CIBC Ditches Microsoft Copilot to Build In-House AI for 45,000 Staff

Inside CIBC’s GenAI Transformation: From 500 Pioneers to 45,000 Empowered Employees With CIBC — Ai4

Ai4

Building proprietary generative AI delivers major cost savings over vendor software, but capturing productivity gains depends on driving grassroots behavior change rather than shipping technical features.

AI-assisted summary of Inside CIBC’s GenAI Transformation: From 500 Pioneers to 45,000 Empowered Employees With CIBC — original source. Check the original for full context.

Source published: 2025-08-27

Detailed notes

## Background Published on YouTube by Ai4 on August 27, 2025, this presentation features Chris Patterson, Head of Enterprise AI Platforms and Solutions at CIBC. The session details CIBC's year-in-review journey of scaling its internal generative AI platform from an initial 500-person pioneer pilot to over 45,000 empowered employees, highlighting user engagement, champion-led AI literacy, and their strategic roadmap toward Agentic AI.

## TL;DR Rejecting commercial off-the-shelf tools like Microsoft Copilot, CIBC engineered an in-house generative AI platform equipped with tailored workflow buttons, inline diagramming, self-serve RAG, and an internal agent marketplace. Platform access is governed by mandatory compliance training, with granular telemetry tracking usage across five distinct fluency tiers where expert users report saving approximately 20 hours per month. Patterson emphasizes that while token and infrastructure costs are trivial, the primary hurdle in enterprise GenAI is human change management rather than technical development. Sustainable scaling ultimately relied on shifting from centralized instruction to decentralized, grassroots community enablement that cemented AI as an indispensable daily workflow tool.

## Atmosphere & style Patterson delivers a candid, pragmatic, and highly engaging keynote that blends the scrappy mindset of an internal tech startup with the governance realities of a major financial institution. His presentation style is conversational and refreshingly transparent, using humor and relatable anecdotes—such as non-technical lawyers writing Python scripts and employees refusing to work during platform outages—to demystify enterprise AI. The talk carries a dynamic, forward-looking energy that balances hard-won operational lessons with clear pride in achieving enterprise-wide cultural transformation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) CIBC bypassed off-the-shelf solutions like Microsoft Copilot to build an in-house enterprise AI platform, giving the bank full engineering control and custom workflow integration. - [01:53](https://www.youtube.com/watch?v=t_26JjS6C6w&t=113s) To bridge the AI literacy gap, the interface incorporated pre-built workflow buttons for summarization, email drafting, and coding rather than relying on raw prompting. - [04:02](https://www.youtube.com/watch?v=t_26JjS6C6w&t=242s) In-house ownership enabled rapid feature deployment within days, including automated Mermaid.js diagram generation and inline HTML prototyping for product teams. - [05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) CIBC's strategic roadmap prioritizes secure in-house meeting notetakers, project-level collaborative workspaces, real-time advisor audio assistants, and agentic workflows with strict spend guardrails. - [08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) Self-serve RAG removes engineering bottlenecks by letting business units auto-index data sources like SharePoint, supported by an internal storefront that compensates employees for sharing custom recipes. - [09:41](https://www.youtube.com/watch?v=t_26JjS6C6w&t=581s) Enterprise access is gated by a mandatory 1-hour compliance and literacy course, with granular telemetry tracking interaction depth to calculate productivity lift per team. - [11:55](https://www.youtube.com/watch?v=t_26JjS6C6w&t=715s) Employees progress across a five-tier AI fluency framework over 3–4 months, with expert users validating an average time savings of 20 hours per month. - [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) LLM operational token costs are minimal relative to enterprise scale, and CIBC utilizes a multi-model gateway integrating OpenAI and Google for redundancy and cost optimization. - [16:29](https://www.youtube.com/watch?v=t_26JjS6C6w&t=989s) Human adoption and behavioral change management represent the hardest challenge in enterprise AI transformation, outweighing pure technical feature development. - [17:05](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1025s) Scaling to 45,000 employees required shifting from central platform training to a decentralized model of business-led working groups and peer-to-peer coaching circles. - [19:01](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1141s) Demonstrating the power of enterprise RAG, CIBC AI used internal historical presentation decks to answer 9 of 12 criteria questions for a winning industry award submission.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) Building an In-House AI Startup at CIBC

Chris Patterson introduces himself as the Head of Enterprise AI Platforms and Solutions at CIBC, framing his team's mandate as operating like a "micro AI startup within the bank." Given a modest amount of initial funding, the team was tasked with experimenting freely with large language models to explore what was possible, operating in a fast-moving, highly innovative environment focused on driving enterprise efficiency and tangible value through Generative AI. Reflecting on his presentation from the previous year, Patterson recalls that CIBC had spent the better part of a year testing Microsoft Copilot. However, he remained unimpressed with the product, feeling unable to justify the business value behind what it offered.

Determined to find a better alternative, Patterson approached executive leadership to pitch building a bespoke internal platform. He told the decision-makers, "give me some funding, let me build our own thing... I'm sure I could do something better than Copilot." While the leadership initially laughed at the audacity of the proposal, they granted the budget with the directive to "have at it." CIBC launched its internal tool just one week prior to the previous year's conference. Over the subsequent 12 months, the platform evolved from an initial 500-person pioneer pilot to scaling across tens of thousands of employees—a transition Patterson notes is far more grueling than spinning up a 10- or 50-person prototype due to the immense security, safety, and governance hurdles inherent in a major financial institution.

- **Initial Scope:** Piloted initially with 500 users before scaling toward an enterprise rollout targeting 40,000+ employees. - **Strategic Motivation:** Rejecting off-the-shelf Copilot solutions to retain full engineering control, build customized features, and deliver measurable ROI.

---

[01:53](https://www.youtube.com/watch?v=t_26JjS6C6w&t=113s) Designing CIBC AI: Workflow Buttons and Surprising Adopters

Patterson candidly describes the underlying interface as essentially a "clone of ChatGPT," admitting he was not aiming to reinvent the wheel from a pure product design standpoint. Instead, the team differentiated the tool by outfitting it with custom buttons and CIBC-tailored workflows designed to guide employees toward predictable, high-value outcomes. The engineering team had originally operated under the assumption that most bank employees were already well-versed in prompt engineering and large language models; in reality, a vast portion of the workforce had never interacted with ChatGPT before launch. To bridge this AI literacy gap, the interface introduced explicit workflow buttons such as document summarization, drafting emails, and answering customer correspondence.

The platform was quickly expanded to include coding assistance features. While the team expected developer teams to be the primary users of the coding module, adoption spread to unexpected corners of the enterprise. Patterson shares an anecdote of discovering a bank lawyer using the platform to generate complex Python scripts intended to scrape and parse his email inbox. When the lawyer encountered execution errors and reached out for help, it turned out he lacked local administrator rights and the required Python runtime libraries on his corporate laptop. Patterson highlights this as proof that LLMs democratize access to technical capabilities for non-technical staff across the enterprise.

```mermaid flowchart TD accTitle: CIBC AI Core Ecosystem and Integrations subgraph EnterpriseData["Enterprise Context & Data Layer"] M365["Inbox & Outlook Calendar"] Docs["Internal Documents & SharePoint"] Policies["CIBC Policies & Procedures (RAG)"] end

subgraph PlatformCore["CIBC AI Platform Engine"] LLM["Foundation LLMs"] PromptEngine["Tailored Prompt & Workflow Logic"] end

subgraph GuidedWorkflows["Guided UI & Workflows"] BtnSumm["Document Summarizer"] BtnMail["Email Drafting & Response"] BtnCode["Software & Code Generation"] BtnVisual["Visual Artifacts (Mermaid / HTML)"] end

EnterpriseData --> PlatformCore PlatformCore --> GuidedWorkflows ```

- **Core Grounding Integrations:** Seamless connectivity to user inboxes, calendars, internal documentation, and policy repositories via self-service Retrieval-Augmented Generation (RAG). - **Behavioral Onboarding:** Action-oriented buttons designed to prompt employees on specific tasks without requiring advanced prompting skills.

---

[04:02](https://www.youtube.com/watch?v=t_26JjS6C6w&t=242s) Rapid Feature Rollouts: Mermaid Diagrams and HTML Prototyping

Because CIBC owns and operates its underlying platform rather than relying on a rigid third-party SaaS vendor, Patterson's team can design, test, and ship custom features within days. One major capability introduced is native diagram generation: users can prompt the model to generate Mermaid syntax on the fly, rendering process flowcharts, Gantt charts, and architecture diagrams instantly. After a tech team member demonstrated the proof-of-concept, it was immediately released to production, rapidly accumulating thousands of active users who generate diagrams in seconds without dedicated visualization software.

Patterson details another rapid deployment that originated the Friday before the conference. During conversations with internal product managers, several staff voiced frustration over the communication friction between business teams and engineering, noting that developers "don't always hear me, they don't always understand what I want." Recognizing how easily LLMs handle front-end code, Patterson's team implemented a "Prototype Builder" button that generates raw HTML and renders working visual mockups directly inside the chat interface. Launched over the weekend, the feature immediately generated unsolicited positive feedback from internal teams by Tuesday morning, demonstrating the agility of maintaining an in-house enterprise AI stack.

- **Diagramming Support:** Automated generation and rendering of Mermaid.js diagrams, Gantt charts, and operational process flows. - **Prototype Builder:** Instant HTML generation and inline rendering allowing product managers to visually communicate requirements to engineering teams.

---

[05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) The Product Roadmap: In-House Notetaking, Collaborative Workspaces, and Agentic AI

Looking forward, CIBC's AI roadmap focuses on four primary pillars: meeting intelligence, collaborative workspaces, real-time audio, and agentic workflows. Addressing meeting notes, Patterson emphasizes that the bank cannot use popular commercial tools like Fathom due to data exfiltration and model-training security concerns. In response, CIBC built its own internal notetaker bot, scheduled for broader enterprise deployment in September. Patterson emphasizes that eliminating manual, handwritten meeting minutes will be life-changing for project managers, product leads, and governance committees, with past transcripts indexed via quasi-RAG to serve as searchable institutional memory.

