Telus Reaps $100M From AI by Pricing Project Risk Like a Venture Fund

Demonstrating Clear ROI from AI Initiatives With TELUS — Ai4

Ai4

Enterprises generate outsized returns from artificial intelligence not by forcing higher project success rates, but by managing pipelines as dynamic probability-weighted portfolios that extract cash from early operational fixes.

AI-assisted summary of Demonstrating Clear ROI from AI Initiatives With TELUS — original source. Check the original for full context.

Source published:

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 24, 2025, this presentation features Alexandre Guilbault, Head of Artificial Intelligence at TELUS. According to the video description, the session explores frameworks and methodologies for quantifying the value of AI investments, aligning implementations with strategic business objectives, and presenting real-world case studies demonstrating measurable ROI across enterprise sectors.

## TL;DR Alexandre Guilbault breaks down how TELUS scaled its AI practice to generate over $100 million in annual business value despite the industry-wide 80% failure rate of enterprise AI projects. He introduces a dynamic portfolio management model inspired by the multi-armed bandit problem, balancing exploratory OPEX and risk-adjusted expected value rather than chasing unvetted headline figures. Using a call center case study that unlocked $3 million in immediate savings through basic root-cause analysis before training a single machine learning model, Guilbault illustrates the power of prioritizing 'proof of value' over speculative proofs of concept. The talk outlines five core principles for enterprise leaders, emphasizing a shift from IT delivery to strategic consulting and urging organizations to make AI practically useful rather than merely technologically novel.

## Atmosphere & style The presentation carries a pragmatic, disciplined, and refreshing executive tone that directly confronts generative AI hype with hard business realities. Guilbault delivers his insights with confident clarity and consulting-grade structure, using crisp mental models and financial trade-offs rather than abstract technical jargon. Watching the session feels like sitting in an elite corporate strategy briefing where complex data science theory is seamlessly translated into operational common sense and rigorous capital allocation.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) Only 20% of enterprise AI projects succeed in creating measurable business value, creating tension between CFO risk aversion and the fear of falling behind. - [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) TELUS scaled its AI practice to over 100 practitioners generating over $100 million in annual value by optimizing project governance rather than forcing higher base success rates. - [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) Project delivery frameworks should match problem certainty: Waterfall for known goals and paths, Agile for unknown technical paths, and Fail Fast when both path and destination are unknown. - [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) Under the Fail Fast framework, teams must stress-test a project's single weakest link—such as data privacy compliance or data availability—on day one rather than months into development. - [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) Applying a Multi-Armed Bandit model allows organizations to dynamically reallocate capital based on continuously updated expected ROI (Estimated Value multiplied by Probability of Success). - [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) Effective dynamic AI portfolio management requires a diversified pipeline, dedicated exploratory OPEX funding, and early incremental monetization. - [09:45](https://www.youtube.com/watch?v=NBUSofH1XHo&t=585s) Evaluating initiatives via risk-adjusted expected value often demonstrates that lower-headline projects in mature domains yield higher expected returns than large, highly uncertain bets. - [11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) In a $50M repeat-call challenge, root-cause descriptive analytics captured $3M in immediate operational savings through updated training before building any predictive machine learning models. - [11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) Delivering early minimum viable insights de-risks future machine learning deployment by building stakeholder trust, securing privacy approvals, and validating data pipelines. - [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) Enterprise AI teams should function like strategic business consultancies focused on problem diagnosis and process redesign rather than IT delivery shops pushing technology for its own sake. - [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Sustainable AI transformation requires adopting a venture capital portfolio mindset, falling in love with business problems, starting rapidly, monetizing iteratively, and prioritizing Proof of Value over Proof of Concept.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) Introduction and the Paradox of AI Investment

Alexandre Guilbault introduces himself and TELUS, noting that while the company is a household name in Canada, international audiences might only know it as a telecommunications carrier. While TELUS delivers mobile, home, and security solutions, it has diversified aggressively across industries—establishing TELUS Health, TELUS Agriculture, and global IT and AI solutions. Guilbault sits at the center of these diverse lines of business to deploy artificial intelligence across all fields.

Turning directly to the core challenge, Guilbault emphasizes that while AI will undoubtedly define the decade and possesses universally recognized potential, enterprises face a brutal reality: "Only 20% of AI projects are successful in creating value in business." Pitching an initiative to a CFO and telling them there is only a one-in-five chance of seeing a return on investment would get almost any proposal rejected. Yet CFOs are simultaneously terrified of falling behind, creating an intense organizational tension between risk aversion and the fear of missing out (FOMO). Over the past decade, TELUS navigated this exact journey, growing its AI practice from two or three machine learning engineers to over 100, generating more than $100 million in realized annual business benefits. This was not a linear path, but rather a process of learning why AI projects fail and adapting operational frameworks accordingly.

- AI acts as a fundamental disruptor requiring alignment across three pillars: people, process, and technology. - Common failure points stem from human resistance to change (fear of AI), compounding operational complexity, and fragmented business data. - Organizations do not necessarily need to force an unnatural jump in base project success rates to win; even with an industry-standard 20% success rate, disciplined management can extract astronomical value.

```mermaid flowchart LR accTitle: The Corporate AI Investment Dilemma subgraph Tension["CFO Investment Dilemma"] direction TB A["Risk Aversion<br/>(80% Failure Rate)"] B["Fear of Missing Out<br/>(Falling Behind on AI)"] end Tension --> C["Dynamic Portfolio Management"] C --> D["Over $100M Annual Value<br/>(TELUS Practice)"] ```

---

[03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) Matching Delivery Frameworks: Waterfall, Agile, and Fail Fast

Guilbault outlines three foundational delivery methodologies, cautioning any program managers present that he will oversimplify to illustrate clear conceptual boundaries. None of these methodologies are obsolete; each fits a specific profile of technical and business certainty. Traditional Waterfall works when both the destination and the exact path to get there are fully known. In this model, teams calculate an upfront ROI, thoroughly map and mitigate risks, and fund the entire project end-to-end. When the destination is known but the technical execution or path remains uncertain, Agile methodology takes over through incremental deliveries, fast feedback loops, and iterative adjustments.

```mermaid flowchart TD accTitle: Project Delivery Framework Matrix subgraph Quadrants["Delivery Methodology Selection"] W["Waterfall<br/>• Known Destination<br/>• Known Path"] A["Agile<br/>• Known Destination<br/>• Unknown Path"] FF["Fail Fast<br/>• Unknown Destination<br/>• Unknown Path"] end

W --> ExW["Example: Enterprise RAG Chatbot"] A --> ExA["Example: Multi-Agent System / MCP Architecture"] FF --> ExFF["Example: Call Center Repeat Call Reduction"] ```

However, a huge portion of enterprise AI falls into the third quadrant: neither the final destination nor the path to get there is known at the outset. For these dark, ambiguous initiatives, organizations increasingly lean on the "fail fast" methodology. Under Fail Fast, teams still estimate an initial theoretical ROI, but immediately identify and stress-test the project's single weakest link on day one instead of waiting six months.

- **Weakest link examples:** - *Data privacy & regulation:* In a hyper-personalized marketing campaign, using sensitive PII or crossing privacy regulations can kill the initiative. Testing compliance and governance on day one avoids six months of wasted model development and stakeholder alignment. - *Data availability:* For unmapped business operations, validating whether usable, historical data even exists must precede any modeling or architectural considerations. - **Delivery archetype examples across AI:** - *Waterfall:* Building an internal enterprise Retrieval-Augmented Generation (RAG) chatbot against curated internal documentation, where the architectural path and target outcome are established and reliable. - *Agile:* Constructing an agentic workflow (e.g., configuring Model Context Protocol / MCP servers, agent tooling, and orchestration), which is primarily a software engineering problem requiring iterative sprints. - *Fail Fast:* Ambiguous operational challenges, such as tackling root causes behind repeat calls in contact centers without pre-specifying the algorithmic solution.

---

[07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) Beyond Fail Fast: The Multi-Armed Bandit Approach to Value

While failing fast avoids dragging out costly errors, Guilbault argues that organizations can do better than simply killing off projects. At TELUS, the team evolved their operating model by borrowing a foundational data science concept: balancing exploration versus exploitation, modeled after the multi-armed bandit problem. Rather than treating project viability as a static, binary gate (kill vs. keep), the team treats ROI dynamically.

Teams establish a potential return on investment and multiply it by a continuously updated probability of success:

$$\text{Dynamic Expected ROI} = \text{Estimated Total Return} \times P(\text{Success})$$

As exploratory work uncovers data flaws, technical barriers, or quick wins, this probability shifts. Rather than prematurely canceling a project or blindly over-investing, the team constantly re-evaluates whether to exploit a high-probability initiative or explore alternative projects in the pipeline.

To operate this dynamic allocation engine successfully, Guilbault identifies three indispensable prerequisites: 1. **A diverse project portfolio:** Organizations cannot optimize across a single project. Because virtually every business operation can be augmented with AI, teams must maintain a rich pipeline of initiatives across multiple departments. 2. **Dedicated exploratory OPEX funding:** Traditional corporate budgeting demands fully capitalized business cases for predictable builds. Without designated operational expenditure (OPEX) to fund early discovery, teams cannot run the small experiments needed to measure and de-risk early failure points. 3. **Early and incremental monetization:** Deferring value realization to the tail end of a multi-year project spikes delivery risk. Big problems must be decomposed into bite-sized operational chunks, shifting the team's operating mindset from an abstract "proof of concept" (POC) to a concrete "proof of value" (POV).

