Stride CEO Says 70% of Enterprise AI Projects Fail to Deliver ROI

Ai4

Companies face a 36-month window to fix broken workflows before forced AI adoption ruins operational margins and leads to business failure.

AI-assisted summary of The AI Profit Formula With Stride — original source. Check the original for full context.

Source published: 2025-08-27

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 27, 2025, this 17-minute video features Francisco Martin, CEO of Stride. According to the page description, the presentation explores how AI implementations often destroy business value when deployed without business context or guardrails, and outlines how to convert legacy systems into competitive AI knowledge powerhouses.

## TL;DR Despite projections of a $15.7 trillion economic impact, 70% of enterprise AI initiatives fail because companies attempt to automate broken legacy processes without proper context or outcomes in mind. Francisco Martin outlines 'The AI Profit Formula,' a strict execution hierarchy prioritizing People first, Process second, and Technology third. By establishing robust guardrails, modernizing legacy system logic, and rolling out autonomous 'beacon teams,' organizations can avoid catastrophic public AI failures and secure bottom-line ROI within a critical 36-month transformation window.

## Atmosphere & style The presentation carries an urgent, pragmatic, and authoritative tone, blending high-stakes executive warnings with actionable corporate advice. Martin delivers a grounded reality check on the pitfalls of corporate 'automation theater' and top-down mandates, contrasting absurd AI missteps with disciplined operational strategies. The session feels like a sharp strategic briefing designed to shake business leaders out of hype-driven complacency and motivate immediate structural action.

## Key takeaways - [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) While AI is projected to add $15.7 trillion to the global economy, 70% of organizational AI initiatives fail to deliver any ROI. - [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) Shoving AI into a broken legacy process creates 'automated dysfunction,' accelerating flawed procedures at a high cost. - [02:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=121s) Organizations must focus on financial outcomes rather than raw output or replacing headcount, recognizing that some problems do not require AI. - [04:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=241s) The AI Profit Formula relies on a strict three-tier hierarchy: People first, Process second, and Technology third. - [05:28](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=328s) Top-down mandates and proxy output metrics like lines of code or story points lead to employee resistance and low-quality results. - [08:26](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=506s) Implementing input and output guardrails is essential to prevent costly public failures, hallucinations, and regulatory liabilities. - [08:26](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=506s) Legacy systems should be viewed as a competitive advantage; AI can extract proprietary business logic to build defensible modern software. - [12:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=766s) Companies should transition from traditional, manual workflows to modern, AI-enabled models centered on outcome-based results. - [12:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=766s) Deploying multidisciplinary 'beacon teams' with specific targets and scaling them via a 'split and seed' strategy accelerates organizational AI adoption. - [15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) Organizations face a 36-month window to successfully implement an outcome-focused AI strategy before falling behind competitors and facing business obsolescence.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) The $15.7 Trillion Opportunity and the "Value Destruction Formula"

Francisco Martin introduces himself as the CEO of two distinct companies: Stride, an AI consultancy that helps clients adopt AI pragmatically, modernize legacy systems, and build software using AI, and Conductor, a product that automates code reviews for web performance and accessibility. Addressing the core challenge facing modern enterprises, Martin promises to reveal how to actually generate return on investment (ROI) from AI adoption. He opens with a stark contrast: while analysts project AI will contribute $15.7 trillion—"trillion with a T"—to the global economy in just a few years, 70% of current AI initiatives are failing outright. Seven out of ten companies attempting AI integration fail to generate a single cent in ROI, faring worse than breaking even.

``` +-----------------------------------------------------------------------+ | VALUE DESTRUCTION FORMULA | | | | [ Broken Legacy Process ] + [ Shove AI Into It ] | | = AUTOMATED DYSFUNCTION | +-----------------------------------------------------------------------+ ```

The underlying cause of this failure rate stems from what Martin calls the "value destruction formula." Organizations take an existing legacy process that was already broken or inefficient—playfully noting to younger audience members that the graphic depicts a floppy disk—and blindly "shove AI into that process." The result is merely "automated dysfunction," where companies incur the immense cost and technical complexity of AI while accelerating flawed procedures.

- **Market projections:** $15.7 trillion projected economic contribution from AI. - **Failure rate:** 70% of organizational AI initiatives fail to deliver ROI. - **Root cause:** Automating bad processes amplifies flaws rather than fixing them.

---

[02:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=121s) Reframing the AI Dialogue: Outcomes Over Output

Executives regularly collapse into failure traps by asking the wrong core questions. Typical missteps include asking "How do we AI the crap out of this process?", "How do we replace people with LLMs?", or getting bogged down in narrow technical debates regarding the best model or stack. In contrast, high-performing organizations completely reframe their inquiry around business fundamentals. They prioritize evolving operational processes to achieve clear financial outcomes—making or saving money—before deciding what tools to implement. Martin stresses that "some business outcomes don't need AI," requiring process re-engineering or traditional tooling instead.

Furthermore, LLMs are not currently positioned to replace human labor across most domains; rather, their real utility lies in scaling and amplifying human cognition. Crucially, leadership must ask how to establish guardrails to protect the company. AI possesses a distinct talent for executing "really, really, really dumb" actions with incredible efficiency if left unsupervised. The goal of AI transformation should never be maximizing raw throughput or lines of code, but driving bottom-line profitability.

- **Wrong questions:** Focus on replacing headcount, forcing LLM usage, or picking models prematurely. - **Right questions:** Focus on business outcomes, maximizing human cognition, and implementing protective guardrails.

---

[04:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=241s) The AI Profit Formula: People, Process, Technology

To consistently extract value from AI investments, Martin outlines "The Profit Formula," structured as a strict three-tier hierarchy where order of execution is paramount. At the foundational base are **People**. Frontline employees possess the clearest insight into operational bottlenecks and daily friction points. When leadership empowers and educates these workers, they can directly apply AI toward meaningful goals.

``` / \ / \ 3. TECHNOLOGY (Tooling, Models, Guardrails) /-----\ / \ 2. PROCESS (Outcome-Oriented Workflows) /---------\ / \ 1. PEOPLE (Empowered Frontline Teams) /-------------\ ```

The middle layer represents **Process**. Workflows must be fundamentally reoriented toward business outcomes rather than raw outputs. Measuring intermediate artifacts like lines of code or story points encourages teams to produce low-quality volume that actively harms the organization. At the peak of the pyramid sits **Technology**. Tool selection, model choices, and automated safety boundaries should only be addressed once the people and process foundations are firmly set.