```mermaid flowchart LR accTitle: CIBC AI Strategic Capabilities Roadmap subgraph NearTerm["Production & Near-Term"] direction TB N1["In-House Notetaker Bot<br/>(September Launch)"] N2["Quasi-RAG Project Notebooks<br/>(Jira / Confluence / SharePoint)"] end

subgraph Horizon["Emerging Capabilities"] direction TB H1["Real-Time Audio Assistant<br/>(Live Advisor Data Fetching)"] H2["Agentic AI Workflows<br/>(Guardrails & Cost Governance)"] end

NearTerm --> Horizon ```

Beyond notetaking, the platform is expanding into collaborative team workspaces—a concept inspired by Google NotebookLM—allowing cross-functional squads to ground an AI assistant in shared artifacts from Jira, Confluence, and SharePoint across an entire sprint or deliverable. The team is also exploring low-latency audio capabilities where an AI acts as a real-time "junior assistant" during client-advisor calls, proactively fetching backend data the moment a client asks a question. Regarding agentic AI, Patterson acknowledges high demand across leadership but urges caution: while internal proofs-of-concept are active, productionizing autonomous agents requires solving critical reliability and guardrail challenges, such as preventing runaway agents from "going rogue and spending 50 grand on a task that was probably only really should have been five bucks."

- **Enterprise Notetaker:** Secure, internally hosted meeting bot launching bank-wide in September, fully searchable via RAG. - **Collaborative Workspaces ("Notebooks"):** Project-level grounding connected to Jira, Confluence, and SharePoint to support team-based delivery sprints. - **In-Meeting Audio Copilot:** Real-time speech monitoring capable of instantly fetching client and product data during live advisor discussions. - **Agentic AI Governance:** Rigorous testing on output accuracy, deterministic behavior, and financial spend guardrails prior to broad scaling.

[08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) Feature Usage, Self-Serve RAG, and the Agent Marketplace

Employees across CIBC have moved well beyond basic prompts. While common use cases like text summarization and email drafting remain prevalent, teams are now actively producing custom information diagrams and relying heavily on self-serve Retrieval-Augmented Generation (RAG). With self-serve RAG, technical teams no longer act as gatekeepers or bottlenecks: the system automatically scrapes, auto-indexes, and connects directly to sources such as SharePoint so business units can launch their own retrieval pipelines autonomously.

```mermaid flowchart LR accTitle: Self-Serve RAG and Agent Marketplace Flow subgraph Data ["Data & Self-Serve RAG"] A["SharePoint / Documents"] --> B["Auto-Index & Scraping Engine"] B --> C["User RAG Pipeline"] end subgraph Marketplace ["Workflow Marketplace"] C --> D["Custom Agent / Recipe Creation"] D --> E["Storefront / Marketplace"] E --> F["Enterprise-Wide Sharing & Adoption"] end ```

To harness user ingenuity, CIBC is launching an internal agent marketplace. Although Patterson notes his personal hesitation around overusing the buzzword "agent," the company has rebranded custom workflows under that term to align with industry nomenclature. The storefront allows employees who design high-impact workflows or recipes to publish and share them across the entire bank. To supercharge participation, CIBC has even introduced a compensation scheme:

> "We're going to create a storefront where you can actually share this and I'll even pay them. We have a scheme where I can compensate people for sharing this stuff. Why not? Why not? I want to drive up adoption."

- Summarization and email drafting are standard baselines, with dynamic diagram creation emerging as a popular new visual feature. - Self-serve RAG removes the technology team from the critical path by automatically indexing document repositories like SharePoint. - An internal marketplace storefront allows creators to share agentic recipes, with financial and formal incentives provided to employees who contribute valuable workflows.

---

[09:41](https://www.youtube.com/watch?v=t_26JjS6C6w&t=581s) Adoption Velocity and Granular Productivity Telemetry

Since officially launching at the end of May, the enterprise platform has seen rapid adoption. Daily active usage stands at roughly 7,500 employees, with total platform registrations climbing from 24,000 to around 26,000, on track toward a broader goal of 40,000 to 45,000 employees. To balance speed with bank governance, CIBC enforces a strict prerequisite: every employee must complete a mandatory one-hour training course covering AI fundamentals and risk management before platform access is provisioned.

```mermaid flowchart TD accTitle: User Onboarding and Telemetry Analysis subgraph Onboarding ["Governance & Access Gate"] A["1-Hour Mandatory AI Training"] --> B["Platform Access Provisioned"] end subgraph Telemetry ["Exhaustive Data Collection"] B --> C["Capture Clicks, Chats, Prompts & Depth"] C --> D["Productivity Calculation Models"] end subgraph Reporting ["Managerial Insights"] D --> E["Daily Employee-Level Reports"] E --> F["Identify Innovators vs. Laggards"] F --> G["Mobilize Change Champions"] end ```

To quantify the platform's return on investment, Patterson's team captures comprehensive granular telemetry on every user interaction—logging button clicks, chat length, conversational depth, prompt patterns, and specific recipes or agents executed. These data points feed mathematical models that estimate productivity lift down to the individual employee and daily level. Managers receive tailored insight reports that benchmark team progress, pinpoint adoption laggards, and highlight key innovators who can be recruited into a change champion network.

- **Active Users:** ~7,500 daily active users; over 26,000 employees onboarded since the end-of-May rollout. - **Access Gating:** A mandatory 1-hour compliance and literacy training gate ensures safe usage within a regulated banking environment. - **Target Reach:** Active trajectory targeting 40,000+ empowered enterprise employees. - **Granular Telemetry:** Captures full interaction depth (clicks, prompts, conversation flows) to calculate mathematical productivity lift and generate team-level managerial scorecards.

---

[11:55](https://www.youtube.com/watch?v=t_26JjS6C6w&t=715s) AI Fluency Progression and Behavior Transformation

CIBC evaluates employee capability using a five-tier AI fluency framework ranging from novices to advanced practitioners:

Progression across these tiers typically takes 3 to 4 months of sustained usage. Transitioning from novice to expert is not automatic; it requires fundamental behavioral change. To support varied learning styles across multi-generational demographics—from Gen Alpha onward—CIBC deployed a diverse enablement strategy. A dedicated full-time trainer leads weekly Ask-Me-Anything (AMA) open forums that train 200 to 300 employees per session, supplemented by enterprise Workplace discussion channels and business-line champions. Over half the company has already completed voluntary upskilling courses.

> "Experts [save] 20 hours a month... When folks really use this product, they're getting some real value from it. I have spoken to hundreds of our experts to validate this claim... I want to know what are you doing? How are you using this product? Show me."

---

[15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) Platform Economics and LLM Infrastructure Strategy

The infrastructure and operational token costs of running enterprise generative AI are described as "peanuts" and a mere "rounding error" relative to the scale of CIBC's operations. Token pricing has dropped exponentially while model reasoning capabilities have surged. When CIBC initially launched using OpenAI's GPT-3.5 Turbo, the per-token cost was significantly higher than advanced reasoning models like o3 are today.

```mermaid flowchart TD accTitle: Multi-Model Gateway and Evaluation Architecture subgraph Providers ["Model Providers"] M1["OpenAI (Enterprise Endpoint)"] M2["Google / Gemini (In Onboarding)"] end subgraph EnterpriseGateway ["CIBC Secure Platform Gateway"] M1 --> GW["Security, Compliance & Routing Layer"] M2 --> GW end subgraph Evaluation ["Evaluation & Staging"] GW --> T1["Dev Sandbox Testing (e.g., GPT-5 / Frontier)"] T1 --> T2["Production Certification"] T2 --> T3["45,000 Enterprise Employees"] end ```

CIBC continuously tests frontier systems in internal development sandboxes—including early testing with next-generation iterations such as GPT-5—though Patterson expresses skepticism over early external benchmarks, suspecting some releases function as model-routed wrappers over existing foundations like o3 and GPT-4.1. While current enterprise-grade security pipelines run primarily through OpenAI models, CIBC is actively partnering with Google to onboard additional frontier architectures to ensure provider redundancy, flexibility, and cost optimization.

---

[16:29](https://www.youtube.com/watch?v=t_26JjS6C6w&t=989s) Lessons Learned: Centering Humans Over Technical Features

Reflecting on the rollout, Patterson highlights that human adoption, rather than software development, represents the hardest challenge in generative AI transformation. Engineering teams naturally lean toward shipping advanced technical features, but shipping tools in isolation yields minimal return if employees do not evolve their daily routines.

> "Adoption is hard. This is where I spend almost all my time. I would much rather be innovative. I would much rather be building cool features, cool product ideas. And I did focus a lot on, you know, the next best feature, the next best product. And I realized, you know, fairly early on, this is a waste of time. If I don't bring the people along with me, they won't use it anyway."

Sustainable platform expansion demands spending disproportionate time understanding specific employee workflows, eliminating behavioral friction, and continuously aligning platform development with real-world end-user needs.

[17:05](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1025s) Enterprise Adoption, Cultural Shift, and Community-Driven AI Literacy

The training programs and enablement strategies have driven a material lift across CIBC, moving staff steadily toward the "expert" proficiency tier. The genuine excitement and reliance on CIBC AI across the workforce is unmistakable. Whenever the platform experiences even a brief one-hour outage, inboxes flood with urgent inquiries from employees whose daily workflows have fundamentally transformed. In fact, some staff have reached the point where they refuse to perform certain tasks manually, preferring to wait until the AI is back online to complete them—a dynamic the team welcomes as proof of genuine utility.

Recognizing that a central team cannot individually train 45,000 employees, CIBC pivoted to a decentralized, community-driven model where business units lead their own enablement. Teams have organically established working groups, peer-learning circles, and impromptu sharing sessions to exchange effective prompting techniques and showcase how models drive operational efficiency. Successfully launching and scaling an internal generative AI product across an enterprise of this size is a major milestone, particularly within a strict regulatory environment that initially responded to ChatGPT by blocking it entirely. Compared to expensive third-party vendor platforms, building this in-house capability has delivered substantial productivity gains at a fraction of the cost.

```mermaid flowchart TD accTitle: CIBC Community-Driven AI Scaling Model subgraph Central["Central Enablement"] A["Core AI Platform Team"] --> B["Enterprise Frameworks & Tooling"] end subgraph Business["Business-Led Communities"] B --> C["Department Working Groups"] C --> D["Peer-to-Peer Training & Office Hours"] C --> E["Use-Case Discovery & Sharing"] end subgraph Outcomes["Enterprise Outcomes"] D --> F["Material Lift in AI Literacy"] E --> G["Grounded Productivity Gains"] end ```

- **Operational Dependency:** Brief system downtime generates immediate employee feedback, demonstrating that generative AI has become an indispensable daily productivity tool. - **Grassroots Enablement:** Business-led working groups and informal peer-coaching networks sustain enterprise-wide literacy without overburdening central platform teams. - **Cost Efficiency:** Developing an internal, grounded solution delivered enterprise-grade capabilities at a trivial cost relative to external commercial vendor tools.