[09:45](https://www.youtube.com/watch?v=NBUSofH1XHo&t=585s) Portfolio Prioritization and Expected ROI: The Project A vs. Project B Dilemma

To consistently deliver meaningful financial returns, organizations must evaluate AI initiatives by balancing their estimated return on investment against their actual probability of success. Rather than viewing stalled initiatives as outright failures, the team implements a rapid "fail fast" or pivot strategy: a project that is unviable today due to missing data or unready processes is simply deprioritized, leaving open the possibility that it might become the best bet in the future. By continuously prioritizing projects with the highest risk-adjusted expected value and monetizing early, TELUS has successfully generated over $100 million annually in value over the past three to four years.

To illustrate this decision framework, consider a choice between two potential opportunities: Project A, valued at $1 million, and Project B, valued at $500,000. Under traditional decision-making, an organization would instinctively pursue Project A due to its larger headline number. However, if Project A represents a brand-new domain with unknown stakeholders, unmapped data, and unproven business processes, its odds of success align with the standard industry baseline of roughly one in five (20%), giving it a risk-adjusted expected value of only $200,000. Conversely, if Project B is within an established line of business—such as finance, where data schemas are well understood, stakeholder relationships are solid, and operational adoption is assured—the probability of success can easily reach 50%, resulting in an expected value of $250,000. Under this lens, Project B is the objectively superior project to pursue first.

```mermaid flowchart TD accTitle: Project Expected Value Prioritization subgraph ProjA ["Project A: High Value / High Uncertainty"] A1["Potential Value: $1,000,000"] A2["Unknown Data, Processes & Stakeholders (20% Odds)"] A3["Risk-Adjusted Expected Value: $200,000"] A1 --> A2 --> A3 end

subgraph ProjB ["Project B: Moderate Value / High Certainty"] B1["Potential Value: $500,000"] B2["Known Domain, Familiar Data & Partners (50% Odds)"] B3["Risk-Adjusted Expected Value: $250,000"] B1 --> B2 --> B3 end

A3 -.->|Lower Expected Return| Dec{"Decision"} B3 -->|Higher Expected Return| Dec Dec --> Choice["Prioritize Project B via Fail-Fast MVP"] ```

- **Annual AI Impact:** TELUS generates over $100 million annually through this disciplined portfolio approach. - **Expected Value Metric:** $E(V) = \text{Potential Value} \times P(\text{Success})$. - **Pivot Strategy:** Deprioritizing unready projects preserves capital while leaving the door open for future execution when conditions improve.

---

[11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) Case Study: Resolving Call Center Repeat Callers Through Early Proof of Value

Call center operations frequently suffer from repeat callers—customers who reach out again within a day, two days, or a week because their underlying issue was not resolved during the initial contact. At TELUS, repeat callbacks represented a massive $50 million annual operational expense. The business unit originally approached the AI team with a technical request: build a predictive machine learning model to identify which customers were likely to call back so that agents could intervene proactively. While a high-value opportunity, immediately embarking on model development would have required extensive data engineering, complex workflow modeling, and lengthy turnaround times, with no guarantee that frontline operations would ultimately adopt the solution.

Instead of jumping straight into machine learning development, the AI team initiated a root-cause investigation using simple descriptive analytics. By analyzing caller category breakdowns, they discovered that billing inquiries represented the largest driver of repeat contacts, and first-time bills accounted for the overwhelming majority of those disputes. Specifically, new customers were confused by activation fees appearing on their opening statements.

Armed with this "minimum viable insight" (proof of value), the team addressed the root problem directly without deploying complex algorithms: they enhanced call center agent training regarding fee disclosures and equipped retail store staff to explain the first bill directly to customers at the point of sale. This operational adjustment immediately recovered $3 million of the $50 million cost pool, while simultaneously clearing privacy approvals, establishing data pipelines, and earning critical stakeholder trust for downstream predictive modeling.

```mermaid flowchart LR accTitle: Incremental Value Realization in Call Centers subgraph Phase1 ["Descriptive Root-Cause Analysis"] D1["$50M Repeat Caller Problem"] --> D2["Analyze Call Reasons"] D2 --> D3["Identify First-Bill Activation Fee Confusion"] end

subgraph Phase2 ["Operational Fix & Early ROI"] O1["Update Agent Training & Retail Explaners"] O2["Capture $3M Immediate Value"] O3["De-Risk: Secure Privacy, Data & Stakeholder Buy-in"] D3 --> O1 --> O2 --> O3 end

subgraph Phase3 ["Advanced Machine Learning"] M1["Build Predictive Callback Model"] M2["Target Remaining $47M Opportunity"] O3 --> M1 --> M2 end ```

- **Problem Scope:** Repeat customer calls cost TELUS approximately $50 million per year. - **Initial Finding:** The primary catalyst for callbacks was first-billing confusion regarding account activation fees. - **Immediate Value Delivered:** Simple process updates and agent enablement captured $3 million in savings before a single machine learning model was trained. - **Foundational De-risking:** Early iterations secured stakeholder engagement, data access, and privacy approvals ahead of advanced AI deployment.

---

[13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) Shifting from Software Delivery to Strategic Business Consulting

Transforming an enterprise with AI requires moving away from a traditional IT software delivery mindset toward strategic business consulting. Rather than operating like a technology implementation firm building AI for its own sake, the TELUS AI organization functions more like McKinsey—focusing primarily on uncovering and understanding the fundamental business challenge before evaluating technical solutions.

The engagement process begins with design thinking workshops aimed at discovering problem spaces rather than mapping where AI might fit into existing workflows. Because modern AI handles both structured and unstructured data across enterprise environments, technical components naturally emerge once the problem is framed properly. In many scenarios, descriptive statistics, basic advanced analytics, or process re-engineering unlock substantial business value long before machine learning models enter production.

- **Consulting Mindset:** Focus on business problem diagnosis rather than tech-push delivery ("operating more like a McKinsey than an Accenture"). - **Design Thinking Principle:** Avoid asking where AI fits; identify broken processes and high-impact operational friction points first. - **Value Realization Hierarchy:** Descriptive statistics and workflow redesign often capture early baseline value prior to full model development.

---

[15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Five Core Principles for Enterprise AI Transformation

Successfully executing enterprise AI initiatives requires adhering to five strategic management practices that maximize business impact while minimizing wasted investment:

1. **Finance a Portfolio of Initiatives:** CFOs must transition from an individual project-tracking mindset to a venture capitalist mentality. Rather than gambling on single high-risk bets, organizations should fund a balanced portfolio, ensuring that the 20% of projects that succeed deliver maximum aggregate value. 2. **Fall in Love with the Problem, Not the Solution:** Teams must avoid being blinded by the desire to apply artificial intelligence everywhere. AI will inevitably play a role, but defining the correct business problem must precede any technology choice. 3. **Start Early and Avoid Overplanning:** Prolonged discovery phases aimed at theoretically mapping every possible risk before writing code waste critical time. Teams should launch rapidly to uncover real-world edge cases through live execution. 4. **Monetize Iteratively:** Reject long development cycles that promise value only upon final deployment of a high-accuracy model. Break initiatives into incremental milestones that yield proofs of value and minimum viable insights early. 5. **Differentiate Proof of Value from Proof of Concept:** While proofs of concept often serve merely to generate artificial hype, proofs of value establish measurable operational returns.

Alexandre Guilbault concludes with a call to action for enterprise leaders: "Let's stop chasing AI projects for the sake of applying AI. Let's make the cool useful."

## Notable quotes > [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) Only 20% of AI projects are successful in creating value in business. — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Let's stop chasing AI projects for the sake of applying AI. Let's make the cool useful. — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Finance a Portfolio of Initiatives — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Fall in Love with the Problem, Not the Solution — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Start Early and Avoid Overplanning — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Monetize Iteratively — Alexandre Guilbault

> [15:13](https://www.youtube.com/watch?v=NBUSofH1XHo&t=913s) Differentiate Proof of Value from Proof of Concept — Alexandre Guilbault

> [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) operating more like a McKinsey than an Accenture — Alexandre Guilbault

## Mentioned Alexandre Guilbault, TELUS, TELUS Health, TELUS Agriculture, Ai4, McKinsey, Accenture, Model Context Protocol

## Video description Alexandre Guilbault, Head of Artificial Intelligence, TELUS

Abstracts

In this 20-minute session, panelists will delve into frameworks and methodologies for quantifying the value of AI investments. We’ll explore strategies for aligning AI implementation with broader business objectives, and present real-world case studies that illustrate measurable returns on AI adoption across sectors.