- **Layer 1 (People):** Frontline empowerment, education, and bottom-up ownership. - **Layer 2 (Process):** Workflow design centered on business outcomes rather than output metrics. - **Layer 3 (Technology):** Tool selection, model integration, and governance guardrails.

---

[05:28](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=328s) Overcoming Top-Down Mandates and Output Traps

Top-down AI mandates—such as forcing entire engineering departments to use Copilot—frequently backfire. Employees feel disempowered, perceive that management does not understand their day-to-day workflows, and actively resist adoption. Martin highlights an ironic pattern where corporate leaderboards tracking raw AI usage often highlight top performers who deliver the worst overall business results. In software development, encouraging engineers to uncritically accept every LLM code suggestion yields a flood of broken features that damage customer outcomes.

Transforming this dynamic requires shifting from output volume to strategic value. Generative models can easily produce "10 million lines of code" on command, but volume is meaningless if the code fails in production. Organizations must measure business-oriented metrics such as customer satisfaction, retention, revenue, operating costs, and cycle times. Empowering teams means providing clear business objectives and available tools, then stepping back to let frontline workers re-engineer their daily operations.

- **Mandate pitfalls:** Rigid usage requirements drive resistance and artificial compliance metrics. - **Engineering failure mode:** Accepting wholesale LLM code output increases bug density and debt. - **Core metrics to measure:** Retention, revenue, customer satisfaction, operational costs, and cycle times.

---

[08:26](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=506s) Guardrails and Unlocking Legacy Systems

Once empowered teams and outcome-focused processes are established, organizations can turn to technical guardrails and legacy modernization. Skipping guardrails leads to catastrophic public failures. Martin cites a high-profile incident where McDonald's spent hundreds of millions of dollars with IBM on an automated drive-thru ordering system, only to scrap it after viral videos showed the AI repeatedly adding hundreds of chicken nuggets—up to 560 in one instance—to customer orders. Similarly, Air Canada's chatbot provided false information regarding bereavement fares, resulting in regulator fines and mandatory payouts.

``` +--------------------------------------------------------------------+ | GUARDRAIL LAYERS | | | | [ Input Guards ] --> Prevents jailbreaks, prompt injection, | | and PII leakage. | | | | [ Output Guards ] --> Halts hallucinations, false commitments, | | and compliance violations. | +--------------------------------------------------------------------+ ```

Guardrails act like race car barriers, allowing organizations to move faster safely rather than slowing them down. Effective architectures incorporate input guards against jailbreaks or PII leaks, alongside output guards to stop hallucinations. Crucially, AI should be used to monitor AI; shifting software engineers from writing code to manually reviewing endless LLM pull requests creates severe bottlenecks and worker fatigue.

Concurrently, AI provides a powerful mechanism to extract decades of business logic trapped inside legacy infrastructure. McKinsey estimates that Fortune 500 IT departments spend 70% of their budgets simply maintaining legacy systems, such as aging COBOL banking backends whose original developers are retiring. Rather than viewing legacy code as technical debt, companies should treat it as their core competitive advantage. AI tools can analyze and extract slices of this proprietary logic, modernizing it into a modern software stack that competitors cannot easily replicate.

- **McDonald's drive-thru failure:** Scrapped multi-million dollar AI initiative due to runaway order hallucinations. - **Air Canada legal liability:** Chatbot misstated bereavement policy, leading to legal enforcement. - **Legacy maintenance overhead:** 70% of Fortune 500 IT budgets are consumed by maintaining legacy systems. - **AI-driven review:** Automated AI code reviews prevent human review bottlenecks.

---

[12:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=766s) The Operating Model Shift and Beacon Implementation

To successfully adopt AI, enterprises must transition from legacy operational models to an agile, outcome-driven operational structure. Martin contrasts the traditional model with the modern AI-enabled paradigm across several dimensions:

``` +--------------------------------------------------------------------+ | BEACON TEAM IMPLEMENTATION ROADMAP | | | | [ Baseline Current State ] --> Audit output metrics vs. outcomes. | | [ Deploy Beacon Teams ] --> Give business targets & AI tools. | | [ Establish Sharing ] --> Incentivize successful practices. | | [ "Split and Seed" ] --> Disperse veterans to scale org. | +--------------------------------------------------------------------+ ```

Execution begins by baselining the current state to identify where teams track meaningless outputs instead of business value. Next, organizations deploy targeted "beacon teams"—representative multi-disciplinary units tasked with achieving specific business targets. These teams are granted tooling, training, and operational freedom to redesign their workflows. Successful experiments are recognized through incentive structures and knowledge-sharing platforms. Finally, companies scale this transformation using the classic agile "split and seed" approach: breaking up seasoned beacon teams and distributing experienced members into new business units to seed the culture company-wide.

- **Baseline audit:** Identify where teams track proxy metrics (e.g., story points) over real business value. - **Beacon team deployment:** Isolate representative teams, arm them with tools and training, and grant autonomy. - **Incentive alignment:** Reward successful process redesigns with cash incentives or internal recognition. - **Scale via "split and seed":** Disperse experienced beacon team members across the wider organization.

---

[15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) Strategic Summary and the 36-Month Horizon

Martin concludes by reiterating that AI adoption is fundamentally a people challenge first, a process challenge second, and a technology challenge third. Leaders must grant frontline workers autonomy while holding them strictly accountable for business targets. Organizations should actively extract proprietary logic from legacy applications to build defensible moats, while engineering modular, upgrade-friendly technical architectures that can adapt as new models and agentic systems emerge.

Closing with a reference to a trending industry perspective, Martin emphasizes that companies face a tight window: "You have 36 months to make it." This sense of urgency is not driven by the time required to deploy AI tools, but by the rapid pace at which competitors adopting this profit formula will widen the efficiency gap.

- **Core priorities:** Empower people, hold teams accountable, re-engineer processes, and extract legacy logic. - **Architecture design:** Build upgrade-friendly, modular frameworks with continuous guardrails. - **Timeline:** A 36-month window exists before AI-driven operational advantages create an insurmountable gap.