---

[19:01](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1141s) The Agentic Roadmap and Winning an Award with CIBC AI

Looking to the future, CIBC is charting a path toward Agentic AI, with plans to return to the stage next year to share progress on autonomous workflows. The presentation concluded with a celebration of industry recognition: CIBC has now won the award for Best Generative AI Initiative twice.

The story behind the latest submission highlights the practical power of grounded AI. Tasked on the final day of the deadline with submitting the award application, the team turned to CIBC AI's "Smart Library"—the platform's grounded Retrieval-Augmented Generation (RAG) feature. By uploading a year’s worth of presentation decks and the 12 evaluation criteria questions, the internal AI successfully generated high-quality, fully grounded responses for 9 of the 12 questions. The remaining 3 were completed manually where source data was ungrounded. The application went on to win the award, proving that when large language models are provided with the right context and enterprise data grounding, they can deliver exceptional real-world results—literally writing their own winning submission.

- **Agentic AI Roadmap:** CIBC is actively developing autonomous agent architectures to solve higher-order enterprise tasks. - **Two-Time Award Winner:** CIBC has won the Best GenAI Initiative award twice for its internal enterprise deployment. - **RAG in Practice:** Using Smart Library, CIBC AI answered 9 out of 12 complex award evaluation questions directly from historical PowerPoint presentations.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) give me some funding, let me build our own thing... I'm sure I could do something better than Copilot. — Chris Patterson

> [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) micro AI startup within the bank — Chris Patterson

> [01:53](https://www.youtube.com/watch?v=t_26JjS6C6w&t=113s) clone of ChatGPT — Chris Patterson

> [04:02](https://www.youtube.com/watch?v=t_26JjS6C6w&t=242s) don't always hear me, they don't always understand what I want.

> [05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) going rogue and spending 50 grand on a task that was probably only really should have been five bucks. — Chris Patterson

> [08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) We're going to create a storefront where you can actually share this and I'll even pay them. We have a scheme where I can compensate people for sharing this stuff. Why not? Why not? I want to drive up adoption. — Chris Patterson

> [11:55](https://www.youtube.com/watch?v=t_26JjS6C6w&t=715s) Experts [save] 20 hours a month... When folks really use this product, they're getting some real value from it. I have spoken to hundreds of our experts to validate this claim... I want to know what are you doing? How are you using this product? Show me. — Chris Patterson

> [16:29](https://www.youtube.com/watch?v=t_26JjS6C6w&t=989s) Adoption is hard. This is where I spend almost all my time. I would much rather be innovative. I would much rather be building cool features, cool product ideas. And I did focus a lot on, you know, the next best feature, the next best product. And I realized, you know, fairly early on, this is a waste of time. If I don't bring the people along with me, they won't use it anyway. — Chris Patterson

## Mentioned Chris Patterson, CIBC, CIBC AI, Microsoft Copilot, ChatGPT, OpenAI, GPT-3.5 Turbo, GPT-4.1, GPT-5, o3, Google, Gemini, Google NotebookLM, Fathom, Mermaid.js, Jira, Confluence, SharePoint, Outlook, Smart Library

## Video description Chris Patterson, Head of Enterprise AI Platforms and Solutions, CIBC

Abstracts

Join me at my talk for a year-in-review of CIBC’s Generative AI journey. Last summer, we introduced our AI product CIBC AI to 500 employees; today we’ve scaled its reach to over 45,000, reshaping enterprise-wide productivity. Discover the strategic milestones that paved our path—spanning user engagement, champion-led & training backed AI literacy, and our roadmap toward Agentic AI. Walk away with real-world insights, lessons learned, and a first-hand look at the transformative potential of generative AI at scale.

[Source](https://www.youtube.com/watch?v=t_26JjS6C6w)

中文简报

CIBC弃用微软Copilot，为4.5万名员工自研AI平台相比采购供应商软件，自建专有生成式AI能大幅节省成本，但要实现生产力提升，关键在于推动基层行为转变，而非单纯交付技术功能。 CIBC放弃了微软Copilot等商业现成工具，自主研发了一套企业生成式AI平台，配备定制化工作流按钮、内联图表绘制、自助式RAG以及内部Agent市场。该平台的访问权限由强制合规培训管控，并通过细粒度遥测技术追踪五个不同熟练度层级的使用情况，其中专家级用户报告每月可节省约20小时。Patterson强调，尽管Token和基础设施成本微不足道，但企业GenAI面临的主要障碍在于人员变革管理，而非技术开发。实现可持续的规模化落地，最终依赖于从集中式指导转向去中心化的基层社群赋能，从而将AI巩固为日常工作流中不可或缺的工具。 CIBC 放弃了 Microsoft Copilot 等现成方案，自主构建内部企业 AI 平台 CIBC AI，从最初 500 人试点稳步扩展至覆盖 45,000 名员工，并通过预设工作流按钮降低全员使用门槛。 (0:04) 自主掌控技术栈使团队能在数天内快速上线 Mermaid.js 图表生成和 HTML 原型渲染等定制功能；未来路线图重点涵盖内部会议记录机器人与具备预算防失控机制的 Agentic 工作流。 (4:02) 平台推出自助式 RAG 自动对接索引 SharePoint 等数据源以消除技术团队的对接瓶颈，并通过内部工作流商店对分享实用 Agent 配方的员工给予奖励。 (8:28) 企业要求全员必须通过 1 小时的合规培训方能开通权限；通过精细化遥测数据追踪，在 3 至 4 个月内达到专家级熟练度的员工每月平均可节省 20 小时工时。 (11:55)

Reading summary

After finding commercial solutions like Microsoft Copilot insufficient to justify their business value, CIBC developed an in-house generative AI platform that expanded from an initial 500-user pilot to more than 26,000 employees. Chris Patterson, Head of Enterprise AI Platforms and Solutions at CIBC, argues that enterprise value depends on guided workflow design, community-driven enablement, and granular productivity telemetry rather than standalone feature development. In-House Development Over Off-the-Shelf Tools (0:04) Rather than relying on external software like Microsoft Copilot, CIBC funded an internal team operating like a micro-startup within the bank to build a tailored platform. While the user interface mirrored ChatGPT, the team differentiated the system by addressing an enterprise AI literacy gap with structured workflow buttons for document summarization, email drafting, and coding assistance. Owning the underlying software stack enabled the team to rapidly deploy internal requests within days, including Mermaid.js diagram generation and an inline HTML Prototype Builder that product managers use to visually communicate specifications directly to engineering teams. Mandatory Enablement and Granular Productivity Telemetry (9:41) To balance rapid rollout with banking regulations, CIBC mandates a one-hour compliance and AI fundamentals training course before provisioning user access. Since its launch at the end of May, the platform reached roughly 7,500 daily active users and approximately 26,000 onboarded staff, tracking toward a goal of 40,000 to 45,000 employees. To measure return on investment, the bank captures exhaustive interaction telemetry—logging prompt patterns, conversation depth, and button usage—to model individual and team-level productivity gains. CIBC evaluates employee progress across a five-tier fluency scale ranging from Novice to Professor. Patterson noted that progressing to higher tiers requires sustained behavioral shifts over three to four months, with verified Expert users saving an estimated 20 hours per month. Grassroots Upskilling and an Incentivized Workflow Marketplace (8:28) Because a centralized team cannot independently support tens of thousands of users, CIBC shifted to a decentralized, business-led adoption model. Departments established peer-coaching networks, discussion channels, and weekly open forums alongside dedicated training sessions. On the engineering side, self-serve Retrieval-Augmented Generation (RAG) was integrated to automatically scrape and index sources like SharePoint, eliminating technical bottlenecks for departmental pipelines. To encourage internal innovation, the bank is launching an enterprise marketplace where employees can publish custom workflow recipes. CIBC even introduced compensation incentives to pay staff who create widely adopted agentic workflows. Demonstrating the efficacy of this grounding architecture, the team used the platform's internal Smart Library RAG feature to answer 9 of 12 questions on historical slide decks for a submission that won CIBC a Best GenAI Initiative award. Infrastructure Economics and Agentic Governance Constraints (5:20) Patterson characterized operational token expenses as a rounding error relative to the bank's operational scale, noting that declining token prices have coincided with higher model reasoning capabilities. The system operates primarily on enterprise OpenAI endpoints and is onboarding Google Gemini models to establish redundancy and lower costs. CIBC's product roadmap includes an internally hosted meeting notetaker bot to prevent data exfiltration, project notebooks grounded in Jira and SharePoint, and real-time audio copilots for customer advisors. However, Patterson urged caution regarding fully autonomous agentic workflows, noting that internal rollouts require strict financial and deterministic guardrails to prevent unmonitored agents from incurring thousands of dollars in unnecessary compute expenses. CIBC's experience indicates that an in-house generative AI platform can achieve broad enterprise adoption and cost efficiency within a regulated environment if supported by guided UI and decentralized enablement. However, moving beyond assisted workflows to autonomous agents remains constrained by unresolved safety, accuracy, and financial spend guardrails.