[Source](https://www.youtube.com/watch?v=NBUSofH1XHo)

中文简报

TELUS借鉴风投模式评估项目风险，从AI中创收1亿美元 企业从人工智能中获取超额回报，并非依靠强行提高项目成功率，而是将项目管线作为动态概率加权投资组合进行管理，从而从早期的运营改进中实现现金收益。 Alexandre Guilbault 详细剖析了在全行业企业 AI 项目失败率高达 80% 的背景下，TELUS 如何扩展其 AI 实践并创造超 1 亿美元的年商业价值。他介绍了一种受多臂老虎机问题启发的动态投资组合管理模型，旨在平衡探索性运营支出与经风险调整后的预期价值，而非盲目追逐未经核实的表面数据。Guilbault 以一个呼叫中心案例展示了将“价值证明”置于投机性“概念验证”之上的效果——该案例在训练任何机器学习模型之前，仅通过基础的根本原因分析就立即节省了 300 万美元。本次演讲为企业领导者概述了五项核心原则，强调要从单纯的 IT 交付转向战略咨询，并敦促各类机构让 AI 具备实际效用，而不仅停留在技术层面的新奇感。 尽管业内仅约 20% 的企业级 AI 项目能创造可衡量的商业价值，TELUS 仍通过建立严格的治理体系将团队扩展至超百人规模，每年实现逾 1 亿美元的业务回报。 (0:04) 交付框架应与问题确定性相匹配：路径与目标已知采用瀑布模型，技术路径未知采用敏捷开发；而在目标与路径皆未知的场景下推行快速失败（Fail Fast），在第一天即对数据隐私或数据可用性等最薄弱环节展开压力测试。 (3:03) TELUS 借鉴多臂老虎机（Multi-Armed Bandit）模型，基于动态预期 ROI（预估价值乘以成功概率）持续重配资金，优先推进经风险调整后预期回报更高、成熟度更明确的业务项目。 (7:20) 在每年耗资 5000 万美元的客服重复来电挑战中，团队在训练任何机器学习模型之前，仅通过根因描述性分析与前线流程优化，便立即锁定了 300 万美元的运营成本节约。 (11:18)

Reading summary

Enterprises struggle to justify artificial intelligence investments because roughly 80% of projects fail to generate business value, placing chief financial officers between acute risk aversion and fear of falling behind. Alexandre Guilbault details how TELUS built an AI practice generating over $100 million in annual value by shifting from speculative, large-scale deployments to dynamic portfolio management and incremental proofs of value. Aligning Delivery Frameworks to Project Uncertainty (3:03) Guilbault argues that project delivery must match the level of business and technical certainty. Traditional Waterfall delivery remains viable when both the objective and implementation path are clear, such as building an internal enterprise Retrieval-Augmented Generation (RAG) chatbot against curated internal documents. When the end goal is defined but technical execution is evolving—such as configuring agentic workflows and Model Context Protocol (MCP) architectures—iterative Agile sprints are required. For ambiguous initiatives where neither the destination nor the technical path is established, organizations should use a "fail fast" approach. Instead of spending months building models, teams immediately isolate and stress-test the project's single weakest operational link on day one, such as regulatory compliance, personal data privacy constraints, or basic data availability. Prioritizing Value Through Expected ROI Models (7:20) Rather than treating project viability as a static decision to either fund or kill an initiative, TELUS models project selection on the multi-armed bandit problem from data science. Teams balance exploration against exploitation by calculating an initiative's dynamic expected value: estimated total return multiplied by the updated probability of success. Stalled projects facing missing data or unready processes are deprioritized rather than abandoned, preserving capital while leaving room to revisit them later. This framework prevents organizations from chasing large headline numbers that carry high execution risk. For example, a $1 million opportunity with unmapped data and an industry-standard 20% success rate yields an expected value of $200,000. By contrast, a $500,000 project in an established operational area with clean data and known stakeholders carries a 50% success probability, yielding a superior expected value of $250,000. Managing this portfolio requires maintaining a diversified pipeline across departments and providing dedicated operational expenditure (OPEX) for early discovery. De-Risking Initiatives Through Early Operational Wins (11:18) To minimize delivery risk, TELUS emphasizes shifting from abstract "proofs of concept" to operational "proofs of value" that capture revenue or cost savings early. In a TELUS case study, repeat customer callbacks cost the organization $50 million annually. Frontline stakeholders initially requested a predictive machine learning model to flag repeat callers for proactive outreach, which would have required extensive engineering and change management before showing any return. Instead, the AI team ran simple descriptive analytics on call categories and discovered that new customer confusion regarding first-bill activation fees caused the bulk of repeat inquiries. By updating agent training and equipping retail staff to clarify activation fees at the point of sale, TELUS immediately eliminated $3 million in operational costs without deploying a machine learning model. This early win established data pipelines, cleared privacy requirements, and built frontline trust to pursue the remaining $47 million opportunity with advanced predictive models. Operating as a Strategic Business Consultancy (13:55) Guilbault asserts that internal enterprise AI groups must function like strategic consultancies rather than pure IT delivery shops. The team conducts design thinking workshops to isolate core business friction points before considering whether machine learning is required, as descriptive analytics or process redesign often solve problems faster than algorithmic modeling. TELUS structures its enterprise AI strategy around five management principles: financing a venture-capital-style portfolio rather than isolated bets; focusing on the business problem rather than the technological tool; launching rapidly to encounter live operational edge cases instead of overplanning; monetizing incrementally; and prioritizing operational value creation over technical demonstrations. Guilbault concludes that maximizing enterprise AI returns requires moving away from novelty and focusing strictly on practical utility. The model relies heavily on corporate willingness to fund exploratory OPEX and continuously deprioritize unready projects, acknowledging that high baseline failure rates can still yield massive financial returns under disciplined portfolio management.

精炼阅读摘要

TELUS的Alexandre Guilbault指出，尽管业界普遍面临AI项目仅有20%成功创造价值的现实，但企业无需盲目追求整体胜率的跃升，而应通过动态投资组合管理、以期望回报率排序优先级并践行小步快跑的“价值验证”，化解财务部门在风险厌恶与落后焦虑之间的拉锯，从而实现年化超1亿美元的业务收益。 交付框架匹配：瀑布、敏捷与快速失败 (3:03) Guilbault指出，各类交付方法并无绝对优劣，关键在于技术路径与业务目标的确定性。传统的瀑布模式（Waterfall）适用于目标和路径均完全明确的场景，例如基于经过整理的内部文档构建企业检索增强生成（RAG）聊天机器人；当业务目标明确但技术实现路径充满不确定性时，则应采用敏捷模式（Agile），通过迭代冲刺构建多代理系统或配置模型上下文协议（MCP）。但在大量业务与技术双重未知的AI探索场景中，团队必须转向“快速失败”（Fail Fast）模式。 在快速失败模式下，团队虽然会预估理论投资回报率，但其核心是在项目启动首日即对“最脆弱环节”进行极限施压，而非耗费半年开发后再验证。例如在超个性化营销项目中，涉及敏感个人隐私或监管合规的风险可能直接阻断项目，第一天验证合规边界能避免后续无效投入；在未被系统记录的业务流程中，确认是否存在可用历史数据必须先于任何模型架构设计。 动态期望ROI：多臂老虎机式的组合投资 (7:20) TELUS突破了非生即死的二元项目关停模式，借鉴数据科学中平衡探索与利用的“多臂老虎机”机制，对项目投资回报率实施动态调整。团队将动态期望回报率定义为预估总回报与即时成功概率的乘积。当探索性工作发现底层数据缺陷或快速获益点时，成功概率会随之调整，从而动态决定是继续开发高胜率项目，还是转向储备池中的其他方案。 Guilbault通过对比指出：一个潜在价值100万美元但处于全新业务领域的项目，因利益相关方未知和数据缺失，成功率仅有行业基准的20%，其风险调整后的期望价值为20万美元；相反，一个位于财务等成熟部门、潜在价值50万美元的项目，因数据结构清晰且合作成熟，成功率可达50%，其期望价值为25万美元，因而具备更高的执行优先级。有效运转该机制需要三大前提：跨业务部门的多元项目池、专门支持早期探索性实验的运营支出（OPEX）预算，以及将庞大业务难题拆解为可阶段性变现模块的能力。 从POC转向POV：呼叫中心减负案例 (11:18) 在解决年耗费约5000万美元的呼叫中心客户重复来电问题时，业务部门最初希望AI团队构建预测性机器学习模型以提前干预。然而，AI团队并未直接启动复杂的模型研发，而是先通过基础的描述性统计展开归因分析。分析发现，大量重复来电属于账单咨询，其中绝大多数是新客户对首期账单中的激活费产生困惑。 基于这一“最小可行洞察”，团队并未上线任何高级算法，而是通过强化客服话术并让线下零售门店在销售环节主动向客户解释首期账单，直接消除了混乱来源。该纯运营手段迅速挽回了300万美元的成本损失。更重要的是，这一过程以实打实的“价值验证”（Proof of Value, POV）取代了虚浮的“概念验证”（Proof of Concept, POC），不仅在低风险下回收了早期投资，还顺带完成了隐私审批、搭建了底层数据管线，为后续针对剩余4700万美元机会开发预测模型奠定了坚实基础。 转型战略咨询：驱动企业AI转型的核心准则 (13:55) Guilbault主张企业AI团队应当从传统软件交付模式转向战略商业咨询定位，以类似麦肯锡而非纯技术外包的方式运作。团队应借助设计思维工作坊优先界定痛点与低效环节，而非预设技术方案去生搬硬套。现代AI兼具处理结构化与非结构化数据的能力，只要核心业务问题界定精准，技术路径便会自然浮现，且基础统计与流程重构往往能在模型投产前就释放巨大价值。 结合TELUS的实践，Guilbault归纳了企业AI转型的五项核心准则：第一，CFO应以风险投资（VC）组合思维资助项目，依靠胜出的20%高价值项目实现整体超额回报；第二，专注于解决实际问题而非盲目迷信算法方案；第三，尽早启动以在实战中捕捉边缘案例，避免过度规划导致的时间浪费；第四，分拆里程碑以实现递进式变现；第五，坚决剔除用于制造声量的概念验证，专注于能够量化业务回报的价值验证，让前沿技术真正转化为实用生产力。 Guilbault的论证展示了成熟企业如何通过运营框架与预算模式的转型跨越“AI投资悖论”，但该方案有效运转的前提在于企业需具备跨多元产业的丰富业务场景，以及能够支撑试错的专属探索性OPEX预算；对于业务单一或预算受传统资本支出严格管控的组织，动态组合调度的实际落地仍面临制度性挑战。