[17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) The 36-Month Deadline and the Recipe for AI Success

Francisco Martin delivers a stark warning and a clear call to action regarding the rapidly shifting business landscape: over the next 36 months, the business world is set to transform so dramatically that failing to execute an effective AI strategy will literally put companies out of business. There is an urgent, existential clock ticking for organizations to get this right, as improper implementations risk destroying value while falling behind competitors who adapt intelligently.

Despite the high stakes and the tight 36-month timeline, Martin reassures the audience that navigating this transition does not require an overwhelmingly complex formula. Success comes down to a straightforward, disciplined recipe executed in the proper sequence. The fundamental order of operations must always prioritize culture and operational structure before touching the tech stack: "People, process, then technology." He concludes his talk by thanking the audience for their time and attention.

- **Transformation Window:** 36 months to effectively integrate AI before facing business obsolescence. - **Core Execution Formula:** People $\rightarrow$ Process $\rightarrow$ Technology.

## Notable quotes > [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) trillion with a T — Francisco Martin

> [02:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=121s) How do we AI the crap out of this process? — Francisco Martin

> [02:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=121s) some business outcomes don't need AI — Francisco Martin

> [02:01](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=121s) really, really, really dumb — Francisco Martin

> [05:28](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=328s) 10 million lines of code — Francisco Martin

> [12:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=766s) split and seed — Francisco Martin

> [15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) You have 36 months to make it. — Francisco Martin

> [17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) People, process, then technology. — Francisco Martin

## Mentioned Francisco Martin, Stride, Conductor, McDonald's, IBM, Air Canada, McKinsey, Copilot, COBOL

## Video description Francisco Martin, CEO, Stride

Abstracts

Most AI implementations don't just fail, they actively destroy business value by flooding organizations with meaningless high-throughput noise. I'll expose why AI without business context and proper guardrails becomes a business liability that burns money while degrading decision quality.

This talk will reveal how to transform your legacy systems into AI knowledge powerhouses that create genuine competitive advantage instead of expensive automation theater. I'll show you the guardrails that separate value-creating AI from value-destroying implementations: automated performance monitoring that prevents costly failures, and compliance shields that protect against regulatory disasters.

All and all, you'll discover how to build AI that doesn't just work faster, it works intelligently with your business context, turning potential liabilities into profit engines.

[Source](https://www.youtube.com/watch?v=NmajE2zD5Dk)

中文简报

Stride CEO称70%的企业AI项目未能实现ROI 企业面临36个月的窗口期来修复残缺的工作流程，否则强行应用AI将侵蚀运营利润率并导致业务失败。尽管预测显示AI将带来15.7万亿美元的经济影响，但仍有70%的企业AI项目失败，原因在于企业在缺乏适当业务背景或明确成果考量的情况下，试图将受损的旧有流程自动化。Francisco Martin提出了“AI盈利公式”（The AI Profit Formula），这是一个严格的执行层级体系，遵循以人员为先、流程居次、技术排第三的原则。通过建立稳健的防护栏、实现遗留系统逻辑的现代化并部署自主的“灯塔团队”，企业可以在关键的36个月转型窗口期内，避免发生灾难性的公开AI失败并确保取得实质性的ROI。尽管预计 AI 将为全球经济贡献 15.7 万亿美元，但高达 70% 的企业 AI 举措未能带来任何投资回报（ROI）；将 AI 盲目塞入残缺的传统流程只会加速错误，演变为昂贵的“自动化功能失调”。 (0:04) 实现 AI 盈利必须遵循严格的执行次序：以赋能一线员工的“人”为基石，以业务成果为导向的“流程”居中，具体工具与模型等“技术”则置于最后。 (4:01) 自上而下强制推行 AI 或考核代码行数等代理指标极易引发抵触并堆积技术债务，企业应转向考核客户留存、营收和交付周期等真实业务成果。 (5:28) 建立自动化输入与输出护栏是防止 AI 幻觉和合规风险的底线；同时应利用 AI 提炼占据企业 70% IT 维护预算的遗留系统专属逻辑，将其转化为竞争壁垒。 (8:26) 企业仅有约 36 个月的转型窗口期；通过组建跨职能“灯塔团队”验证最佳实践，再以“拆分播种”方式在全公司推广，是平稳完成 AI 运营模式重塑的关键。 (15:46)

Reading summary

Francisco Martin, CEO of Stride and Conductor, argues that despite an anticipated $15.7 trillion economic impact, 70% of enterprise AI projects fail because organizations blindly apply models to broken workflows. He asserts that generating actual return on investment requires executing a strict hierarchy that prioritizes frontline staff and outcome-oriented processes over tooling. The Value Destruction Formula and Output Traps (0:04) Martin highlights a stark divide in enterprise adoption: while analysts project AI will add $15.7 trillion to the global economy, seven out of ten corporate initiatives fail to yield any return on investment. This failure rate stems from what he terms the "value destruction formula," where organizations take broken legacy workflows and insert AI, producing only "automated dysfunction" that accelerates operational errors at high technical expense. Leadership teams frequently stumble by asking how to replace staff with large language models or obsessing over model stacks rather than business fundamentals. Focusing on raw productivity artifacts, such as generating millions of lines of code or completing story points, incentivizes low-quality output. Martin stresses that models cannot universally substitute for human labor; instead, their real value lies in amplifying human cognition toward bottom-line financial goals like reducing costs and increasing revenue. The Hierarchy: People, Process, and Technology (4:01) To counter widespread implementation failure, Martin outlines "The Profit Formula" as a strict three-tier execution hierarchy. At the foundation are frontline employees, who best understand daily operational friction. Top-down mandates, such as forcing engineering departments to use tools like GitHub Copilot, routinely trigger employee resistance and result in artificial compliance metrics where high tool usage correlates with poor code quality. The second layer demands re-engineering processes around tangible business outcomes rather than volume metrics. Organizations must evaluate performance based on cycle times, customer satisfaction, retention, and operating expenses. Technology occupies the top of the pyramid: model selection, tech stacks, and tool integration should only be addressed after the people and process layers are firmly established. Mitigating AI Hallucinations and Unlocking Legacy Code (8:26) Implementing technical guardrails is essential to prevent costly operational and legal liabilities. Martin points to high-profile failures, such as McDonald's scrapping an IBM automated drive-thru that added hundreds of unwanted chicken nuggets to customer orders, and Air Canada facing legal penalties after its chatbot misstated bereavement fares. Dual-layer guardrails—input filters against prompt injection and PII leakage, paired with output verification to block hallucinations—act as safety barriers that enable organizations to move faster. Furthermore, using AI to monitor AI outputs prevents human engineers from becoming overwhelmed by manual review bottlenecks. Concurrently, Martin urges enterprises to view legacy infrastructure as a proprietary asset rather than technical debt. McKinsey estimates that Fortune 500 IT departments allocate 70% of their budgets to maintaining legacy environments, such as aging COBOL mainframes. AI tools can analyze and extract slices of this trapped, proprietary business logic, allowing enterprises to modernize it into newer architectures that competitors cannot easily replicate. Operational Shift and the Beacon Team Model (12:46) Transitioning to an AI-enabled enterprise requires replacing linear workflows and proxy metrics with parallel execution and outcome-focused goals. Martin details a structured implementation path beginning with a baseline audit of existing metrics to eliminate output-tracking traps. Next, companies deploy "beacon teams"—autonomous multidisciplinary units supplied with AI tooling and clear business targets, granting them the freedom to redesign their operational processes. To institutionalize success, organizations must reward effective workflow innovations through recognition or financial incentives. Companies can then scale these capabilities using an agile "split and seed" approach, dismantling established beacon teams and dispersing experienced members across other business units to seed the new operational culture organization-wide. Martin concludes with an urgent warning that enterprises face an approximate 36-month window to implement an effective AI operating model. He asserts that organizations failing to master this transition risk severe competitive obsolescence as peer firms widen the efficiency gap, though surviving this shift depends on disciplined execution rather than technical complexity: people first, process second, and technology last.