精炼阅读摘要

加拿大帝国商业银行（CIBC）企业人工智能平台主管 Chris Patterson 表示，该行放弃了现成软件方案，通过在内部构建受控的生成式 AI 平台，在 12 个月内从 500 人试点扩展至数万名员工，以更低成本驱动了业务自动化与效率提升。自建平台替代商业套件的战略抉择 (0:04) CIBC 在经过近一年的测试后对 Microsoft Copilot 并不满意，认为其难以证明预期的商业价值。为此，Patterson 带领团队在银行内部以类似初创公司的方式运作，争取到预算自主研发了一套生成式 AI 平台。与直接采购第三方 SaaS 产品相比，自建架构让技术团队拥有完全的工程控制权与定制灵活性，能迅速应对金融机构严格的安全性、合规性与治理审查。平台最初仅面向 500 名先锋用户进行概念验证，并在上线后的一年内迅速规划向全行 40,000 多名员工全面推广。定制工作流与快速敏捷的功能迭代 (1:53) 为了消除员工对提示词工程的认知门槛，平台在类 ChatGPT 的界面中引入了预设的工作流按钮，覆盖文档摘要、邮件起草、代码编写等具体场景。非技术人员的技术门槛被大幅降低，甚至有法务人员利用平台生成复杂的 Python 脚本来抓取并解析邮件。由于底层技术完全自研，团队能够以极短周期上线新特性，例如在几天内推出 Mermaid.js 图表即时生成工具，以及专为产品经理设计的“原型构建器（Prototype Builder）”，可在聊天窗口内直接生成并渲染可交互的 HTML 界面，从而缓解业务端与技术端的沟通阻力。自助式 RAG 与内部智能体市场激励 (8:28) CIBC 的应用场景已从基础文本处理扩展到深度业务数据联动。平台推出了自助式检索增强生成（RAG）管道，无需技术团队作为中介，系统可自动抓取并索引 SharePoint 等内部知识库。此外，银行正在建立内部智能体市场（Storefront），员工可以将自己设计的定制工作流或食谱向全行发布并共享。为了推动采纳率，CIBC 还制定了专门的激励计划，对贡献出高价值工作流的员工提供实际报酬。而在团队申报行业奖项时，平台借助 Smart Library 功能自动调取一整年的演示文稿，独立完成了 12 项评选标准中 9 项的高质量作答并最终获奖。量化遥测追踪与分级行为转型 (9:41) 自 5 月底正式发布以来，平台日活跃用户约 7,500 人，注册量已超过 26,000 人，并向 40,000 至 45,000 人的目标迈进。所有员工必须通过 1 小时的人工智能合规与安全培训方可获取权限。团队通过细粒度遥测技术记录每次点击、对话深度与提示词模式，利用数学模型推算个人层面的生产力提升，并向管理层提供日报以识别创新者与滞后者。银行建立了从“新手”到“教授”的五级熟练度体系，数据显示“专家”级员工每月可节省约 20 小时；为推动持续转型，CIBC 辅以每周数百人参与的 AMA 培训，并依靠业务部门内部的工作小组开展去中心化的同伴互助。基础设施成本与代理式 AI 演进路线 (5:20) 在底层技术层面，Patterson 指出运行大模型的 Token 与基础设施成本在银行的总体业务体量下几乎可以忽略不计。CIBC 在沙箱环境中持续评估前沿模型，并在 OpenAI 端点之外积极引入 Google Gemini 等提供商，以确保冗余度与架构弹性。平台未来的路线图聚焦于四个方向：计划于 9 月全行上线、具备 RAG 检索能力的自研会议记录机器人；整合 Jira 和 Confluence 的协作工作空间；面向客户顾问的实时低延迟音频辅助；以及自主智能体（Agentic AI）。但 Patterson 提醒，代理式 AI 必须先解决失控支出与确定性防护栏等治理挑战，以避免单次简单任务产生高额异常开销。 Patterson 总结认为，企业生成式 AI 转型中最艰巨的挑战是促成组织与人员的日常行为改变，而非单纯堆砌软件功能。尽管 CIBC 自建平台获得了显著的使用率和生产力提升，但如何平衡智能体自主化运作与财务风险防范，仍是全面扩展前需审慎解决的核心议题。