Detailed notes — zh

## Background 本视频由 Ai4 于 2025 年 8 月 24 日发布在 YouTube 平台上，主讲嘉宾为 TELUS 人工智能负责人 Alexandre Guilbault。视频内容围绕量化 AI 投资价值的框架与方法论展开，深入探讨了如何将 AI 项目与企业整体商业目标对齐，并通过跨领域实战案例阐释可衡量的投资回报率（ROI）。 ## TL;DR 针对业内仅有 20% 的 AI 项目能产生实际商业价值的残酷现状，TELUS 分享了其团队十年来如何通过 AI 实现每年超 1 亿美元商业回报的实战体系。演讲剖析了从瀑布模式、敏捷开发到快速失败（Fail Fast）的范式演变，提出利用多臂老虎机（Multi-Armed Bandit）与风投思维进行动态期望价值评估和资源仲裁。通过一个耗费 5000 万美元的客服中心重复呼叫案例，团队展示了如何不写一行 AI 代码、仅凭基础数据洞察与业务流程修复便提前斩获 300 万美元收益并建立业务互信。最后，主讲人提炼出以投资组合管理、聚焦真实痛点、及早变现为核心的五大转型法则，主张以价值验证（POV）替代盲目炒作（POC）。 ## Atmosphere & style 整场演讲呈现出冷静务实且极具穿透力的战略咨询底色，彻底褪去了当下 AI 领域的浮躁狂热与术语堆砌。演讲者以坦诚锐利的姿态直击 CFO 的风险规避心理与行业低成功率痛点，情绪曲线由冰冷的现实审视逐步转向清晰明朗的破局解法。其叙事兼具高管顶层设计的从容与业务前线摸爬滚打的笃定，通过生动的案例拆解化繁为简，带给听众一种豁然开朗的务实感与强大的实战信心。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) AI 投资面临严峻落地困境：全行业仅约 20% 的 AI 项目能产生实际商业回报，CFO 与决策层往往在极度风险规避与错失恐惧（FOMO）的张力之间艰难推进。 - [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) TELUS 通过兼顾人员、流程与技术，十年间将 AI 团队从 2-3 人拓展至 100 余人规模，如今每年通过 AI 稳定创造超过 1 亿加元/美元的直接商业效益。 - [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) AI 项目管理应因地制宜：RAG 等成熟场景适合瀑布模式，智能体架构适合敏捷迭代，而终点与路径皆未知的复杂痛点则应采用快速失败（Fail Fast）模式。 - [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) Fail Fast 的精髓在于首日直击“最脆弱环节”：在项目启动第一天就对隐私合规或关键技术假设进行压力测试，避免耗费半年开发后才被法务或业务叫停。 - [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) 引入多臂老虎机（Multi-Armed Bandit）机制：摆脱单一静态 ROI 评估，依据“预期收益 × 动态成功概率”的期望价值在探索（Exploration）与利用（Exploitation）间动态调配资源。 - [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) 动态资源分配的三大前提：企业必须拥有丰富的业务项目组合、划拨专属日常运营支出（OPEX）支持试错探索，并坚持及早进行阶段性商业变现。 - [09:45](https://www.youtube.com/watch?v=NBUSofH1XHo&t=585s) 以风投思维评估期望价值：高潜在价值但在陌生领域、成功率仅 20% 的项目，其优先级应让位于深耕领域、数据成熟且期望回报更高的项目。 - [11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) 拆解客服中心 5000 万美元重复呼叫痛点：团队未盲目从头构建端到端预测模型，而是先通过简单数据分析定位“首期账单激活费困惑”这一核心根因。 - [11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) 践行“最小可行洞察（MVI）”提前变现：在未写一行 AI 代码的情况下，仅靠改进坐席话术与门店说明便挽回 300 万美元损失，并借此打通数据合规与业务信任。 - [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) AI 团队需转型为战略咨询角色：采用设计思维深挖真实业务痛点，通过基础分析修复破碎的业务流程，其价值往往远胜盲目堆砌算法模型。 - [15:14](https://www.youtube.com/watch?v=NBUSofH1XHo&t=914s) 落实 AI 商业转型的五大核心法则：以投资组合分散风险、聚焦真实痛点而非技术本身、尽早启动避免过度规划、敏捷迭代变现，并从概念验证（POC）转向价值验证（POV），让技术真正实用化。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) AI 落地困境与 TELUS 的规模化经验 TELUS 作为一家在加拿大乃至全球知名的企业，不仅提供移动通信、家庭解决方案和安防服务，还跨足了 TELUS Health（医疗健康）、TELUS Agriculture（农业）以及全球 AI IT 解决方案等多元化领域。在如此广泛的业务版图中，集中部署 AI 面临着巨大的机遇与挑战。AI 必将定义当下的十年，但一个不容忽视的残酷现实是：目前业内只有 20% 的 AI 项目能够真正为企业创造商业价值。 ```mermaid flowchart LR accTitle: AI 投资中的高管决策困境 subgraph Dilemma["CFO 的心理博弈"] Risk["极度风险规避<br/>(仅20%成功率/五中选一)"] <--> FOMO["错失恐惧症 FOMO<br/>(害怕成为最后一个拥抱 AI 的企业)"] end Dilemma --> Investment["艰难推进投资与立项"] ``` 如果向首席财务官（CFO）汇报一个项目并直言其“只有五分之一的概率能带来投资回报（ROI）”，几乎没有 CFO 愿意掏钱；但与此同时，也没有任何 CFO 愿意在 AI 浪潮中落后于人。这种在“风险规避”与“错失恐惧（FOMO）”之间的剧烈张力，构成了当下 AI 投资的核心矛盾。TELUS 在过去十年中亲历了这一过程：团队从最初的 2 到 3 名机器学习工程师，一路扩张至超过 100 人的规模，如今每年通过 AI 产生超过 1 亿加元的商业效益。成功并非一蹴而就，任何颠覆性的 AI 落地都必须兼顾“人员（People）、流程（Process）与技术（Technology）”三要素，克服员工对新技术的抗拒心理以及复杂业务流程和数据链路带来的挑战。 - **TELUS 业务矩阵**：覆盖电信通信、智慧家庭、医疗健康（TELUS Health）、智慧农业（TELUS Agriculture）及全球 AI 解决方案。 - **团队规模演进**：过去 10 年间从 2~3 名 ML 工程师成长为 100+ 人的专业团队。 - **商业化价值输出**：每年为公司稳定创造超过 1 亿美元/加元的直接效益。 - **行业痛点**：全行业 AI 项目仅有约 20% 能产生实际业务回报。 --- ### [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) 项目管理范式演变：从 Waterfall、Agile 到 Fail Fast 要理解 AI 项目如何创造价值，首先需要梳理主流的项目管理方法论。不同的项目管理方式并非“非此即彼”的淘汰关系，即便在当今的 AI 场景下，每种方法也都有其适用的土壤： ```mermaid flowchart TD accTitle: 三种项目管理方法论对比 subgraph Waterfall["瀑布模式 (Waterfall)"] W1["明确终点 (Destination)"] --> W2["清晰路径 (Path)"] W2 --> W3["前期测算 ROI 并全面资助"] end subgraph Agile["敏捷模式 (Agile)"] A1["明确终点 (Destination)"] --> A2["未知路径 (Unknown Path)"] A2 --> A3["小步快跑、迭代反馈"] end subgraph FailFast["快速失败 (Fail Fast)"] F1["未知终点 (Unknown Destination)"] --> F2["未知路径 (Unknown Path)"] F2 --> F3["首日直击最脆弱环节 (Weakest Link)"] end ``` 1. **瀑布模式（Waterfall）**：适用于“目标明确且实现路径清晰”的场景。团队能够预先精确测算 ROI、识别并缓解潜在风险，随后获得全额预算。 2. **敏捷模式（Agile）**：适用于“目标明确但实现路径未知”的软件交付场景。通过增量开发与快速反馈，团队在行进过程中不断调整优化路线。 3. **快速失败（Fail Fast）**：在 AI 领域中，大量项目处于“终点未知且路径未知”的完全摸黑状态。Fail Fast 的核心思想是优先找出项目的“最脆弱环节（Weakest Link）”，并在第一天而非耗费六个月后去对其进行压力测试。如果一个营销项目涉及极度敏感的个人身份信息（PII），存在极大的隐私或合规被否决风险，就应优先验证合规边界，而不是花大半年去洗数据、建模型、调流程最后才被法务叫停。 - **RAG 知识库问答**：更适合 **瀑布模式**，因为技术成熟度高、已有企业文档明确，目标和实施步骤清晰。 - **智能体架构开发（Agentic Systems）**：更偏向 **敏捷模式**，本质上是软件工程问题（例如是否引入 MCP、MCP 服务器托管位置、模型选型探索等）。 - **复杂业务痛点攻坚**：如客服中心的大量重复来电（Repeat Calls）问题，由于不知道最终形态，最适合采用 **Fail Fast** 模式。 --- ### [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) 进阶机制：基于 Multi-Armed Bandit 的探索与利用平衡 Fail Fast 机制虽然有效，但其重点在于“砍掉项目”。TELUS 在此基础上更进一步，引入了数据科学中经典的多臂老虎机（Multi-Armed Bandit）思想，将管理重点从单纯的“终止项目”转变为“在探索（Exploration）与利用（Exploitation）之间实现动态平衡”。 ```mermaid flowchart LR accTitle: 动态 ROI 驱动的资源分配模型 subgraph Inputs["输入要素"] ROI["初始预期 ROI"] Prob["动态成功概率 P(Success)"] end Inputs --> Model["动态期望收益评估<br/>Expected Value = ROI × P(Success)"] Model --> Decision{"动态优先级仲裁"} Decision -->|"高确定性收益"| Exploit["利用 (Exploitation)<br/>全速推进并早期变现"] Decision -->|"高不确定性潜力"| Explore["探索 (Exploration)<br/>OPEX 试错并测试弱点"] Decision -->|"劣质/死胡同"| Shift["无痛转移资源至备选方案"] ``` 在传统的项目管理中，ROI 通常是一个固定不变的静态数字；而在该框架下，ROI 是动态更新的——即由“预期商业价值 × 随认知演化的成功概率”计算所得。随着对风险和技术可行性的逐步摸清，成功概率会实时调整，团队据此动态决定是继续追加投入，还是平滑地将资源切换到其他替代方案。落地该方法论需要三大核心先决条件： - **优质项目组合（Project Portfolio）**：如果手中只有一个项目，该方法毫无用武之地；企业必须梳理出能够被 AI 赋能的丰富业务管线。 - **探索阶段的专属资金（Exploratory Funding / OPEX）**：不能只资助已具备资本化条件的成熟项目，必须划拨日常运营支出（OPEX）专门用于前期探索和弱点验证。 - **早期商业化变现（Monetize Early）与 Proof of Value**：告别等到全流程交付才验证价值的传统 POC（概念验证）思维，转而以 POV（价值验证，Proof of Value）拆解大问题，在最短时间内实现阶段性价值变现以对冲风险。 ### [09:45](https://www.youtube.com/watch?v=NBUSofH1XHo&t=585s) 评估成功概率：告别单点押注的“风险投资”思维 企业在评估人工智能项目的投资回报率（ROI）时，不仅要看预期的绝对收益，还必须结合项目的成功概率来综合决策。如果某个项目当前条件不成熟，理性的做法是果断降低其优先级并快速调整方向（pivot），而不是一条路走到黑；等到未来数据或业务条件完备时，它可能又会重新成为最佳选择。这种动态评估、及早变现并不断循环迭代的策略，正是 TELUS 在过去三到四年间能够每年通过 AI 稳定创造超过 1 亿美元业务价值的核心秘诀。 ```mermaid flowchart TD accTitle: 项目期望价值评估对比流程 subgraph S1["传统决策视角 (仅看潜在价值)"] A1["项目 A<br>潜在价值: $1,000,000"] -->|表面回报更高| R1["优先立项并全力投入"] B1["项目 B<br>潜在价值: $500,000"] end subgraph S2["加权成功率视角 (风险投资思维)"] A2["项目 A<br>全新领域 / 陌生数据 / 利益方未知<br>成功率: 20%"] --> E1["期望价值: $200,000"] B2["项目 B<br>熟悉财务业务 / 团队互信 / 数据成熟<br>成功率: 50%"] --> E2["期望价值: $250,000"] E2 -->|期望回报更高且风险可控| R2["首选推进项目 B (快速见效)"] end ``` 通过一个直观的对比案例能够清晰说明这种逻辑差异：假设手头有两个备选项目，项目 A 潜在价值 100 万美元，项目 B 潜在价值 50 万美元。