精炼阅读摘要

Stride 与 Conductor 首席执行官 Francisco Martin 提出，尽管分析师预测人工智能将为全球经济贡献 15.7 万亿美元，但目前 70% 的企业 AI 项目因陷入盲目套用工具的“价值毁灭公式”而未能实现投资回报。他主张企业必须摒弃唯产出论，严格遵循“人员、流程、技术”的优先级顺序重塑运营架构，以应对未来 36 个月的生存窗口期。价值毁灭公式与产出陷阱的根源 (0:04) Martin 指出，当前高达七成的企业 AI 项目无法收回成本，其核心症结在于企业落入了“价值毁灭公式”：直接将 AI 技术生硬地塞进原已脱节、低效的陈旧业务流程中，最终制造出成本高昂且加速错误运作的“自动化功能失调”。高管层在推进转型时往往纠结于错误的问题，例如执着于寻找替代人力的途径、盲目追逐底层大语言模型或技术栈评测，而忽视了商业回报的核心本质。自上而下的行政命令（如强行要求整个工程部门使用 Copilot）往往会激起基层抵触，催生单纯追求使用率的表面合规行为。在软件工程领域，片面考核代码行数、Story Points 等中间交付物会引发严重的负面效应。生成式模型虽然能轻松生成千万行代码，但若缺乏对业务实际价值的考核，全盘接受 AI 建议只会导致软件缺陷率激增和技术债务累积。Martin 强调，企业必须将评估标准从单纯的代码交付量转向商业成果，包括客户满意度、留存率、营收规模、运营成本缩减以及交付周期的实际改善。人员、流程与技术的收益金字塔架构 (4:01) 为了让 AI 投资切实转化为利润，Martin 提出了严格按先后顺序执行的三层“收益公式”金字塔。最底层的基石是“人员”，一线员工对日常业务堵点和操作摩擦拥有最直接的洞察，管理层必须首先为基层赋能和培训，建立自下而上的归属感与自主权。中间层是“流程”，工作流的设计必须以商业成效而非交付量为导向进行彻底重构，唯有理顺流程后才能有效降低损耗。位于金字塔顶端的才是“技术”，工具选型、模型接入与自动化防护规则必须建立在扎实的人员与流程基础之上。该架构的核心逻辑在于正确定位 AI 与人类的关系。Martin 认为，大语言模型目前并非用来全盘取代各领域的人类员工，其真正的战略价值在于规模化增强和放大人类的认知能力。某些具体的商业目标甚至根本不需要引入 AI，通过传统的流程重构或现有工具即可达成；如果脱离了成熟的人员认知与流程规范，AI 容易在无人监督的情况下以极高效率执行灾难性的错误决策。技术护栏防护机制与核心传统资产萃取 (8:26) 在技术落地层面，Martin 将安全护栏比作赛道防护墙，强调其作用是保障企业高速前进而非制造阻碍。忽视防护机制已导致多起严重的公开事故：例如麦当劳与 IBM 合作耗资数亿美元开发的自动得来速点餐系统，因 AI 在订单中屡次失控添加数百个麦乐鸡块（单笔最高达 560 块）而被彻底叫停；加拿大航空的聊天机器人因向乘客误报丧亲票价政策，导致航空公司招致监管罚款和强制赔偿。有效的安全架构必须构建双向机制：输入端护栏拦截提示词注入、越狱攻击与个人身份信息（PII）泄露，输出端护栏阻止模型幻觉、虚假承诺与合规违规，并应引入自动化 AI 监督机制以避免人工代码审查过载。与此同时，企业应当将传统系统视为不可替代的核心壁垒而非单纯的技术债务。麦肯锡的数据显示，《财富》500 强企业的 IT 预算中有 70% 仅仅用于维护传统遗留系统，例如大量原开发人员已退休的底层 COBOL 银行架构。Martin 指出，企业应利用 AI 分析并逐层萃取这些遗留代码中沉淀了数十年的核心专有商业逻辑，将其平滑迁移至现代软件技术栈，从而建立竞争对手难以轻易复制的技术护城河。灯塔团队运作机制与敏捷规模化路径 (12:46) 企业要摆脱线性、缓慢且依赖人工代码审查的传统模式，需通过敏捷的“灯塔团队”路径建立现代 AI 运营架构。实施流程首先从现状基线审计切入，识别组织内部将代理指标（如 Story Points、工具调用次数）与实际商业价值混淆的环节。随后，企业需要组建由跨学科人员构成的代表性灯塔团队，明确赋予具体的业务经营目标、AI 工具与自主权，允许其自由重构日常工作流。在灯塔团队探索出有效模式后，企业应通过内部共享平台与现金激励机制鼓励流程创新。最终的规模化扩散则依托经典的敏捷“拆分与播种”策略：将成熟灯塔团队的核心成员拆分并派遣至各个全新的业务单元，将验证有效的文化、流程重塑经验与 AI 协作规范快速移植至整个组织，从而推动全公司从追求交付量转向追求营收增长与周期缩短。 Francisco Martin 警告称，商业竞争格局正迎来关键节点，企业仅有 36 个月的窗口期来有效重塑 AI 战略，未能跟上步伐的企业将因效率差距被彻底淘汰。不过，该论断及其收益模型主要基于其实践经验与咨询观察，对于不同监管强度和技术成熟度行业的具体落地差异，仍有待更长期的量化检验。