Detailed notes — zh

## Background 该内容由 YouTube 频道 Ai4 于 2025 年 8 月 27 日发布。视频为加拿大帝国商业银行（CIBC）企业级 AI 平台与解决方案负责人 Chris Patterson 的主题演讲，回顾了 CIBC 的生成式 AI 转型历程，展示其产品如何从 500 名试点员工扩展到覆盖全行 45,000 名员工，并深入剖析了用户参与策略、AI 素养培训及 Agentic AI 演进路线。 ## TL;DR CIBC 团队以内部“微型初创团队”模式放弃了外部商业套件，自主研发并上线了支持 45,000 名员工的企业级 GenAI 平台。平台通过预设引导按钮、自助式 RAG 知识检索与原型构建器降低使用门槛，并基于全量交互埋点构建了精细化的员工 AI 熟练度量化模型。面对金融强监管环境，CIBC 依托社区驱动的培训网络与 AMA 机制攻克组织行为转变的难点，以极高的成本效益实现生产力重塑，目前正加速向实时会议副驾与 Agentic AI 智能体演进。 ## Atmosphere & style 演讲风格坦诚务实、生动接地气，洋溢着大企业内部敏捷创新的极客热情与创业活力。主讲人毫不避讳地分享了自研决策中的质疑与挑战，并穿插了诸如非技术律师写脚本、短暂停机引发全员催更、以及利用 RAG 临时突击“让 AI 为自己赢下大奖”等诙谐真实的案例。整体氛围既有攻克传统银行合规壁垒的成就感，更充满了对“技术采纳核心在于人而非工具”的清醒认知与真诚反思。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) 组织机制与自研路线：CIBC 组建内部敏捷“微型 AI 初创团队”，因外部商业方案（如 Copilot）ROI 论证不足而选择自主研发，一年内从 500 人原型扩展至服务 45,000 名员工。 - [01:53](https://www.youtube.com/watch?v=t_26JjS6C6w&t=113s) 引导式 UI 破除使用门槛：通过内置标准化操作按钮（如文档摘要、邮件起草、Mermaid 流程图生成及 HTML 原型即时渲染），固化高频工作流，让法务等非技术人员也能触达自动化能力。 - [05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) 受监管金融环境下的自研工具栈：为避免敏感数据泄露给外部厂商，自研会议记录助手（Notetaker Bot），并借鉴 NotebookLM 理念打造打通 Jira、Confluence 的项目级协同工作空间。 - [05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) Agentic AI 落地核心挑战：智能体架构在全面推广前必须攻克确定性输出保障、防失控机制（防止低成本任务演变成无序调用）以及严格的 API 成本控制。 - [08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) 自助式 RAG 与内部智能体市场：推出直连 SharePoint 的自助知识检索方案，并建立内部工作流市场与物质激励机制，驱动员工自发构建并复用优质 Agent 资产。 - [09:41](https://www.youtube.com/watch?v=t_26JjS6C6w&t=581s) 安全门槛与全量数据量化看板：全员须通过 1 小时强制合规培训方可获得使用权限；系统全量追踪交互行为并构建生产力看板，精准识别业务部门中的采纳滞后者与领跑者。 - [11:55](https://www.youtube.com/watch?v=t_26JjS6C6w&t=715s) AI 熟练度阶梯与效率价值：将员工划分为从新手到教授级五档，专家级员工每月可节省约 20 小时工时，普通员工通常需 3 至 4 个月跨越进阶门槛。 - [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) 算力成本大幅下降与极致性价比：随着大模型性能提升与 Token 成本骤降，全行 AI 平台的运行支出对于大型银行而言仅是微不足道的零头，成本效益远超第三方方案。 - [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) 核心教训——组织采纳远难于技术开发：技术团队极易陷入功能自嗨，唯有倾听员工真实需求、推动全员工作习惯与业务行为转变，才能释放 GenAI 的规模化价值。 - [17:05](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1025s) 社区驱动赋能与深层工作流嵌入：通过业务部门自发组织的工作坊与变革倡导者网络实现去中心化培训，员工对 AI 产生高度依赖，重塑了银行业务日常运转方式。 - [19:01](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1141s) 上下文注入与数据溯源的实战验证：内部 RAG 工具“Smart Library”基于沉淀 PPT 自主完成了 12 道奖项申报题目中的 9 道，证明高质量数据基础与上下文注入在严谨业务场景下的惊人生产力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) CIBC 内部的“微型 AI 初创团队”与自研起步 CIBC（加拿大帝国商业银行）企业级 AI 平台与解决方案负责人 Chris Patterson 开门见山地将自己的团队定位为“银行内部的微型 AI 初创公司”。管理层给了他们一笔预算，允许他们大胆尝试、摆弄各种大语言模型，看看究竟能为银行创造什么价值。整个团队运作敏捷、高度创新，核心目标就是通过生成式 AI（GenAI）驱动企业效率。回顾一年前的历程，CIBC 当时已经花了大半年时间测试微软的 Copilot。然而 Chris 对 Copilot 并不满意，甚至直言“我现在依然不喜欢它”，原因在于其投入产出比和实际业务价值很难得到合理解释。于是，他直接找到高层领导提议：“给我一笔资金，让我们自己造一个，我确信能做出比 Copilot 更好的东西。”管理层虽然笑了笑，但还是批准了经费，对他说：“去放手干吧（Have at it. Go for it.）”。团队在上一届大会前一周正式上线了该产品；如今整整一年过去，产品已从早期原型扩展至支持 45,000 名员工。Chris 指出，为 5 到 50 个人做原型很容易，但要搭建一个企业级安全、被全行上下理解并接受的产品，过程极其艰难。 - **组织角色**：定位为大行内部创新的“微型 AI Startup”，自主掌控技术栈。 - **立项背景**：因外部商业方案（如 Copilot）价值论证不足，决定自研替代方案。 - **规模跃升**：从一年前的初步上线，扩展到支撑全行 45,000 名员工日常使用。 --- ### [01:53](https://www.youtube.com/watch?v=t_26JjS6C6w&t=113s) 产品形态与场景化引导：降低员工使用门槛在产品设计上，Chris 坦言该工具本质上类似于 ChatGPT 的企业定制克隆版，并没有自称颠覆性的底层创新，但针对 CIBC 的实际工作流增加了大量专属功能。最核心的设计之一就是各类“预设功能按钮”。在产品上线前，团队曾误以为多数员工已经熟悉大模型交互，但现实情况是，大量员工此前从未接触过 ChatGPT，根本不知道 LLM 能做什么。为了引导员工获得可预期的输出，系统内置了多种标准化操作按钮：一键文档摘要、邮件起草与回复等。随后，团队还推出了代码生成功能。Chris 分享了一个有趣的真实案例：他原本以为该功能只会吸引程序员，结果在查看用户反馈时，发现法务部的一名律师居然用它写了一大串 Python 脚本去抓取和分析自己的邮箱数据；该律师因缺乏本地管理员权限和所需 Python 库导致脚本无法运行，并在群里求助。这生动证明了生成式 AI 正在打破技术壁垒，让非技术人员也能触达以往难以企及的自动化能力。 ```mermaid flowchart TD accTitle: CIBC AI 核心架构与功能流向 subgraph Input["企业数据与上下文集成"] A["收件箱 & 日历 (Inbox & Calendar)"] B["企业文档 & 政策库 (CIBC Data / RAG)"] end subgraph Core["CIBC AI 平台引擎"] C["预设引导按钮 (Workflow Buttons)"] D["企业级大语言模型 (LLMs)"] end subgraph Output["业务功能场景"] E["文档摘要 & 邮件生成"] F["代码编写 & 自动化脚本"] G["Mermaid 流程图 / 甘特图生成"] H["HTML 视觉原型即时渲染 (Prototype Builder)"] end Input --> Core Core --> Output ``` - **引导式 UI（Guided Buttons）**：解决员工“面对空白提示框不知道问什么”的痛点，固化高频工作流。 - **跨系统连接**：无缝对接员工的收件箱、日历和本地文档，并基于 CIBC 内部数据进行检索增强（RAG）。 - **流程图与图表绘制**：支持直接通过自然语言生成 Mermaid 流程图和甘特图，上线后已有数千名员工日常使用。 - **极速原型构建器（Prototype Builder）**：针对业务人员与技术团队沟通需求不畅的问题，新上线的按钮允许业务人员直接生成 HTML 并在界面内实时渲染界面原型。 --- ### [05:20](https://www.youtube.com/watch?v=t_26JjS6C6w&t=320s) 演进路线图：会议助手、团队协作工作区与 Agentic AI 谈及产品路线图，Chris 介绍了团队正在推进的几项核心能力。首先是计划于 9 月全面推广的**自研会议记录助手（Notetaker Bot）**。虽然市面上已有类似 Fathom 的第三方会议工具，但作为受监管的金融机构，CIBC 绝不能将敏感会议数据暴露给外部厂商去训练未知模型，因此团队选择自建。该功能将彻底解放项目经理、产品人员及合规团队的手写笔记负担，并结合类 RAG 能力实现会议记忆回溯。其次是**团队协同工作空间（Team Workspace）**。团队借鉴了 Google NotebookLM 的交互理念，打造面向企业敏捷交付的协同前端。AI 可以挂载到特定项目或 Sprint，连接 Jira、Confluence 和 SharePoint，围绕专属上下文为技术和业务团队提供深度协作支持。在音频与实时交互方面，Chris 描绘了从“会后总结”升级为“会中实时副驾”的愿景：例如在投资顾问与客户沟通时，AI 充当后台初级助理，实时监听对话并在客户提问时主动秒级检索对应数据推送给顾问。最后，针对当前备受关注的 **Agentic AI（智能体）**，Chris 表示团队已建立大量 PoC（概念验证），但推向全行前仍需攻关诸多严苛挑战，尤其是如何确保执行结果的准确性，防止 Agent 出现“失控跑偏（going rogue）”——比如把一项仅需 5 美元成本的任务演变成消耗 5 万美元的无序调用。 - **自研 Notetaker**：确保数据合规与隐私安全的前提下，提供全量会议自动记录与历史上下文检索。 - **Notebook 协同工作区**：借鉴 NotebookLM 模式，针对特定项目或交付目标打通 Jira/Confluence/SharePoint 文档池。 - **会中实时助手（In-meeting Copilot）**：从异步处理演进为同步实时监听与主动信息检索。 - **Agentic AI 落地卡点**：核心攻坚方向包括确定性输出保障、防失控机制以及严格的 API 成本/调用控制。 ### [08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) 员工使用场景与自助式 RAG 赋能在日常工作流中，员工对 CIBC 内部 AI 工具的使用正在从基础场景向高阶能力拓展。最普遍的使用场景依然是文档总结（Summarization）和邮件起草（Email Drafting）。与此同时，生成定制化信息图表（Diagrams）等新场景也开始涌现。为了彻底打破技术壁垒，团队推出了自助式检索增强生成（Self-serve RAG）方案：员工无需技术团队介入，系统即可自动对数据进行索引、抓取、提取，并直接连接至 SharePoint，实现开箱即用的企业知识检索，极大加速了全行范围内的技术采纳。 ```mermaid flowchart LR accTitle: 自助式 RAG 与智能体市场流通架构 subgraph S1["数据接入层"] A["SharePoint / 文档数据"] --> B["自动抓取与索引引擎"] end subgraph S2["能力构建与共享"] B --> C["自助式 RAG 知识检索"] C --> D["定制化工作流 Agent"] D --> E["内部应用市场 (Marketplace)"] end subgraph S3["组织推广"] E --> F["同行复用与激励分红"] end ``` 为了进一步激发员工的创造力，团队正在推出一个内部应用市场（Marketplace）。虽然演讲者调侃自己并不太喜欢言必称“智能体”（"I don't like using the word agent, but I have to rebrand around that"），但团队允许员工构建简单的工作流智能体，并将其发布到内部应用商店中供其他同事复用。更具突破性的是，CIBC 甚至设计了一套激励机制，对分享优质工作流的员工给予物质补偿，以此全力推动内部采纳。 ### [09:41](https://www.youtube.com/watch?v=t_26JjS6C6w&t=581s) 采纳规模与生产力价值量化模型在上线推广方面，CIBC 设定了严格的安全门槛：作为一家受到高度监管的银行，任何员工若想获得该产品的访问权限，必须先完成至少 1 小时的强制性合规培训，以了解 AI 的底层工作原理与潜在风险，确保安全可控。自 5 月底正式发布以来，用户规模呈现出爆发式增长。 - **日常活跃用户**：每天有 7,500 名员工活跃使用。 - **登录注册总量**：截至上周达到 24,000 人，目前已突破约 26,000 人（超过全行员工总数的一半），未来目标是拓展至 40,000 人。 - **全量埋点追踪**：系统采集用户在平台内的每一次点击、每一场对话、每一次提示词交互、所使用的模板配方（Recipe）以及对话深度。 - **生产力模型测算**：基于真实交互数据建立生产力计算模型，为各业务线管理者提供团队维度的细分看板，精确识别出采纳滞后者（Laggards）与创新领跑者，并借此在各业务线建立“变革倡导者”（Change Champions）网络。 ### [11:55](https://www.youtube.com/watch?v=t_26JjS6C6w&t=715s) AI 熟练度阶梯与分层培训体系 CIBC 依据全量数据将员工的 AI 熟练度（AI Fluency / AI Literacy）划分为不同层级。如果员工一个月仅使用一次文档总结，无论频率如何都只能算作新手（Novice）；而进入进阶（Intermediate）、从业者（Practitioner）乃至专家（Expert）级别后，员工则彻底重塑了自身的工作流与工作行为习惯。最高层级的“教授级”（Professors）员工几乎完全重构了自己的岗位职责，成为推动全行拥抱技术的领军者。 | 熟练度层级 | 行为特征与价值体现 | |:--- |:--- | | **新手 (Novice)** | 仅偶尔用于单一场景（如每月总结一次文档），未体现深层理解 | | **进阶 / 从业者 (Intermediate / Practitioner)** | 定期登录系统，将其融入多项日常辅助任务中 | | **专家 (Expert)** | 彻底重塑工作流程与行为模式；**每月可节省约 20 小时工时** | | **教授级 (Professors)** | 用 AI 全面重构自身岗位职责，并作为业务先锋引领组织变革 | 根据真实访谈验证，普通员工平均需要 3 到 4 个月才能进阶到专家水平，但并非所有人都能顺利跨越这一门槛。为了攻克“人与行为改变”这一最艰难的环节，团队构建了立体化的培训体系： - 专门聘请培训专家每周组织 AMA（Ask Me Anything）答疑直播，单场覆盖 200 至 300 名员工，提供开放式提问平台。 - 搭建跨业务线的“变革倡导者”网络与 Workplace 交流社区。 - 针对从 Z 世代/阿尔法世代（Gen Alpha）到资深员工的不同学习习惯，提供自适应、多元化的教学媒介。 ### [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) 运行成本效益、模型演进与核心经验教训谈及系统运行成本时，演讲者直言其花费对于如此体量的大型银行来说简直是“微不足道的九牛一毛”（"Peanuts, honestly it's a rounding error"）。回顾技术演进历程，算力性能指数级上升而 Token 成本却在指数级下降：当初启动项目时使用的 GPT-3.5 Turbo，其单位调用费用甚至比如今能力强大的 o3 模型还要高昂。 ```mermaid flowchart TD accTitle: 模型演化与部署接入架构 subgraph Cloud["模型生态与演进"] M1["早期: GPT-3.5 Turbo (高成本/基础能力)"] M2["当前生产: o3 / GPT-4.x (低成本/高能力)"] M3["开发环境: GPT-5 / 新一代模型测试"] M4["拓展中: Google 系列模型接入"] end subgraph Platform["CIBC 企业级安全平台"] SEC["企业级合规与安全防护网 (Enterprise-grade Secure Gate)"] M1 -.->|逐步淘汰| SEC M2 --> SEC M3 -.->|认证评估后上线| SEC M4 -.->|适配集成| SEC end SEC --> BIZ["赋能全行业务场景与员工 Agent"] ``` 目前 CIBC 已经在开发环境中测试 GPT-5 等前沿模型，并同步与 Google 合作接入多元化模型矩阵。演讲者分享的核心教训在于：**技术采纳远比技术本身更难**。技术团队很容易陷入不断开发新功能的自嗨中，但如果不能带领全行员工共同成长、转变工作习惯，再炫酷的功能也只是开发者“在本地环境自娱自乐”。倾听员工真实需求、推动组织变革，才是释放生成式 AI 规模化价值的核心所在。 ### [17:05](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1025s) 社区驱动赋能与深植银行业务的工作流重塑 CIBC 在推广生成式 AI 工具时发现，全员 AI 素养与能力的提升确实带来了向“专家级”熟练度的实质性飞跃。员工对该产品的热情与依赖度极高，甚至达到了戏剧性的程度：一旦系统偶发停机仅一个小时，团队收件箱就会被海量求助邮件淹没；更有员工开玩笑表示，在 AI 恢复上线并代劳之前，他们宁愿暂缓处理某些特定任务。面对 45,000 名员工的庞大规模，仅凭中心化的 AI 团队显然无法承担全部培训重任，因此 CIBC 转向了“社区驱动”（Community-driven）的赋能模式——让各个业务部门主导，自发组织工作组和非正式交流会，互相分享如何利用模型驱动效率与业务变革。在强监管的传统银行业环境中成功上线并规模化推广此类工具，是一项极具挑战性的突破。回顾 ChatGPT 最初出现时，银行业的标准反应往往是直接拦截封锁（"No, you can't have this tool"）。而 CIBC 能够打破常规，在大型受监管金融机构中实现全员级 GenAI 部署，充分证明了其战略路线的成功。 - **依赖与工作习惯重塑**：员工工作方式发生根本转变，系统短暂停机即引发强烈反馈，甚至出现“等 AI 恢复再干活”的深度工作流嵌入。 - **业务自主扩散**：通过业务部门自发的兴趣小组与工作坊，实现去中心化的知识共享与用例挖掘。 - **突破监管壁垒**：从最初对生成式 AI 的普遍封禁，跨越到合规、安全地为数万名员工提供企业级内部 AI 支撑。 --- ### [18:31](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1111s) 成果转化、极致成本效益与面向 Agentic AI 的展望员工 AI 熟练度的提升直接转化为了显著的业务成果与真实的生产力价值。从投资回报率来看，自主研发与平台化运营所花费的成本，与市面上高昂的第三方供应商工具相比几乎微不足道（"trivial in comparison to any of the vendor tools"），展现了极高的性价比和经济效益。谈及未来规划，团队正全力攻坚智能体架构，并期望在下一年重返舞台时，能向大家分享 CIBC 在 Agentic AI（智能体 AI）领域的落地实践与解决方案。 ```mermaid flowchart TD accTitle: CIBC AI 规模化赋能与演进路径 subgraph S1["能力建设与赋能"] A["社区驱动培训"] --> B["全员 AI 熟练度提升"] end subgraph S2["业务价值落地"] B --> C["高黏性工作流重塑"] B --> D["可量化的高生产力产出"] end subgraph S3["平台演进与未来"] C --> E["极低自研成本优势"] D --> E E --> F["下一阶段：Agentic AI 智能体架构"] end ``` --- ### [19:01](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1141s) “让 AI 为自己赢下大奖”：Smart Library RAG 的实战趣闻 CIBC 的生成式 AI 实践两次荣获“最佳生成式 AI 倡议”（Best GenAI Initiative）大奖。在此背后，还有一个极具戏剧性且幽默的真实故事：在奖项申报截止的当天，团队才临时接到提交申请的紧急通知。面对时间紧迫的死线，主讲人并没有慌乱手写，而是直接调用了内部的检索增强生成（RAG）工具——“Smart Library”。他将过去一年里制作的全部项目 PPT 作为上下文知识库，连同奖项评审标准及 12 道申报问题一并输入系统。Smart Library 表现惊艳，极为出色地直接回答了其中的 9 道题；主讲人仅针对缺乏背景数据的其余 3 道题进行了人工补充。最终，这份申报材料成功折桂，“CIBC AI 凭借自己的能力为自己赢得了大奖”（"I had the CIBC AI win its own award"）。这一实战案例生动地证明：只要为大语言模型提供准确、真实的数据、充分的上下文以及可靠的信息溯源能力，大模型就能在企业严谨场景下爆发出不可思议的生产力。 - **获奖荣誉**：CIBC 两次斩获“最佳 GenAI 倡议奖”（Best GenAI Initiative）。 - **紧急申报应对**：申报截止当日临时受命，借助内置检索增强生成的 Smart Library 完成材料撰写。 - **人机协同分工**：12 道申报题目中，AI 基于沉淀 PPT 完美作答 9 道，人工仅补充未收录背景的 3 道。 - **核心启示**：企业级 LLM 应用的关键在于高质量的数据基础、精确的上下文注入（Context Grounding）与可追溯的溯源机制。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) 给我一笔资金，让我们自己造一个，我确信能做出比 Copilot 更好的东西。 — Chris Patterson > [00:04](https://www.youtube.com/watch?v=t_26JjS6C6w&t=4s) 去放手干吧（Have at it. Go for it.） — CIBC 管理层 > [08:28](https://www.youtube.com/watch?v=t_26JjS6C6w&t=508s) I don't like using the word agent, but I have to rebrand around that — Chris Patterson > [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) Peanuts, honestly it's a rounding error — Chris Patterson > [15:03](https://www.youtube.com/watch?v=t_26JjS6C6w&t=903s) 技术采纳远比技术本身更难 — Chris Patterson > [17:05](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1025s) No, you can't have this tool — Chris Patterson > [18:31](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1111s) trivial in comparison to any of the vendor tools — Chris Patterson > [19:01](https://www.youtube.com/watch?v=t_26JjS6C6w&t=1141s) I had the CIBC AI win its own award — Chris Patterson ## Mentioned Chris Patterson, CIBC, CIBC AI, Copilot, ChatGPT, Python, Mermaid, Fathom, Google, NotebookLM, Jira, Confluence, SharePoint, GPT-3.5 Turbo, o3, GPT-4.x, GPT-5, Smart Library ## Video description Chris Patterson, Head of Enterprise AI Platforms and Solutions, CIBC Abstracts Join me at my talk for a year-in-review of CIBC’s Generative AI journey. Last summer, we introduced our AI product CIBC AI to 500 employees; today we’ve scaled its reach to over 45,000, reshaping enterprise-wide productivity. Discover the strategic milestones that paved our path—spanning user engagement, champion-led & training backed AI literacy, and our roadmap toward Agentic AI. Walk away with real-world insights, lessons learned, and a first-hand look at the transformative potential of generative AI at scale. [Source](https://www.youtube.com/watch?v=t_26JjS6C6w)