按照传统思维，团队必定毫不犹豫地选择项目 A。然而，如果项目 A 是一个全新的领域，团队不熟悉业务流程、没有现成的数据资产、也不了解业务干系人，从零起步的成功几率可能仅有五分之一（20%），其实际期望价值折算下来仅为 20 万美元；反观项目 B，如果它处在团队深耕已久的领域（例如财务部门），数据质量高且与业务方同事配合默契，其成功率可以达到 50%，折算期望价值为 25 万美元。在这种评估体系下，团队应当果断锁定项目 B 作为主攻方向，通过快速试错（fail fast）与快速迭代的方式启动推进。 --- ### [11:18](https://www.youtube.com/watch?v=NBUSofH1XHo&t=678s) 拆解复杂业务：客服中心 5000 万美元重复呼叫案例 面对宏大的业务痛点，成功的关键在于将其拆解为细小且可快速验证的环节，在逐步提升胜率的同时提前兑现商业价值。TELUS 的客服中心曾面临一个普遍存在的痛点——用户重复拨打电话（repeat callers）。很多用户在致电后的一两天甚至一周内再次来电，这说明他们的问题在首次通话中并未得到解决；光是处理这些重复呼叫，每年就给公司带来高达 5000 万美元的额外成本。业务部门最初提出的需求非常典型：“你们能不能建一个预测模型，预测谁有可能再次拨打电话？这样我们就可以提前主动联系客户。” ```mermaid flowchart LR accTitle: 客服中心重复呼叫问题拆解与分步价值兑现 subgraph Step1["问题拆解与根因分析"] P1["分析呼叫原因饼图"] --> P2["定位首要矛盾: 计费问题"] P2 --> P3["深挖根因: 首次账单激活费"] end subgraph Step2["最小可行洞察 (MVI) 落地"] P3 --> A1["加强坐席话术培训"] P3 --> A2["线下门店交付账单说明"] A1 & A2 --> V1["提前锁定 $3,000,000 收益"] end subgraph Step3["AI 深度建模阶段"] V1 --> M1["数据与隐私审批就绪"] M1 --> M2["业务方深度信任协同"] M2 --> M3["构建主动预测干预模型"] end ``` 尽管这是一个绝佳的业务场景，但团队并没有立即着手搭建复杂的机器学习模型。因为从头开发模型耗时漫长，数据准备、流程理解等环节充满不确定性，且业务部门往往容易产生变革阻力甚至中途转移优先级，导致耗费巨资打造的模型最终被束之高阁。团队选择反其道而行之，先从最基础的数据探索入手——通过一张简单的呼叫原因饼图，发现大量重复呼叫集中在计费（billing）环节；进一步下钻后发现，最大的症结在于“第一期账单”（first billing），客户普遍对账单中的账户激活费（activation fee）感到困惑与不满。 基于这一“最小可行洞察”（Minimum Viable Insight），团队在没有写一行 AI 算法代码的情况下迅速采取了业务干预措施： - 优化客服坐席的话术培训，确保在开户时向客户清晰沟通激活费用的构成。 - 联动线下零售门店，在客户办理业务现场提供第一期账单的详细书面说明。 - 仅凭这两项基础的流程改进，就在 5000 万美元的总损失中直接挽回了 300 万美元的成本。 - 在此过程中，AI 团队不仅提前兑现了商业价值，更顺利完成了数据治理、隐私合规审批以及业务干系人的深度绑定，为后续引入高级预测模型奠定了坚实信任基础。 --- ### [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) 战略咨询导向：从设计思维到流程修复 在推动企业级 AI 转型时，必须摒弃传统的“软件交付型”思维。AI 项目的本质不是为了炫耀技术，而是为了解决实际业务问题。TELUS 的 AI 团队将自身定位为“业务战略咨询机构”，其运作模式更偏向麦肯锡（McKinsey）而非传统的系统集成商埃森哲（Accenture）——核心目标不是为了做 AI 而做 AI，而是寻找最值得用 AI 赋能的真实业务场景。 项目通常以设计思维（Design Thinking）作为起点，深入挖掘业务痛点。在运用设计思维探索时，切忌一开始就陷入“AI 到底能插在哪个环节”的思维陷阱。当前的人工智能技术极其强大，无论是处理结构化还是非结构化数据都游刃有余，随着企业数据资产的不断丰富，AI 迟早会在各个业务链路中发挥作用；但如果过早聚焦于 AI 技术本身，反而会蒙蔽团队的视野。在很多情况下，通过高级分析（Advanced Analytics）跑出一些基础统计数据，或者通过战略咨询手段修复一个原本支离破碎的业务流程（fixing a broken process），其所创造的商业价值可能远比盲目堆砌算法模型要大得多。 --- ### [15:14](https://www.youtube.com/watch?v=NBUSofH1XHo&t=914s) 实现 AI 商业转型的五大核心法则：“让酷炫变为实用” 在分享的最后，演讲者结合 TELUS 的实战经验总结了推动 AI 业务转型的五大核心法则，呼吁行业转变对 AI 项目的评估与落地模式： ```mermaid flowchart TD accTitle: 推动有意义的 AI 商业转型的五大核心法则 L1["1. 组合式投资<br>像风投一样思考，分散风险并扩大优质产出"] L2["2. 聚焦真实痛点<br>深挖高价值问题，切勿迷恋单一技术方案"] L3["3. 快速启动实践<br>避免过度规划，在实践中探索未知风险"] L4["4. 敏捷迭代变现<br>拒绝等待最终模型，提早拆解并兑现收益"] L5["5. 拒绝纯粹炒作<br>从 POC 转向 POV，让酷炫的技术真正创造价值"] L1 --> L2 --> L3 --> L4 --> L5 ``` - **以投资组合视角进行资金分配（Finance a portfolio of initiatives）**：首席财务官（CFO）需要改变逐个项目死扣预算、进度和风险的传统会计思维，像风险投资家（Venture Capitalist）而非赌徒那样管理项目池。不要把所有筹码押在一个项目上；假设整体成功率只有 20%，核心目标应是从这成功的 20% 项目中榨取最大的商业价值。 - **从好问题出发，切勿迷恋解决方案（Start with a good problem, don't fall in love with the solution）**：不要被“AI 无所不能、必须随处应用”的狂热冲昏头脑，找准真正痛点，AI 自然会在适当时机登场发挥作用。 - **尽早启动，避免过度规划（Start early, avoid overplanning）**：许多企业动辄耗费数月时间去推演和罗列 AI 带来的所有潜在风险，但在项目尚未启动前，团队往往“不知道自己不知道什么”。唯有尽早下场实践，才能在探索中学习和纠偏。 - **小步快跑，迭代式变现（Monetize iteratively）**：绝不接受“等模型精度达到某种指标后才能创造价值”的空头支票。团队必须将庞大的构想拆解为立即可验证的价值证明（Proof of Value）和最小可行洞察，分阶段持续回收投资。 - **告别炒作，让酷炫变为实用（Focus on Proof of Value over Hype）**：概念验证（Proof of Concept, POC）往往只是为了制造话题与炒作（hype），而企业真正需要的是价值证明（POV）。"让我们停止为了应用 AI 而追逐 AI 项目，把酷炫的技术变成真正有用的价值（Let's make the cool useful）。" ## Notable quotes > [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) 目前业内只有 20% 的 AI 项目能够真正为企业创造商业价值。 — Alexandre Guilbault > [00:04](https://www.youtube.com/watch?v=NBUSofH1XHo&t=4s) 如果向首席财务官（CFO）汇报一个项目并直言其“只有五分之一的概率能带来投资回报（ROI）”，几乎没有 CFO 愿意掏钱；但与此同时，也没有任何 CFO 愿意在 AI 浪潮中落后于人。 — Alexandre Guilbault > [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) Fail Fast 的核心思想是优先找出项目的“最脆弱环节（Weakest Link）”，并在第一天而非耗费六个月后去对其进行压力测试。 — Alexandre Guilbault > [03:03](https://www.youtube.com/watch?v=NBUSofH1XHo&t=183s) 如果一个营销项目涉及极度敏感的个人身份信息（PII），存在极大的隐私或合规被否决风险，就应优先验证合规边界，而不是花大半年去洗数据、建模型、调流程最后才被法务叫停。 — Alexandre Guilbault > [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) 在传统的项目管理中，ROI 通常是一个固定不变的静态数字；而在该框架下，ROI 是动态更新的——即由“预期商业价值 × 随认知演化的成功概率”计算所得。 — Alexandre Guilbault > [07:20](https://www.youtube.com/watch?v=NBUSofH1XHo&t=440s) 不能只资助已具备资本化条件的成熟项目，必须划拨日常运营支出（OPEX）专门用于前期探索和弱点验证。 — Alexandre Guilbault > [09:45](https://www.youtube.com/watch?v=NBUSofH1XHo&t=585s) 如果某个项目当前条件不成熟，理性的做法是果断降低其优先级并快速调整方向（pivot），而不是一条路走到黑；等到未来数据或业务条件完备时，它可能又会重新成为最佳选择。 — Alexandre Guilbault > [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) 核心目标不是为了做 AI 而做 AI，而是寻找最值得用 AI 赋能的真实业务场景。 — Alexandre Guilbault > [13:55](https://www.youtube.com/watch?v=NBUSofH1XHo&t=835s) 在很多情况下，通过高级分析（Advanced Analytics）跑出一些基础统计数据，或者通过战略咨询手段修复一个原本支离破碎的业务流程（fixing a broken process），其所创造的商业价值可能远比盲目堆砌算法模型要大得多。 — Alexandre Guilbault > [15:14](https://www.youtube.com/watch?v=NBUSofH1XHo&t=914s) 以投资组合视角进行资金分配（Finance a portfolio of initiatives）：首席财务官（CFO）需要改变逐个项目死扣预算、进度和风险的传统会计思维，像风险投资家（Venture Capitalist）而非赌徒那样管理项目池。 — Alexandre Guilbault > [15:14](https://www.youtube.com/watch?v=NBUSofH1XHo&t=914s) 从好问题出发，切勿迷恋解决方案（Start with a good problem, don't fall in love with the solution）。 — Alexandre Guilbault > [15:14](https://www.youtube.com/watch?v=NBUSofH1XHo&t=914s) 让我们停止为了应用 AI 而追逐 AI 项目，把酷炫的技术变成真正有用的价值（Let's make the cool useful）。 — Alexandre Guilbault ## Mentioned Alexandre Guilbault, TELUS, TELUS Health, TELUS Agriculture, Ai4, McKinsey, Accenture, Multi-Armed Bandit, RAG, Agentic Systems, MCP, Proof of Value (POV), Proof of Concept (POC) ## Video description Alexandre Guilbault, Head of Artificial Intelligence, TELUS Abstracts In this 20-minute session, panelists will delve into frameworks and methodologies for quantifying the value of AI investments. We’ll explore strategies for aligning AI implementation with broader business objectives, and present real-world case studies that illustrate measurable returns on AI adoption across sectors. [Source](https://www.youtube.com/watch?v=NBUSofH1XHo)