Detailed notes — zh

## Background 该内容由 Ai4 于 2025 年 8 月 27 日发布在 YouTube 平台上。视频记录了 Stride 首席执行官 Francisco Martin 的主题演讲，重点探讨了为何多数企业 AI 项目陷入价值毁灭的泥潭，以及如何通过业务上下文、安全防护栏与激活遗留系统，将 AI 真正转化为盈利引擎。 ## TL;DR Francisco Martin 针对万亿级 AI 市场预期与高达 70% 落地惨败率的巨大反差，直指将 AI 强塞入低效流程所导致的‘自动化功能失调’。为此，他提出了‘人员赋能 $\rightarrow$ 流程演进 $\rightarrow$ 技术与防护栏’的自底向上 AI 盈利金字塔，主张摒弃代码行数等表面产出指标，转以净推荐值、成本与交付周期等商业成果为导向。此外，演讲强调必须借助全链路防护栏防范公关与合规灾难，通过 AI 激活企业遗留系统中的核心业务知识，并警告企业仅有 36 个月的时间窗口来完成这场智能化转型以避免被市场淘汰。 ## Atmosphere & style 演讲风格务实、犀利且充满紧迫感，主讲人开篇即用鲜明的数据反差打破行业对 AI 的盲目狂热。通过剖析麦当劳点餐失控与加航虚假承诺等真实翻车案例，现场弥漫着清醒审慎的警示意味；而在阐述‘标杆团队’与‘裂变播种’方法论时，语气则转为沉稳笃定，在 36 个月生存倒计时的压迫感中为企业管理者提供了一条清晰可行的破局路径。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) 商业预期与落地现实存在巨大落差：尽管 AI 预计将为全球经济带来高达 15.7 万亿美元的增量，但当前仍有整整 70% 的企业 AI 项目以失败告终，无法实现盈亏平衡。 - [01:20](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=80s) 警惕“价值毁灭公式”：将 AI 强行塞入原本就运转不灵的陈旧流程中，只会制造昂贵的“自动化功能失调”，企业提问应从“如何用 AI 裁员”转向“如何重构流程以实现赚钱或省钱”。 - [03:40](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=220s) 确立 AI 盈利金字塔的正确次序：企业落地 AI 必须严格遵循自底向上的执行原则——以“人员（People）”为基石、以“流程（Process）”为中层、以“技术与工具（Technology）”为顶层。 - [05:28](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=328s) 摒弃自上而下的行政强推命令：强制要求全员使用特定 AI 工具往往会引发抵触并导致低质代码泛滥，应转向自下而上赋能一线人员自主探索和重构工作流。 - [06:40](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=400s) 度量衡从表面产出转向真实业务成果：停止对代码行数、Story Points 等过程指标的考核，全面将度量锚定在客户满意度（CSAT）、留存率、营收增长和交付周期等核心商业结果上。 - [08:26](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=506s) 防护栏（Guardrails）是全速创新的安全前提：麦当劳穿梭点餐与加航客服机器人的翻车事件证明，必须在输入层、输出层与合规层构建严密防护，防止越狱、数据泄露与幻觉灾难。 - [10:00](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=600s) 以 AI 治理 AI 避免人工审核瓶颈：过度依赖人工介入（Human-in-the-loop）会使工程师沦为代码审查的单点瓶颈，应借助专业的自动化 AI 审查工具来保障质量与安全。 - [11:30](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=690s) 唤醒遗留系统（Legacy Systems）中的知识护城河：通用基础模型人人可买，而企业沉淀在 COBOL 等老旧系统中的数十年专有业务逻辑，才是 AI 时代不可复制的核心竞争优势。 - [13:34](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=814s) 规模化复制的敏捷演进路线：遵循“建立业务基线 $\rightarrow$ 标杆团队（Beacon Teams）试点 $ ightarrow$ 沉淀激励对齐 $ ightarrow$ 裂变与播种（Split and Seed）”四步法稳步推进企业转型。 - [15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) 权责对等与技术架构解耦：在给予一线人员创新自主权的同时实施严格的业务结果问责，并将底层系统设计为易于升级的松耦合形态以随时适配模型演进。 - [17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) 未来 36 个月的转型生存窗口期：当前正处于企业智能化重塑的分水岭，无法在未来 36 个月内遵循“人-流程-技术”规律完成转型的组织将面临被市场淘汰的风险。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) AI 落地困局：万亿预期与 70% 的惨败率作为 AI 咨询公司 Stride 与代码审查自动化产品 Conductor 的首席执行官，Francisco Martin 开门见山地指出了当前企业在 AI 落地过程中普遍面临的严峻现实与商业悖论。一方面，行业分析师预测 AI 将在短短几年内为全球经济贡献高达 15.7 万亿美元（"trillion with a T"）的惊人产值；但另一方面，现实数据却冷酷地表明，目前有整整 70% 的企业 AI 项目以失败告终。这意味着十家尝试引入 AI 的企业中就有七家颗粒无收，连基本的盈亏平衡（break even）都无法实现，更遑论投资回报率（ROI）。 - **主讲人背景**：Francisco Martin 同时掌舵专注于务实推进 AI 落地与遗留系统现代化的咨询公司 Stride，以及专注于 Web 性能与无障碍代码审查自动化工具 Conductor。 - **宏观经济反差**：预期高达 15.7 万亿美元的经济增量 vs. 现实中 70% 的企业 AI 项目未产生任何正向商业回报。 --- ### [01:20](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=80s) 价值毁灭公式：向失效流程硬塞 AI 的灾难 Francisco 揭示了绝大多数企业陷入泥潭的核心症结——他称之为“价值毁灭公式”（Value Destruction Formula）。许多管理者误将 AI 当作灵丹妙药，直接强行塞入原本就运转不灵的陈旧流程（legacy process）中，结果不仅没有消除原有的低效，反而制造出了极其昂贵的“自动化功能失调”（automated dysfunction）。这背后的根源在于高管们往往在问完全错误的问题：比如“我们如何把这个流程彻底 AI 化？”或者“我们怎样用大语言模型（LLM）取代员工？” 与之相对，真正成功的组织往往回归商业本质，将焦点从“技术手段”转向“业务成果”（Business Outcomes）。管理者应当首先思考如何演进业务流程以实现赚钱或省钱的目标，进而评估包括 AI 在内的哪些工具能够辅助达成该目标。Francisco 强调，并非所有业务瓶颈都适合或需要 AI，很多时候流程重构才是解药。此外，当前阶段的 LLM 根本无法在绝大多数场景下完全替代人类，其真正价值在于“放大人类认知”（amplify human cognition）并赋予业务杠杆。