our CEO is walking around [11:55] reminding people and we see now that people are using AI a lot but when they're using AI a lot [12:03] alone it doesn't mean they're using AI enough in the team process around them and that's where [12:10] we are seeing a bit of a cultural crash at the moment and it's even harder to change the culture [12:17] in a company than just individual habits. So, we are still pushing for that. We're not there yet. [12:25] Uh the AI team has focused on these four parallel tracks. Upskilling as I've mentioned, [12:32] we built the platform for everyone making it as easy as possible for everyone to deploy agents. [12:39] We have interviewed and had workshop with the chiefs and all the heads in the company [12:44] to identify which project have been or are the most valuable for us to implement. you uh [12:54] so we had the bottom up approach with everyone coming up with their own project but we also [12:58] have the top down approach which was we think are the most valuable and finally partnerships with u [13:07] AWS with Antropic and other peers we're trying to stay ahead of the curve when we're creating the [13:15] the the solutions here for for us we want to be more efficient We want to increase productivity [13:24] and year over year we have been uh 20% or more than 20% more efficient and with 700 employees [13:30] that's uh 300 billion region crowns or divide that by 10 or 30 it's it's not as much it's in the [13:38] project we really see the value here. So for us as a fund value also comes from increased returns. [13:46] So we are looking for anything that can help us make good decisions. we are doing kind of some [13:52] alpha generation as well. Uh and this uh both the automation the savings and the AI usage in [13:59] our investment proc processes are virt much more than the efficiency gain we have seen. [14:07] But when we're talking about increased returns, it's we cannot avoid reducing the risk [14:13] and making even more informed decisions by using AI to dig much more into the data than we were [14:21] able to before. For with 700 companies, thousand which we follow really closely, it takes a lot of [14:29] time to dig through everything that's happening. So one example on how we've changed how we work [14:37] we used to work with the old scrum methology when we're doing development [14:41] eight developers one portfolio manager or one project manager [14:47] working together the eight developers of the two pizza team were the bottleneck [14:54] they couldn't deliver code enough to complete the project fast enough etc etc now we see [15:00] that two developers and one from the business working together as a small autonomous team [15:08] which we trust to do the whole project are so much more efficient than the old methology [15:16] and we're struggling to keep it small and autonomous because as in every organization [15:21] there is bureaucracy and these has the tendency to blow up with more and more people because it's [15:28] would be nice to have someone more awards. So that's a cultural thing again and we have to [15:32] work really hard to keep it small and autonomous because that's where we see the efficiency again. [15:40] But on the governance side as well we hold the people accountable. We could run AI portfolios u [15:53] within hours if we wanted to but we want people to monitor everything that is happening. [16:00] We want this to be traceable. We want to know what the AI is actually doing and referring to [16:09] and we want to iterate and challenge the decision or the suggestions the AI are giving us. [16:16] So one example is how do we screen all of these 7,000 companies? We have policies [16:24] or mandates from the government saying that we can and cannot be invested in in certain areas. [16:30] And how do we know that one of these companies hasn't changed what they're doing and are [16:36] let's say uh started producing weapons which is off limits for us. So we take all our vendor data, [16:44] we take everything that's in the news picture, financial report, covenant filings, local media, [16:50] etc., etc., and all our internal data as well. All our yearly meetings with the company [16:58] and we give that to an agent. So at any given time the agent can shoot this and it also [17:06] do it on it does it on the schedule. There are not just one agent there [17:10] are swarm of them doing this for us. So it's available for uh internally all the time. So [17:20] that setup together with what we call a company meeting assistant and simulator helps us have [17:29] more than 3,200 company meetings each year. So that's only the companies we follow more closely. [17:40] Before this, it took between three and four hours [17:44] to prepare for such a meeting for a research analyst and this is an example with meta. [17:52] So we search for the company we are having a meeting with. Uh we select it and we can see [18:00] then okay this is what we did last time. These are resources available and we can then get a draft [18:09] agenda from it. And is in this draft agenda, it's not as easy here, but you can see some green and [18:16] yellow markers in them here as reference point to where the AI is actually digging this information [18:22] out to. And the point here is that there is also a chat on the side here where the the analyst can [18:29] iterate through the agenda change it as you would you know can assume you can use any AI to change [18:38] a working document as this. Once the research analyst is happy with the [18:46] the work from the AI here, the simulation part kicks in and the AI can pretend to be the company [18:53] the analysts are meeting and try and answer as they think it the the company would answer to it. [18:59] So the analyst is even more prepared when they're going to that meeting. [19:05] So these are then three examples of how we have implemented AI and we still use uh still working [19:13] on implementing more and more sources into this and making it as widely uh available in comp in [19:19] internally as as possible. Um and I think that's all I have time for. So if you have any questions [19:28] please try and find me afterwards. I'll be around. Thank you very much. [applause]