Source quotes

[0:04] Chris Patterson: 给我一笔资金，让我们自己造一个，我确信能做出比 Copilot 更好的东西。 [0:04] CIBC 管理层: 去放手干吧（Have at it. Go for it.） [8:28] Chris Patterson: I don't like using the word agent, but I have to rebrand around that [15:03] Chris Patterson: Peanuts, honestly it's a rounding error [15:03] Chris Patterson: 技术采纳远比技术本身更难 [17:05] Chris Patterson: No, you can't have this tool [18:31] Chris Patterson: trivial in comparison to any of the vendor tools [19:01] Chris Patterson: I had the CIBC AI win its own award

Source quotes

[0:04] Chris Patterson: give me some funding, let me build our own thing... I'm sure I could do something better than Copilot. [0:04] Chris Patterson: micro AI startup within the bank [1:53] Chris Patterson: clone of ChatGPT [4:02] don't always hear me, they don't always understand what I want. [5:20] Chris Patterson: going rogue and spending 50 grand on a task that was probably only really should have been five bucks. [8:28] Chris Patterson: We're going to create a storefront where you can actually share this and I'll even pay them. We have a scheme where I can compensate people for sharing this stuff. Why not? Why not? I want to drive up adoption. [11:55] Chris Patterson: Experts [save] 20 hours a month... When folks really use this product, they're getting some real value from it. I have spoken to hundreds of our experts to validate this claim... I want to know what are you doing? How are you using this product? Show me. [16:29] Chris Patterson: Adoption is hard. This is where I spend almost all my time. I would much rather be innovative. I would much rather be building cool features, cool product ideas. And I did focus a lot on, you know, the next best feature, the next best product. And I realized, you know, fairly early on, this is a waste of time. If I don't bring the people along with me, they won't use it anyway.