如果没有合适的防护栏（guardrails），AI 在执行愚蠢决策时同样具备极高的“效率”。 ```mermaid flowchart LR accTitle: 价值毁灭公式与价值创造模式对比 subgraph Destruction["价值毁灭模式（Top-down / 关注产出）"] A["陈旧失效流程"] --> B["强行灌入 AI 工具"] B --> C["自动化功能失调 / 成本激增"] end subgraph Creation["价值创造模式（Bottom-up / 关注成果）"] D["明确商业成果目标"] --> E["重构业务流程与认知放大"] E --> F["配置技术与安全防护栏"] end ``` - **错误提问方向**：过度聚焦于“如何用 LLM 裁员”、“如何最大化代码行数”或“哪款底层模型参数最强”。 - **正确提问方向**：聚焦于“如何演进流程以达成商业盈利/降本”、“如何借助工具放大前线认知”以及“如何构建护栏保护业务安全”。 --- ### [03:40](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=220s) AI 盈利金字塔：人、流程与技术的正确次序为了逆转 AI 投资的颓势，Francisco 提出了切实可行且至关重要的“AI 盈利公式”（Profit Formula）。该框架呈三层金字塔结构，其核心逻辑在于严格自底向上的执行顺序：底层为“人员”（People），中间层为“流程”（Process），顶层才是“技术与工具”（Technology）。 ```mermaid flowchart TD accTitle: AI 盈利金字塔架构 subgraph Pyramid["AI 盈利金字塔（自底向上）"] T["【顶层】技术与防护栏 (Technology & Guardrails) 工具选型、模型解耦、自动化防御体系"] P["【中层】流程演进 (Process) 以商业结果为导向，而非盲目衡量产出指标"] H["【基石】人员赋能 (People) 前线业务人员自下而上重构认知与工作流"] H --> P P --> T end ``` 作为金字塔基石的“人员”，指的是那些身处业务一线的员工。一线团队最清楚痛点所在，一旦获得认知赋能和技术自主权，他们才能真正找到用 AI 解决业务问题的支点。金字塔中层的“流程”必须围绕最终的商业成果重新设计；如果流程的考核指标依然停留在代码行数、Story Points 等过程产出，AI 只会加速制造毫无价值的高通量垃圾。最后，当人员与流程就绪后，企业才进入顶层的“技术与工具”选型，并同步配置严密的安全防护栏。 --- ### [05:28](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=328s) 人员赋能与流程重塑：从过程产出转向商业成果在深入剖析“人员”维度时，Francisco 直言自上而下的行政命令（mandates）往往是 AI 落地的死穴。诸如“全员必须使用 GitHub Copilot”之类的强推指令，不仅让员工感到被剥夺权力，还会引发强烈的抵触情绪。在软件研发场景中，盲目要求工程师全盘接受 LLM 生成的代码，最终往往演变成大量充斥着漏洞的功能被推向生产环境，直接破坏系统稳定性与客户体验。健康的做法是“自下而上”的赋能：向一线团队明确核心商业目标并提供工具箱，让他们自主探索工作流的演进。在“流程”维度上，企业必须彻底摒弃对“表面产出”（outputs）的盲目崇拜。大语言模型可以在几秒内生成上千万行代码或海量文案，但这些并不能等同于商业价值。流程的度量衡必须锚定在核心业务指标上：例如客户满意度（CSAT）、客户留存率（Retention）、营收增长、运营成本削减以及整体交付周期（Cycle Time）。只有当流程以真实的商业价值为驱动时，工具才能成为倍增器，而非混乱的制造机。 - **自上而下强制推行的弊端**：排行榜驱动的工具使用率往往与业务成果成反比，盲目合并 AI 代码导致次生技术债务激增。 - **度量指标转向**：从衡量“代码行数/Story Points/工具打卡率”转向衡量“净推荐值/留存率/成本控制/交付周期”。 --- ### [08:26](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=506s) 技术防护栏与唤醒遗留系统中的知识金矿在金字塔顶层的“技术”环节中，Francisco 特别强调了“防护栏”（Guardrails）的决定性意义。没有防护栏的 AI 极易酿成公关与财务惨剧：例如麦当劳曾斥资数亿美元联手 IBM 研发汽车穿梭餐厅（Drive-thru）AI 点餐系统，结果因模型频频出错并在 TikTok 上引发用户订购 560 块麦乐鸡块的病毒式嘲讽视频，最终不得不草草收场；加拿大航空（Air Canada）的官方聊天机器人因对乘客编造错误的丧亲抚恤票价政策（bereavement fares），导致航司被监管机构裁定承担赔偿与罚款。防护栏绝非拖慢业务的官僚繁文缛节，它就像赛车跑道外侧的防护墙，唯有具备完善的安全边界，团队才敢全速前进。构建防护栏必须涵盖输入层（防注入、防越狱、防敏感数据 PII 泄露）、输出层（防幻觉、事实核查）以及合规治理层。值得注意的是，企业不应将所有审核压力甩给“人类介入”（Human-in-the-loop），否则会让工程师从写代码沦为整天审查 Cursor PR 的“审查瓶颈”；相反，应当运用专业的 AI 工具来实现“以 AI 治理 AI”。 ```mermaid flowchart TD accTitle: AI 全链路防护栏与遗留资产激活机制 subgraph GuardrailArch["多层防护栏体系"] In["输入防护 Prompt 越狱拦截 / PII 隐私脱敏"] --> Core["AI 处理与模型推理"] Core --> Out["输出防护 幻觉检测 / 事实校对 / 业务规则合规"] Out --> AutoRev["自动化 AI 审查 (非人工单点瓶颈)"] end subgraph LegacyData["遗留资产激活"] Cobol["传统银行/保险系统 (COBOL、50年核保沉淀)"] --> AIEx["AI 深度代码与逻辑提取"] AIEx --> CompAdv["转化为现代技术栈下的独家竞争优势"] end ``` 与此同时，Francisco 指出了一项常被忽略的战略技术资产——**遗留系统（Legacy Systems）中的隐性知识**。麦肯锡数据显示，财富 500 强企业的庞大 IT 团队将高达 70% 的预算耗费在维护以 COBOL 等老旧语言构建的传统系统上，而掌握这些系统的老一辈工程师正在陆续退休。市面上的通用模型任何竞争对手都可以买到，但企业深锁在遗留系统中的数十年业务逻辑与专有数据（如保险公司 50 年的核保沉淀）才是真正的护城河。利用 AI 深度解析、切片并现代化这些遗留系统，能够将沉重的维护负担转化为独一无二的核心竞争力。 - **AI 翻车典型案例**：麦当劳穿梭餐厅 AI 因反复强行添加麦乐鸡块（单次暴增至 560 块）导致项目流产；加航因客服机器人虚假承诺丧亲票价被监管重罚。 - **遗留系统维护现状**：头部企业 70% 的 IT 支出被锁定在遗留系统维护中，面临人员断层的知识流失风险，亟需通过 AI 进行知识切片与现代重构。 --- ### [13:34](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=814s) 落地演进路线：从标杆团队到规模化复制企业若想摆脱“为 AI 而 AI”的被动局面并实现真正的 ROI，需要遵循一条清晰的敏捷转型路径： 1. **确立基线（Baseline）**：全面梳理现有团队与流程，严谨界定每个团队为业务带来的实际价值，坚决用商业结果替代 Story Points 等虚荣产出指标。 