Source quotes

[0:04] Alexandre Guilbault: 目前业内只有 20% 的 AI 项目能够真正为企业创造商业价值。 [0:04] Alexandre Guilbault: 如果向首席财务官（CFO）汇报一个项目并直言其“只有五分之一的概率能带来投资回报（ROI）”，几乎没有 CFO 愿意掏钱；但与此同时，也没有任何 CFO 愿意在 AI 浪潮中落后于人。 [3:03] Alexandre Guilbault: Fail Fast 的核心思想是优先找出项目的“最脆弱环节（Weakest Link）”，并在第一天而非耗费六个月后去对其进行压力测试。 [3:03] Alexandre Guilbault: 如果一个营销项目涉及极度敏感的个人身份信息（PII），存在极大的隐私或合规被否决风险，就应优先验证合规边界，而不是花大半年去洗数据、建模型、调流程最后才被法务叫停。 [7:20] Alexandre Guilbault: 在传统的项目管理中，ROI 通常是一个固定不变的静态数字；而在该框架下，ROI 是动态更新的——即由“预期商业价值 × 随认知演化的成功概率”计算所得。 [7:20] Alexandre Guilbault: 不能只资助已具备资本化条件的成熟项目，必须划拨日常运营支出（OPEX）专门用于前期探索和弱点验证。 [9:45] Alexandre Guilbault: 如果某个项目当前条件不成熟，理性的做法是果断降低其优先级并快速调整方向（pivot），而不是一条路走到黑；等到未来数据或业务条件完备时，它可能又会重新成为最佳选择。 [13:55] Alexandre Guilbault: 核心目标不是为了做 AI 而做 AI，而是寻找最值得用 AI 赋能的真实业务场景。 [13:55] Alexandre Guilbault: 在很多情况下，通过高级分析（Advanced Analytics）跑出一些基础统计数据，或者通过战略咨询手段修复一个原本支离破碎的业务流程（fixing a broken process），其所创造的商业价值可能远比盲目堆砌算法模型要大得多。 [15:14] Alexandre Guilbault: 以投资组合视角进行资金分配（Finance a portfolio of initiatives）：首席财务官（CFO）需要改变逐个项目死扣预算、进度和风险的传统会计思维，像风险投资家（Venture Capitalist）而非赌徒那样管理项目池。 [15:14] Alexandre Guilbault: 从好问题出发，切勿迷恋解决方案（Start with a good problem, don't fall in love with the solution）。 [15:14] Alexandre Guilbault: 让我们停止为了应用 AI 而追逐 AI 项目，把酷炫的技术变成真正有用的价值（Let's make the cool useful）。