Original transcript

[0:04] Good morning everyone. Thank you for [0:05] coming to my talk. Uh my name is Chris [0:07] Patterson, head of enterprise AI [0:08] platforms and solutions at CIBC. Think [0:10] of me as the um micro AI startup within [0:14] the bank. So they've given me a little [0:15] bit of funding. They can just have at [0:16] it, play around with language models and [0:18] see what we can do. So my group is kind [0:20] of innovative, fast moving and uh you [0:23] know really designed around you know [0:25] driving efficiency value through uh [0:27] Genai. So today I'm going to speak to [0:30] you about our journey. So I was here [0:32] last year and I spoke to uh you know the [0:35] audience around you know what it is that [0:38] you know CBC was doing at that point in [0:40] time with GEI and at that point we've [0:42] been playing around with C-Pilot for the [0:44] better part of a year. Now I didn't like [0:46] copilot. I still don't like co-cloud. Um [0:48] I just couldn't justify you know the [0:51] value behind uh you know what it was [0:52] trying to do and and so on so forth. And [0:54] so I went and spoke with you know the [0:55] powers that be and said look you know [0:57] give me some funding. let me build our [0:59] own thing, right? And um you know, I I'm [1:02] sure I could do something better than [1:04] co-pilot. And uh you know, they they [1:07] laughed, but they still gave me some [1:08] money and I said, "Have at it. Go for [1:10] it." And so we we launched actually the [1:12] literally the week before I was here [1:14] last year and uh you know, we had turned [1:16] that thing on. So a full year has gone [1:18] by. So today I'm here to talk to you [1:20] about you know, what is it that we [1:21] built? How does it work? You know, what [1:23] are some of the core features? You know, [1:24] what's the road map look like? and uh [1:26] you know really what um you know what [1:29] are some of the key stats you know what [1:31] are the learnings that we've gotten from [1:32] this journey cuz it's been really [1:33] difficult you know scaling a product up [1:35] for 40,000 is not easy you know it's [1:38] easy to build a pro prototype for 5 10 [1:40] 50 20 whatever but getting something [1:42] stood up enterprise safe and um you know [1:45] understood by the enterprise was no no [1:48] straightforward journey so I've got a [1:50] bunch of lessons that I'll share and uh [1:52] you know we'll take you through okay so [1:53] first off what did we build. Honestly, [1:57] it's kind of a clone of Chat GBT. I'm [1:59] not going to say it's all that [1:59] innovative. You know, I'm not that uh [2:02] much of a product guy. I'm going to try [2:03] and create something completely novel. I [2:05] don't have a startup. But that being [2:06] said, uh you know, we have a few bells [2:08] and whistles on here that make it a [2:09] little bit unique. It's tailored to [2:10] CIBC. It uh you know, has a bunch of [2:13] buttons. You know, these buttons do [2:14] different things depending on what you [2:16] know the the workflows are. And they [2:18] were designed specifically to guide [2:21] employees towards predictable outcomes [2:23] with GI. We had to bear in mind a lot of [2:25] folks I had the assumption that a lot of [2:27] employees would know immediately what [2:30] they could do with the language model. [2:31] That was not the case at all. A lot of [2:33] folks never played with chat GPG before [2:35] when we launched this. I was shocked [2:37] honestly like I just assumed by this [2:39] point everyone would be kind of verse [2:40] and understand how these things work. [2:42] They did not. So we created some buttons [2:45] right? You want to summarize a document. [2:46] Oh okay. I didn't know LM could do that. [2:48] Well now you do. Right. Prod uh the [2:50] emails. Right. A lot of folks use it for [2:52] summarizing, for producing emails, [2:54] responding, common use cases. And we've [2:57] extended that a little bit more over [2:58] time. We have software one uh where, you [3:00] know, folks can go in and have it write [3:02] code. I had to laugh the other day [3:04] because I was anticipating that it would [3:06] be developers using that particular [3:07] feature. And I was looking at some [3:09] feedback and I had a lawyer in there and [3:11] he had written a whole bunch of Python [3:12] to scrape through his inbox and he was [3:14] doing some stuff and he's like, I can't [3:15] get this to run. And it's like well you [3:17] know there's a bunch of people trying to [3:18] help him but he needed local admin in [3:20] the Python you know libraries installed [3:22] on his laptop but it it opens the door [3:25] to you know again access to knowledge [3:27] that isn't necessarily you know easy to [3:30] you know to get right and these language [3:32] models are extremely versatile. Uh so we [3:35] we're just constantly rolling new things [3:36] out you enhancing the product. This [3:38] connects to your inbox your calendar [3:41] your documents. So it's really designed [3:43] as just an augmentation productivity [3:45] tool. you have some work to do, throw in [3:47] the tool. Let the tool do it if it can, [3:49] right? Summarize your doc, help you with [3:50] a PowerPoint, you know, review an email, [3:53] respond to a client inquiry, all of that [3:55] sort of stuff. You, you know, it's [3:56] grounded on some of the CIBC data, and [3:59] then you can simply ask it to do things. [4:02] So again, what types of things? You [4:04] know, busy slide, but again, this [4:05] typical, right? You can drop files in, [4:08] summarize, you can um search for policy [4:12] and procedure information. We've [4:13] grounded it on, you know, self-service [4:16] sort of rag implementation. We just [4:17] rolled out this diagram feature. You [4:19] know, AI can now create mermaid diagrams [4:21] and, you know, generate process flows, [4:23] Gant charts, all sorts of things. It [4:25] just does it. And we just, you know, one [4:28] of the my tech guys showed me the [4:29] feature. I thought it was great. Let's [4:30] roll it out. And we have now thousands [4:32] of people producing these types of [4:34] images just like that. Just on Friday [4:36] before I came here, I had this idea. You [4:38] know what? I was talking to some product [4:39] guys and they were telling me, you know, [4:40] it would be nice if we had a way to do [4:42] visual prototypes, right? I have this [4:44] new feature. You know, I want to create [4:46] a a a visual feature, right? Because [4:48] when I'm talking to tech, you know, they [4:50] don't always hear me. They don't always [4:51] understand what I want. And so, I [4:53] thought about it. I thought, actually, [4:54] that's pretty easy. And we spun up a [4:55] button, prototype builder. It generates [4:57] some HTML. It renders it right there. [4:59] And it's I launched that Friday. It's [5:02] only what, Tuesday? I've already had a [5:03] bunch of emails from folks saying, "This [5:05] is really cool feature. I like it. Like [5:07] this tech is amazing at what it can do [5:09] and by having our own product, our own [5:11] platform, we're just able to augment it [5:14] and it stand up these different custom [5:16] bespoke features that are, you know, [5:18] relevant to our enterprise. [5:20] So, you know, in terms of road map, we [5:22] have another couple core capabilities [5:25] that we're looking to bring forward. [5:26] Noteaker being a big one. Uh, this is [5:28] something you've seen probably a lot [5:30] like Fathoms one. There's a bunch of [5:31] vendor tools out there and um [5:33] unfortunately we can't, you know, use [5:36] any of those different products because [5:37] we don't want our data going out [5:39] training who knows what model. We don't [5:40] know how they secure. We don't know how [5:42] any of that stuff works. And so we built [5:44] our own noteaker Bob. Why not? Like we [5:46] we can host this. We have the knowhow. [5:47] We know how we have the tech and it [5:49] works amazingly well. Now, you know, [5:51] we'll launch that more broadly in [5:53] September. And any meeting any meeting [5:55] that we have can now just be auto [5:57] notetaked, right? And honestly, the [5:59] number of people who are sitting there [6:00] doing handwritten notes today, this is [6:02] going to like change their lives, right? [6:04] From PMs to product people to pretty [6:06] much anyone, governance groups, the [6:08] works. Simple simple feature, but it [6:10] will unlock tons of productivity. You [6:12] can go back in time like I'm sure Bob [6:14] asked me something and I I forget, [6:16] right? I my memory is lousy. So just [6:18] having access to those notes in a quasi [6:21] rag situation will make it really easy [6:22] for me to remember and and don't miss [6:24] anything. Team workspace. I'm learning, [6:27] you know, I'm taking the product to the [6:28] next level, looking at how do we now [6:31] create more of a collaborative [6:33] uh front end where teams can work [6:34] together with an AI assistant grounded [6:37] on like a project, right? Like if you're [6:39] in technology, maybe you're working on a [6:41] deliverable, you know, let's ground the [6:43] AI, this notebook, as I'm going to call [6:44] it, borrowing a little bit from Google, [6:46] but I like their product. I like [6:47] notebook LM. It looks it's it's [6:49] interesting. It works in a way that I [6:51] think would also work really well in an [6:53] enterprise setting. And so we're [6:55] building that sort of quasi rag, you [6:57] know, uh, product capability and and [7:00] connecting it to things like Jira to [7:01] Confluence to SharePoint in a way that [7:04] grounds it on a certain specific set of [7:06] materials related maybe to a sprint, to [7:09] a project, to a product, it doesn't [7:11] matter. It's going to be flexible, open [7:12] for, you know, interpretation in terms [7:14] of how it can get used. Uh, audio like [7:17] these models are enhance are advancing [7:19] faster than I can keep up. And so if you [7:21] go to like notetaker for example, it's [7:23] great having the, you know, the AI take [7:25] notes after the meeting's finished, but [7:26] what about while the meeting's [7:27] happening? You know, maybe I want this [7:29] AI to almost act like a junior assistant [7:31] sitting there listening to our advisor [7:33] client thing and like if the client asks [7:35] a question, this guy's proactively going [7:37] out and quickly fetching data like I [7:39] know that I know where that can get [7:40] that. And it can provide all of that [7:42] immediately to the adviser as they're [7:44] having that conversation just like that, [7:46] right? Like all of this capability [7:48] exists today. It's there. We just have [7:50] to kind of wire it up and make it [7:51] available. And then agents, obviously, [7:53] everyone's after my life, but agentic [7:54] AI, I'm working on it. It's what I keep [7:56] saying, we're getting there. But, you [7:58] know, there's a lot of hoops to get [7:59] through before we can launch this stuff. [8:02] There's a lot of challenges with like [8:03] pure proper agents that haven't been [8:06] solved. Like, there's proof of concepts. [8:08] We have a whole bunch of PC's up and [8:10] running. And, you know, the challenge [8:11] with Agentics still is how do you make [8:14] sure that the outcome is correct? You [8:16] know, you don't have these agents going [8:18] rogue and spending 50 grand on a task [8:21] that was probably only really should [8:23] have been five bucks. Like there's [8:24] endless things that we had to solve [8:26] before we can launch this at scale. But [8:28] we're getting there. We're working on [8:29] it. You know, this is my life now all of [8:30] next year, I'm sure. So now, if you look [8:33] at like how are employees using it, like [8:35] I've shown you, summarization is a big [8:36] deal. Like that one's common. Email [8:38] drafting common. Diagrams, that's new. [8:41] All right? like folks are producing [8:42] quick uh you know quick tailor made uh [8:44] information diagrams and whatnot. We [8:46] have self-s serve uh self-s serve rag, [8:49] right? If you want to spin up a [8:50] retrieval, you know, solution for [8:52] yourself, go for it. I don't need to be [8:53] in the way. Like tech doesn't need to be [8:55] there. We can just auto index, scrape, [8:57] pull, connect directly to SharePoint. [8:59] Off you go. It just happens [9:00] automatically. And this is like again [9:02] accelerating our our adoption across the [9:05] organization. You know, there you we [9:07] we've we're launching this, you know, [9:09] thing called a marketplace. What we [9:11] found with our users, we allow them to [9:13] create simple agents. I don't like using [9:15] the word agent, but I have to rebrand [9:17] around that whether I like it or not. [9:19] And we're creating a marketplace for [9:20] this so that those folks who've [9:21] discovered how they can create these [9:24] workflows that work for them. Well, I'm [9:25] sure it's going to work for their [9:26] colleagues. So, let's help them [9:28] publicize that. And we're going to [9:29] create a storefront where you can [9:31] actually share this and I'll even pay [9:33] them. We have a scheme where I can [9:35] compensate people for sharing this [9:37] stuff. Why not? Why not? I want to drive [9:39] up adoption. That's how we're going to [9:41] do it. All right. So, what does our [9:43] adoption look like? I got some [9:44] interesting charts. So we have 7,500 [9:47] folks using it every day. It's pretty [9:49] good, right? 