2. **设立标杆团队（Beacon Teams）**：在企业内部挑选具备代表性的少数团队作为标杆。明确指派具体的业务目标，给予培训与工具自主权，充分授权一线团队自主重构工作流程。 3. **沉淀机制与激励对齐（Incentives & Knowledge Sharing）**：建立跨团队的知识分享体系，设立诸如“月度最佳 AI 商业创新实践”等奖金激励，驱动自发式技能迭代。 4. **裂变与播种（Split and Seed）**：借鉴敏捷转型的“裂变播种”模式，将标杆团队中跑通全流程、取得规模化业务成果的骨干成员分散派驻至其他部门，带动全公司范围的落地开花。 ```mermaid flowchart LR accTitle: AI 转型落地四步演化路线 S1["1. 业务基线梳理 明确真实商业成果"] --> S2["2. 标杆团队试点 充分授权与重构流程"] S2 --> S3["3. 知识沉淀与激励 分享机制与奖金牵引"] S3 --> S4["4. 裂变与播种 骨干扩散带动全局转型"] ``` --- ### [15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) 总结与 36 个月窗口期：权责对等与架构解耦在演讲的第一部分尾声，Francisco 对 AI 盈利的核心逻辑进行了归纳：始终坚持“人员第一、流程第二、技术第三”的法则。在充分赋能一线人员的同时，必须贯彻“权责对等”（With great power comes great responsibility）原则，依据最终业务产出实施严格问责。在技术架构上，系统必须设计为“易于平滑升级”（upgrade-friendly）的松耦合形态，以便随时无缝接入快速迭代的基础模型与 Agent 架构，同时将安全防护栏贯彻始终。最后，Francisco 抛出了一个引人深思的时间窗口判断：当前企业大约只有 36 个月的时间来完成这场智能化跃迁。这并非指有效采用 AI 需要耗费整整 36 个月，而是因为留给落后者的生存窗口正在加速关闭。 ### [17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) 把握未来36个月的转型窗口：人、流程与技术在未来的 36 个月内，商业世界预计将经历极其剧烈而深远的变革。演讲者强调，如果企业无法在此期间高效、切实地推进并落地 AI 战略与系统构建，“你将会被彻底淘汰出局”（"you're going to be out of business"）。这个 36 个月的时间窗口既是紧迫的生存倒计时，也是所有企业重新确立竞争优势的关键机遇期。尽管面临严峻的紧迫感，但转型的路径并非无迹可寻。演讲者指出，好消息在于只要按照正确的逻辑顺序与步骤推进，这其实是一套非常清晰明了的成功法则（"pretty simple recipe"）。实现真正赋能业务而不是制造自动化杂音的核心次序必须是：首先关注“人”（People），其次梳理与优化“流程”（Process），最后才是引入和应用“技术”（Technology）。最后，演讲者对在场所有观众的聆听与支持表达了真诚的感谢。 ```mermaid flowchart LR accTitle: AI转型三步法 A["人员 (People)"] --> B["流程 (Process)"] --> C["技术 (Technology)"] ``` - **关键转型时间窗**：未来 36 个月是企业重构业务模式、融入 AI 核心能力的分水岭，执行不力的企业将面临被淘汰的风险。 - **落地核心法则**：严格遵循“人 $\rightarrow$ 流程 $\rightarrow$ 技术”（People, Process, then Technology）的实施顺序，确保 AI 紧贴业务上下文创造实际价值。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=4s) trillion with a T — Francisco Martin > [01:20](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=80s) amplify human cognition — Francisco Martin > [03:40](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=220s) People, Process, then Technology — Francisco Martin > [15:46](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=946s) With great power comes great responsibility — Francisco Martin > [17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) you're going to be out of business — Francisco Martin > [17:17](https://www.youtube.com/watch?v=NmajE2zD5Dk&t=1037s) pretty simple recipe — Francisco Martin ## Mentioned Francisco Martin, Stride, Conductor, GitHub Copilot, Cursor, 麦当劳 (McDonald's), IBM, TikTok, 加拿大航空 (Air Canada), 麦肯锡 (McKinsey), COBOL ## Video description Francisco Martin, CEO, Stride Abstracts Most AI implementations don't just fail, they actively destroy business value by flooding organizations with meaningless high-throughput noise. I'll expose why AI without business context and proper guardrails becomes a business liability that burns money while degrading decision quality. This talk will reveal how to transform your legacy systems into AI knowledge powerhouses that create genuine competitive advantage instead of expensive automation theater. I'll show you the guardrails that separate value-creating AI from value-destroying implementations: automated performance monitoring that prevents costly failures, and compliance shields that protect against regulatory disasters. All and all, you'll discover how to build AI that doesn't just work faster, it works intelligently with your business context, turning potential liabilities into profit engines. [Source](https://www.youtube.com/watch?v=NmajE2zD5Dk)