Source quotes

[0:04] Alexandre Guilbault: Only 20% of AI projects are successful in creating value in business. [15:13] Alexandre Guilbault: Let's stop chasing AI projects for the sake of applying AI. Let's make the cool useful. [15:13] Alexandre Guilbault: Finance a Portfolio of Initiatives [15:13] Alexandre Guilbault: Fall in Love with the Problem, Not the Solution [15:13] Alexandre Guilbault: Start Early and Avoid Overplanning [15:13] Alexandre Guilbault: Monetize Iteratively [15:13] Alexandre Guilbault: Differentiate Proof of Value from Proof of Concept [13:55] Alexandre Guilbault: operating more like a McKinsey than an Accenture

Original transcript

[0:04] I'm from a company. I'll explain my [0:05] company quickly. It's not a salt pitch [0:07] but it will be important to understand [0:08] the talk. Uh I come from a company [0:10] everybody knows in Canada but might not [0:12] here. Uh it's a t company. So we we [0:15] deliver mobile uh home solution security [0:18] solution but we expanded into various [0:20] other lines of business. So we have [0:22] Telis health, we have Telis agriculture, [0:24] we have Telus IT uh AI solutions all [0:27] around the world. So, we're really a [0:29] multi- uh industry company and I have [0:32] the chance to sit in the middle of all [0:33] of those industries to deploy AI uh in [0:37] all of those fields. Um I'll dive right [0:39] into it because we don't have a lot of [0:41] time. A lot of things I can cover today. [0:43] Um AI will define this decade. You all [0:47] agree that it has tremendous potential. [0:49] That's the reason why you're here. So, I [0:51] don't need to convince you about this. [0:52] But why are companies all failing to [0:54] create value with AI? [0:57] Only 20% of AI projects are successful [1:00] in creating value in business. Imagine [1:03] pitching a project to a CFO and telling [1:06] it telling him it's him or her. That you [1:10] only have one chance out of five that [1:12] you'll be able to have a return on your [1:13] investment. No CFO would like to fund [1:16] that. But at the same time, no CFO wants [1:18] to be the left the last one enabling AI [1:21] for the organization. So it creates a [1:24] tension between risk aversion and fear [1:26] of missing out. At Telus, we've been [1:29] through this journey for the past [1:31] decade. Uh we grew our team from just [1:34] two three machine learning engineers to [1:36] now over 100 and we're generating over [1:38] 100 million in benefits every single [1:40] year. Uh we're transforming every [1:43] business line, every operation with AI. [1:45] um it wasn't a tra a straight path. So [1:47] we learned a lot of things through the [1:49] way and we went through multiple [1:51] hurdles. Uh so what I would like to [1:52] share today is why AI projects fail. Uh [1:56] what we learned about that and how can [1:58] you adapt your projects to it. So why AI [2:01] project failed? Well, there's various [2:03] reasons. Um AI is really a [2:08] transformation. It's a disruptor. So [2:10] like any other IT technology that you're [2:12] bringing in the company, you need to [2:13] consider three key things. Really [2:15] simple. You all know about that, but [2:17] many still forget whenever it's time to [2:19] apply AI. You need people, process, and [2:22] technology, not just technology alone. [2:24] Um AI brings another fair of uh of [2:29] complexities and challenges. There's [2:31] resistance to change because there's [2:33] fear with AI. uh there's enhanced [2:36] complexity because of all the process it [2:38] touches and all the data that you need [2:40] on the the process that you're trying to [2:43] uh improve. [2:45] So AI will most probably fail and the [2:49] issue is not that uh you need to improve [2:51] the success rate and that's what we'll [2:53] go through today. Even if with a 20% [2:56] success rate you can still monetize AI [2:58] and create astronomical value for [3:00] organization. [3:03] So let me go through a little bit how [3:05] projects can be managed. I'll go through [3:07] various ways. They can all be applied [3:09] still today on AI. No one is no of those [3:12] uh project management practices are [3:14] legacy today. Um but I'll oversimplify. [3:17] So if there's any uh program managers in [3:19] the room, you won't like it. But it's to [3:21] make a point just to simplify uh and go [3:23] through this presentation in 20 minutes. [3:25] So you have classic projects that we [3:28] would call waterfall. You know exactly [3:30] the destination where you want to go and [3:31] you know the path to get there. So this [3:33] is about estimating that there will be [3:35] return on investment on your project. [3:37] Map all the risk, mitigate them and if [3:40] fine then you found the full project. [3:42] That's the way you manage a project [3:43] where you know the destination and you [3:44] know the path to get there. But what if [3:47] you know the destination you know where [3:48] what you want to do but you don't know [3:50] exactly how to get there. That's when we [3:52] invented the agile netology that is [3:54] still being used a lot today where you [3:56] will do incremental delivery incremental [3:59] approach so that you iterate quickly you [4:02] collect feedback and then you're able to [4:05] test and change your path along the way. [4:08] So now with AI many projects actually [4:10] you don't know the destination and you [4:12] don't know the path. So you're [4:13] completely in the dark. Um in order to [4:16] tackle those projects many companies [4:17] have started to apply something new that [4:19] is called fail fast. This approach is [4:22] still about first estimating the return [4:24] of investment that a project can make [4:26] but then identify what's the weakest [4:28] link of your project. [4:31] The weakest link is what is likely to [4:32] make a project fail and it's better to [4:35] test that on day one than to wait six [4:37] months before testing it. So the process [4:40] is quite simple again. You identify the [4:42] weakest link. You test it and if your [4:44] project is bound to fail, just make it [4:46] fail on day one and not after six [4:47] months. So what are weakest link and how [4:50] can you identify them? Uh let me give [4:52] you some examples. If you're running a [4:54] marketing campaigns and you will need to [4:56] get as precise information about your [4:58] customer uh as you can. There's some [5:01] privacy risk in there. Uh there's PII [5:03] information. So if your project might [5:06] have uh privacy or regulation uh uh [5:12] problems that might kill your project, [5:14] it's better to test that right away than [5:16] to develop a model, align stakeholders, [5:18] change your process, and then fail after [5:20] six months. It's really costly. Um if [5:23] your challenge is more, you know, [5:24] there's a business process, you don't [5:26] even know if there's data in there, just [5:27] outdated if data is there before even [5:29] going into uh privacy in that case. So [5:32] you really need to test the weakest link [5:34] and then iterate to the next one to the [5:36] next one to the next one and hopefully [5:38] kill this project as early as possible. [5:40] So that's the fail fast approach and [5:42] that helps you to even if you have a 20% [5:45] success rate, you're not spending too [5:46] much time on the failures and much more [5:48] time on the successful ones. So those [5:51] were the three main projects that are [5:53] being used today, project management [5:56] types that are still being used today if [5:58] you and all of them can be applied to [6:00] AI. So it's not that fail fast is always [6:02] the one to be applied in AI. Uh let me [6:04] give you some quick examples um that are [6:06] might not be evident at first but you [6:08] will understand. Um waterfall a good rag [6:13] uh chatbot that you would build in your [6:15] your organization would be more suited [6:17] for waterfall. Um creating a chatbot you [6:20] know it's possible today and you know [6:21] you have documentation so you can just [6:23] build a rack system on top of it. So [6:25] it's it's much more clear where you want [6:27] to go and how to build that. So maybe [6:29] waterfall is the right approach for this [6:31] type of project. Uh for an agent tech [6:34] system uh this is more software delivery [6:35] than training a model. So it's more a [6:38] software problem. You know exactly that [6:40] you want uh an agent that do some task [6:42] in the process but you don't know [6:44] exactly how to build it like do you need [6:46] an MCP? Where will the MCP server be [6:48] hosted? Uh what kind of model will you [6:50] be using? This is a little bit more [6:51] exploratory. So it's better to go in in [6:54] iteration and then agile would be the [6:56] right way. But what if you have a [6:58] business problem and you just don't know [7:00] where to start? Um, I'll go through this [7:03] example a little bit later, but what if, [7:05] for example, you have uh repeat calls in [7:07] your call centers and you want to use AI [7:09] to reduce this amount of calls. Where do [7:11] you start? Um, this is a little bit more [7:13] complex and you don't know exactly what [7:15] will be the end solution. In that case, [7:17] Felas is a good methodology. [7:20] What I'd like to go through now is [7:23] a step further. So those me mythologies [7:25] are great, but there's a way to even [7:28] spend more time on successful [7:30] initiative. And that's what we started [7:32] to apply at Talis and it's been [7:34] incredibly uh um useful for us. So it's [7:37] about balancing exploitation versus [7:40] exploration. Uh fell fast, you really [7:43] kill projects. What we're trying to do [7:44] here is not kill projects. It's actually [7:47] just priorizing and spending our time on [7:49] the most probable uh successful one. So [7:52] this balance between exploration and [7:54] exploitation is a well-known problem in [7:56] data science. So many of you probably [7:58] know about the multi-arm banded approach [8:00] and there's a way to adapt this to [8:02] business uh and and project management. [8:06] So what you need to do is still identify [8:10] the return of investment, validate if [8:12] the project is worth pursuing or not. Um [8:16] multiply that by a by a probability of [8:19] success. So you start a project with a [8:21] probability of success. It's a dynamic [8:22] return of on investment, not a static [8:24] one. So you don't start with one number [8:26] that you will go through the whole [8:28] project. You will adapt this as you [8:30] learn new stuff. So it's really a [8:32] process of learning and then identifying [8:34] if it's the right time to move to [8:35] another alternatives or not. There's [8:38] three key ingredients that you need to [8:39] have in place in order to apply this [8:41] methodology. First one, you need a [8:44] portfolio of good projects. If you have [8:46] a single one, you you cannot apply this. [8:48] That's the your best bet is for sure [8:50] doing this single project. But every [8:52] single business operation can be [8:54] enhanced by AI. So there's there's not a [8:57] lack of ideas and a lack of projects you [8:59] could be doing. You need to identify a [9:01] portfolio of initiative. So that's the [9:02] first ingredient. The second one is [9:05] having funding for an exploratory phase. [9:07] So if you don't have any OPEX, you just [9:10] want to fund capitalizable project. If [9:12] you're you're not you don't have funding [9:14] to just explore and test stuff out then [9:16] you can't start this because you won't [9:18] learn about the risk of each projects [9:19] and you won't be able to uh better [9:21] assess those risk. Um the third key [9:24] point is monetizing early because if you [9:26] wait until the end of the project to [9:28] monetize then the risk is much higher [9:30] and you're not able to validate if it's [9:32] successful or not. So those are the [9:33] three ingredients but they are fairly [9:35] easy to get to. Uh there's always a way [9:38] to break down a big problem into smaller [9:39] chunks. you just need to take a [9:41] different vision from proof of concept [9:43] to a proof of value. What is the minimum [9:45] variable insight that you can deliver to [9:48] uh the team so that they can get going [9:49] and create value with AI. [9:52] So to uh explain the same way this [9:54] project that would be estimate return on [9:56] investment uh and the odds of success [9:59] for each project then choose the best [10:01] bet apply fail fast approach but uh you [10:04] don't fail you pivot if needed. uh a [10:07] project that you don't do today might [10:08] become eventually the the best bet that [10:10] you can make. So you just dep prioritize [10:13] and monetize as early as possible and [10:15] just go through this loop. So that's [10:16] what we have been applying for the last [10:18] three four years and that's why we're [10:20] able to generate over 100 million every [10:21] year at Talis. Um so let's go through an [10:24] example to to make my point. Say you [10:27] have two projects project A worth 1 [10:29] million and project B worth 500K. [10:31] Normally with a classical approach, you [10:33] would go with project A for sure. But [10:36] you can apply here a pro of success to [10:38] project A. If it's a new project that [10:40] you've never seen before, you don't know [10:42] this uh you don't know the people, you [10:44] don't know the stakeholders, you don't [10:46] know the data, you don't know the [10:47] process, you're starting from scratch, [10:50] your best bet is to apply this uh ratio [10:53] we've seen before. Only one project five [10:55] successful. So this project to you is [10:57] worth 200K. But now let's say project B [10:59] is in your line of line of business. [11:02] You're in finance and the data in [11:04] finance, you know it well. You know your [11:05] stakeholders, you're those are [11:07] colleagues. You're pretty sure they will [11:08] be able to do something about the data [11:10] you deliver. Then you're moving on to [11:11] maybe a 50% chance of success. And then [11:14] project B is the one you should be [11:15] tackling, not project A. [11:18] So [11:20] once you identify this, this is where [11:23] you you start to uh tackle this project [11:26] B as a fail fast approach. Now I'll walk [11:28] you through a quick example uh on how [11:30] you can do this, how you can break down [11:32] a a complex problem into smaller chunks [11:34] um and how you can secure some steps uh [11:37] so that your success rate increase. Uh [11:40] so getting back to this call center [11:41] problem uh I I like explaining this one [11:44] because you can probably all relate to [11:45] this. Um in our call centers we have a [11:48] problem of repeat callers. A lot of our [11:51] consumer are calling back after a day or [11:54] two or a week later. Clearly, their [11:56] issue wasn't resolved and it's costing [11:57] us 50 million per year just on call [12:00] backs. So, the business came to us and [12:02] asked, can you build a model that would [12:04] predict who's likely to call back and if [12:07] you do so, we'll be able to proactively [12:09] act on this before they call back. So, [12:11] this is a great idea, great business [12:13] problem, huge opportunity. We should be [12:15] tackling this. This is great. But we [12:17] won't start by building a model because [12:19] building a model will take some time. [12:21] You need to make sure that you have the [12:22] right data. You need to go through [12:23] various situation. You need to [12:25] understand the business process and at [12:27] the end you're not sure that the [12:28] business will really act on it. And [12:29] we've seen many of those uh of those [12:32] requests coming to us. They want to use [12:34] AI but at the end they are resistant to [12:37] change or they have new priorities and [12:38] they never use the model. So what we did [12:41] there is not start with the AI. We [12:43] started by investigating why people are [12:45] calling back which seems obvious but [12:47] that's the first thing we did. We looked [12:49] at a pie chart of why people are are [12:51] calling and a big portion of the uh of [12:54] the call back are because of billing [12:55] issues. So billing is the biggest issue. [12:58] Uh digging deeper into billing we [13:00] discovered that uh the first billing is [13:02] actually the biggest issue people are [13:04] having. Um first billing um people [13:07] didn't understand about the activation [13:09] fees. So that's the main problem we [13:10] should be solving first. This is our our [13:12] proof of value and minimum variable [13:14] insight to get us going. Uh with that we [13:17] can improve uh agent training to make [13:20] sure that they do communicate well uh [13:22] the fees. We can uh also go to the [13:24] stores and maybe provide uh the first [13:26] billing in hand to our our customers and [13:28] already that is tackling today 3 million [13:30] out of the 50 million. So we're we [13:33] haven't solved the full issue and we [13:34] will get to AI eventually. we will build [13:36] a model but already at this point we've [13:38] secured the data we secured the privacy [13:40] approval we've secured the stakeholder [13:42] engagement which is key they're willing [13:44] to change and act on the data we we [13:45] provide so this is one way to monetize [13:48] early and also secure uh some risk uh of [13:52] your project as early as possible [13:55] so our approach hasn't been a software [13:57] delivery approach to AI if you want to [13:59] transform your business it's not just [14:01] about the technology our approach has [14:03] been business strategic consulting to [14:05] AI. So we approach more like a McKenzie [14:08] would than an Accenture. We're not [14:10] trying to develop AI for AI. We're [14:12] trying to find the right problems to [14:14] solve with AI. [14:17] Um so we started the project with design [14:19] thinking trying to find the right [14:21] problems to solve. Uh really important [14:23] when you do design thinking with AI, [14:24] don't try to figure out where AI can fit [14:27] in the company. AI will come into play [14:29] for sure. It's a powerful technology. It [14:31] does so much today with structured or [14:33] unstructured data. All our business are [14:35] generating a lot of data. For sure AI [14:37] will come into play in every project. [14:39] Everything that came to us as always an [14:41] AI component but don't start with AI [14:43] because it's it blindside people to [14:44] think about just AI and where can I [14:46] apply AI. It's much better to find the [14:48] right problem to solve. Then advanced [14:50] analytics can help a lot to get started [14:52] as I I I gave in my example. Just [14:55] getting statistics out can already get [14:57] stuff moving. strategic consulting, [14:59] maybe changing a process actually can do [15:01] much more of the value than the AI will [15:03] eventually fixing a broken process is [15:06] maybe where the value is before applying [15:08] AI. But the AI will come into play. As I [15:10] said, there's always an AI component in [15:12] each of those initiatives. So given I [15:14] don't have a lot of time left, I'll [15:16] finish up with five key lessons for [15:18] driving meaningful AI transformation [15:20] that wraps up a little bit what I've [15:21] been discussing before. So first, you [15:24] need to finance a portfolio of [15:25] initiative. This is a mindset change for [15:27] CFOs that prefer to go project by [15:29] project and and and and track [15:31] everything, every risk and and and be on [15:33] time with the right budget. This is [15:36] about being more like a venture [15:37] capitalist than a gambler. Don't put all [15:39] your eggs into one basket. Don't try to [15:41] make this project successful. Make some [15:43] projects that are portfolio successful. [15:45] If you have a 20% success rate, better [15:48] have the most value out of this 20%. [15:51] Start with a good problem. Uh don't fall [15:53] in love with the solution. Again, don't [15:54] be blindsided by the fact that we want [15:56] to apply AI everywhere. Find the right [15:58] problems to solve. AI will come into [16:00] play. Start early. Avoid overplanning. [16:02] Uh I still see a lot of companies [16:05] looking at multiple months to map all [16:07] the risk possible of the AI. Well, when [16:09] they don't even know what they don't [16:11] know, they will discover when they get [16:12] started. So get just get started. That's [16:14] how you will learn about the initiative. [16:16] Monetize iteratively. Don't wait for the [16:18] end of the project. Anytime somebody [16:20] comes to us and say if you build this [16:22] project uh deliver this model with this [16:24] kind of accuracy I'll be able to create [16:26] value we're we're pushing back and [16:28] exploring further for sure there's a way [16:29] to break down the problem into proof of [16:32] value and a proof of concept there's [16:34] enough hype around AI you'll need to [16:36] create more hype a proof of concept is [16:38] about creating hype what we want here is [16:40] create value with it so focus on proof [16:42] of value minimum variable insights and [16:44] not just creating more hype around AI so [16:48] I'll finish up by saying um let's stop [16:51] chasing AI projects for the sake of [16:53] applying AI. Let's make the cool useful. [16:56] Thank you.