24,000 actually that's [9:52] effective last week. I think we're about [9:53] 26,000 now have logged in, right? We've [9:56] done we have one of our gates is [9:57] mandatory training. So if you want to [9:59] get access to the product, you got to do [10:01] at least uh one hour of training just to [10:03] get a sense of how AI works, what the [10:05] risks are. You know, we are a bank, [10:07] we're regulated, so we don't want people [10:09] going rogue and using it for ways that [10:10] we don't like. uh but we've secured it [10:12] so it it works really well for us and [10:14] it's um you know delivering real value [10:16] and you can see we launched officially [10:18] in end of May and it's been you know [10:20] strong growth uh since then. So I'm [10:23] really pleased with what's coming what's [10:24] been happening and I will grow that [10:26] number to 40,000 over time. Now what [10:29] does that mean in terms of productivity? [10:31] So like material like these numbers I [10:35] have actual math I mean it's real math I [10:37] can't somewhat estimated I will admit [10:39] right but you know nonetheless like I'm [10:41] I what I do is I capture every single [10:44] data point I can within our product [10:46] everything every click every [10:48] conversation every chat every button [10:50] click the works I want it all right I [10:52] want that because I learn from I use it [10:54] and I've used that to build a set of [10:56] models to actually calculate what I [10:58] think the productivity lift is so uh it [11:01] again somewhat estimated however based [11:04] on true math and I actually look at the [11:06] actual clicks the usage the prompts the [11:09] depth of conversation all of it what [11:10] recipe they were using what agent [11:12] whatever and I'd actually come up with [11:14] some numbers that actually showcases uh [11:16] you know really like how much [11:17] productivity this is calc like you know [11:18] creating for the organization I have [11:20] this at the employee level to the day [11:22] all managers have been asking me for [11:24] reports I produce all sorts of reports [11:26] insight sharing them here's what your [11:28] team looks like and they can see like [11:30] who's adopted it, who's not, where the [11:32] laggers are, where the, you know, where [11:34] their um key folks are that are really [11:36] driving innovation. And we've used that [11:37] to create change champions within these [11:39] different networks, different lines of [11:41] business to then drive further growth. [11:43] Tying back to that marketplace, a lot of [11:45] these guys have created meaningful [11:47] recipes that we can now push forward, [11:49] meaningful agents rather, and we're [11:51] going to now again market those. Let's [11:52] get those things out there. All right. [11:55] So from all that data, what I've been [11:57] trying to now do is categorize and [12:00] define what I consider an employees [12:02] fluency is with AI. So this is again I [12:05] think I've heard this a few times since [12:06] I've been here at AI4 now. Like this [12:08] this topic of AI literacy, AI fluency, [12:11] it's it's a thing, right? Like you know, [12:12] not everyone understands how a language [12:14] model works and everyone's, you know, [12:16] uses it for different purposes. So [12:18] someone who uses it only to summarize a [12:20] document once a month, well, they're [12:21] still a novice. I don't care how many [12:23] times they do that, you know, it's not [12:25] really demonstrating any re meaningful [12:27] aptitude with respect to uh, you know, [12:29] genai intermediate. They're at least, [12:31] you know, regularly logging in, doing a [12:33] few things, practitioner, even more [12:35] regular. And, you know, eventually they [12:37] get to expert level where they've like [12:38] literally redefined a lot of their [12:40] workflows. They've changed how they [12:41] work. They have changed their behavior. [12:43] That's been the most important part of [12:45] this journey. Professors are a special [12:47] category. people who I honestly think [12:49] have rewritten their entire job of AI [12:51] should probably be fired. Um, but [12:53] unfortunately I fall in that category. [12:54] So that's not going to happen. But [12:56] nonetheless, you know, we do have a few [12:57] people who stand out who are leading [13:01] from their organization's perspective [13:02] like how to embrace this technology and [13:04] drive it forward. So but who cares? Like [13:06] what does this even mean, right? Well, a [13:08] lot, right? Experts 20 hours a month. [13:11] Like again, when I actually overlay that [13:12] with the productivity data, you can see [13:14] the difference, right? When folks really [13:16] use this product, they're getting some [13:18] real value from it. I have spoken to [13:20] hundreds of our experts to validate this [13:22] claim. This is again estimated to some [13:24] extent. However, I have gone and done [13:26] the hard work to actually interview [13:27] these folks. I want to know what are you [13:29] doing? How are you using this product? [13:32] Show me and I can actually they are [13:34] quite happy to right like take me [13:36] through their workflows. Show me how [13:38] they've actually used this data. You [13:40] know, what it does, how it works for [13:42] them, and what they've done to change [13:43] how they behave. It's amazing. [13:45] And in again from that data I can see [13:47] how long it takes folks to actually [13:49] progress through that journey. It's [13:51] about 3 to four months on average. Uh [13:54] what's unfortunate though is not [13:55] everyone gets there. And this is again [13:57] part of the thing and part of the [13:59] challenge with rolling out a product of [14:01] this nature the people right like [14:04] driving your employees to use it to [14:05] change their behavior that is really [14:07] really hard. We have focused on you know [14:10] building a really thoughtful training [14:12] regimen around this. I even hired a guy [14:14] from uh training to do AMA sessions [14:17] every week. He has uh he runs through at [14:20] least two 300 folks weekly to just teach [14:23] them how to use it, what you can do and [14:25] people just have an open forum where [14:26] they can ask questions and learn. Uh we [14:28] have change champion network being built [14:30] throughout the business throughout the [14:31] networks. We have um you know workplace [14:34] group where people can share. The thing [14:36] is you need different mediums. one [14:37] medium isn't going to solve it because [14:39] we have employees of all sorts of [14:40] different types, right? From Gen Alpha [14:42] all the way up, right? And everyone [14:43] learns in different ways. They want [14:46] access to training materials in [14:47] different ways. So, you have to be [14:49] adaptable uh around how you, you know, [14:51] teach your employees. But that being [14:53] said, again, we have over half the [14:54] organization who's voluntarily taken our [14:57] training to learn more about this [14:58] product and get access to it. So, that [15:00] tells me something. You know, the the [15:01] the need is there. All right. I love [15:03] this slide. What does this cost, right? [15:05] How much does it cost to run this thing? [15:06] peanuts. Honestly, it's a rounding air [15:09] with, you know, the size of our [15:10] organization. It's amazing how much [15:12] value we're getting for practically [15:14] nothing. You know, when you look at the [15:15] the per token fee of the models, it's an [15:19] amazing graph of capability [15:22] exponentially getting better, cost [15:24] exponentially going down. When you when [15:26] we launched with GPT 3.5 turbo way back [15:28] when, it was more expensive than 03 is [15:31] today. Okay. And 03 is obviously [15:33] fundamentally much more powerful than [15:35] the 35 Turbo, but we're now using the um [15:40] the best models available to us. I don't [15:43] limit it all. We have GPT5 already in [15:45] our dev environment. We're testing it [15:46] out, making sure it works, and I'll I'll [15:48] throw them out there the second we can [15:50] certify them and make them available. [15:51] I've seen mixed feedback online, though, [15:53] so I'm not in a rush to actually put [15:54] those things out. I don't know. We'll [15:55] see. But to me, I feel like they're just [15:57] a wrap around 03 4.1. It feels like [15:59] there's a bit of a model redder thing [16:00] going on there. I don't know yet but [16:02] nonetheless you know we are the second [16:04] new model show up they prove their value [16:06] we roll it out why not uh again I want [16:09] the best available to our staff now [16:11] we're currently limited open AI again we [16:13] had to build enterprisegrade secure so [16:16] we have a very safe way of consuming [16:18] those models we are working with Google [16:20] to uh on board uh a further set of [16:22] models so that we can again make this [16:24] product as robust and uh adaptable as [16:27] possible we're getting there all right [16:29] so lessons learned Adoption is hard. [16:33] This is where I spend almost all my [16:35] time. I would much rather be innovative. [16:37] I would much rather be building cool [16:38] features, cool product ideas. And I did [16:40] focus a lot on, you know, the next best [16:43] feature, the next best product. And I [16:45] realized, you know, fairly early on, [16:46] this is a waste of time. If I don't [16:48] bring the people along with me, they [16:50] won't use it anyway. So, what's the [16:52] point? I'm just building it for myself. [16:53] I might as well just keep working my own [16:54] local. Um, so I've had to spend a lot of [16:57] more time and focus driving that change, [17:00] understanding what our like the needs of [17:03] the employee are. What is it that you [17:05] want to learn about like and and teach [17:06] them how to use this software to get the [17:09] value from it? And but it's been [17:10] working, right? Where I'm seeing a [17:12] material lift towards that expert range. [17:14] It's it's working. It's happening and [17:16] we're getting there. The excitement [17:18] around this product is also real, right? [17:21] There are folks who when when I'm out [17:23] like it goes out for like an an hour, I [17:25] get so many emails, right? Like people [17:26] have really changed how they work and um [17:30] you know even to a point where some [17:32] people refuse to do certain tasks until [17:34] the AI is back and they can have the AI [17:36] do it. I mean fine, fair enough. I'm [17:37] happy with that. Um again, we're looking [17:41] at but I can't do it all. How how how [17:43] can I possibly go and train everybody? [17:44] Right? And that's where the [17:45] communitydriven sort of approach is [17:46] coming from. Let's let the businesses [17:48] lead. And we already have tons of [17:49] examples of uh you working groups uh you [17:53] know meeting groups around you know even [17:55] even unscheduled sort of stuff where [17:56] folks are just going and sharing and [17:58] learning how these models are helping [17:59] them drive efficiency and change. It's [18:01] just been an amazing journey for us and [18:03] I'm really pleased to have been able to [18:05] launch a product like this in a bank in [18:08] a regulated environment that typically [18:10] says no you can't do that we don't like [18:13] it right when chatbt showed up what's [18:15] the first thing we did block it no you [18:17] can't have this tool it will not work [18:19] right forget it right so like come on [18:21] like this is like a really difficult [18:24] thing to have launched for for an [18:26] organization of our size and scale and [18:28] it's been an amazing experience getting [18:30] in [18:31] uh the fluency right like this really [18:33] drives outcomes and you know as we've [18:36] seen those people progress through that [18:38] um you know levels of different [18:40] education it's it's again you've seen [18:42] the data right like that actually drives [18:44] real outcomes real value real [18:46] productivity and it's it's just [18:51] amazing how much we've gotten out of it [18:53] for the cost um it's like trivial in [18:56] comparison to any of the vendor tools [18:57] that we've seen so no I'm really pleased [18:59] with how this has played [19:01] And next year I hope to come back and I [19:03] hope to speak to you about our agentic [19:04] AI journey because I'm sure by then we [19:06] will have solved this somehow. But I [19:08] want to end on um we've won an award. [19:12] You're right. I'm really pleased that [19:13] CBC's twice now won uh the best GEI [19:16] initiative. Thank you so much. I [19:17] appreciate that. Now I had to laugh. I [19:19] have to say I was told on the day u oh [19:22] hey we really want you to apply for this [19:24] award. Can you get it done? You have to [19:25] submit today. I'm like seriously and I I [19:27] kid you not. This is just how we work [19:28] sometimes. And so I'm like, "All right, [19:30] fine." So I what I did, I used our tool. [19:32] I created smart library, which is our [19:34] grounded rag. And I say, "Here's a whole [19:35] bunch of PowerPoints I built over the [19:36] last year. Here's the criteria for this [19:39] award. Here's like 12 questions. Just go [19:42] answer it for me." It did nine of them [19:44] really, really well. And I had to answer [19:46] three on my own because it just was [19:47] information ungrounded. I submitted. We [19:49] won. Couldn't believe it. I had the Kai [19:51] win its own award. I even I don't even [19:52] mind sharing that because why not, [19:54] right? Like it did. But again, it was [19:56] grounded on real data, right? It was not [19:58] like I made stuff up like, "Hey, Kai [20:00] could do this, Kai could do that." No, I [20:01] gave it obviously grounded information, [20:03] but this is what LLMs can do if you give [20:05] it the right data, the right context, [20:07] and the ability to uh, you know, source [20:10] that information. But again, super happy [20:12] to have won the award. Uh, it's been a, [20:14] you know, meaningful journey for us, and [20:16] I really hope you guys learned something [20:17] from this. Thank you so much.