Source quotes

[0:04] Francisco Martin: trillion with a T [1:20] Francisco Martin: amplify human cognition [3:40] Francisco Martin: People, Process, then Technology [15:46] Francisco Martin: With great power comes great responsibility [17:17] Francisco Martin: you're going to be out of business [17:17] Francisco Martin: pretty simple recipe

Source quotes

[0:04] Francisco Martin: trillion with a T [2:01] Francisco Martin: How do we AI the crap out of this process? [2:01] Francisco Martin: some business outcomes don't need AI [2:01] Francisco Martin: really, really, really dumb [5:28] Francisco Martin: 10 million lines of code [12:46] Francisco Martin: split and seed [15:46] Francisco Martin: You have 36 months to make it. [17:17] Francisco Martin: People, process, then technology.

Original transcript

[0:04] My name is Francisco Martin. I'm the CEO [0:06] at Stride and Conductor, which are two [0:08] separate companies. Uh, Stride is an AI [0:11] consultancy. So, we help clients adopt [0:13] AI pragmatically. We also help them [0:15] modernize their legacy systems using AI [0:18] and we just generally build AI software [0:20] using AI, lots of AI, right? Uh and then [0:23] conductor is actually a product that [0:25] automates uh uh code reviews for web [0:29] performance and accessibility. And so [0:32] I'm gonna tell you how to make money out [0:35] of uh AI adoption. Everyone is talking [0:37] about this. Everyone is talking about [0:38] how difficult it is. And you know [0:40] through our client work we've figured [0:42] out a way to do it. [0:44] So let's start with something um a bit [0:48] scary. Um, analysts say that AI is going [0:52] to contribute [0:54] 15.7 trillion dollars to the economy in [0:56] just a few years. That's trillion with a [0:59] T. That's a lot of money. Um, but the [1:02] problem is that today 70% of AI [1:05] initiatives fail. [1:07] So that means that seven out of 10 [1:10] companies are trying to adopt AI and [1:12] they're not generating a single ROI. [1:15] It's not break even. No, nothing. [1:20] And why? Well, most organizations are [1:23] doing this thing that you see on the [1:24] screen. This is what I call the value [1:26] destruction formula. [1:29] So, you take a legacy process, you know, [1:31] that process that you already know that [1:33] wasn't working particularly well. Um, [1:36] for those that are young, that's a [1:37] floppy disc in case you didn't recognize [1:39] it. Uh, and then just shove AI into that [1:42] process. [1:44] And what happens? Well, you just get [1:47] automated dysfunction, [1:50] right? You don't realize that the [1:52] process was the problem to start with [1:54] and you're just adding all the [1:56] complexity and all the cost of AI. [2:01] Why does this happen? [2:04] Well, this happens because executives [2:06] are generally asking the wrong [2:08] questions, right? We start thinking [2:10] about, okay, you know, how do we AI the [2:13] crap out of this process? Or even [2:16] better, my favorite, how do we replace [2:18] people with LM? [2:21] And even, you know, let's just very [2:23] detailed technical questions of what is [2:25] the best model, what is the best [2:27] technology, they don't realize that [2:29] those are just the wrong questions to [2:31] ask. So what I've seen successful [2:33] organizations do is start asking the [2:36] questions on your right side which is [2:39] how do we evolve our processes to [2:41] achieve business outcomes. There's [2:43] nothing about AI there. Let's start by [2:45] the basics. How do we make money or how [2:48] do we save money? Those are business [2:50] outcomes. [2:51] What tools including AI can help us get [2:54] there? And please note that I say [2:56] including AI. Some problems, some [3:00] business outcomes don't need AI. They [3:03] may need something else. They may need [3:04] process re-engineering. They may need [3:06] something else. Not everything is fit [3:08] for AI today. [3:10] How do we maximize the impact of human [3:13] cognition? Let's let's be real. LM are [3:17] not in a place where they can replace [3:18] humans for most things. They can help [3:21] human scale, right? This is a problem [3:24] where we have the possibility to amplify [3:26] human cognition. There are plenty of [3:28] companies doing this across the board [3:30] and this is what we should be thinking [3:31] about. How do we guard rail our business [3:34] from AI and you know this is something [3:36] that again companies don't realize AI is [3:40] really really really good at doing [3:42] things that are really really really [3:44] dumb if you don't put guardrails to it. [3:47] So we need to think about how do we [3:48] change our processes? How do we include [3:50] guardrails so we protect our business [3:52] outcomes. That's what we're trying to [3:53] achieve. We're not trying to maximize [3:56] lines of code or throughput. We're [3:58] actually trying to make money. [4:01] So this is what I call the profit [4:03] formula and this is what I found to be [4:07] the most impo impactful to actually [4:09] generate an ROI from uh AI adoption. [4:11] It's a pyramid has three layers and the [4:14] order is extremely important. What is at [4:17] the bottom? The foundation people. It [4:21] all starts with people. It all starts [4:23] with empowering those that know best how [4:26] to use AI to achieve business outcomes. [4:29] You know, I'm going to break it to you [4:30] all. People on the front lines, they're [4:32] the ones that know where problems are. [4:34] They're the ones if given the [4:36] empowerment and given the technology and [4:38] given the knowledge, can actually use AI [4:40] to achieve business outcomes. But [4:42] they're the foundation. [4:44] Second is process. Again, if you have a [4:48] process that is just trying to produce [4:50] outputs, not outcomes, again, lines of [4:53] code, all those things are outputs, then [4:56] you're just going to produce more [4:58] outputs at scale that actually going to [4:59] detract from your business outcomes. So, [5:02] you need to think about how to orient [5:03] your processes to business outcomes. And [5:06] last, at the very top, is technology. [5:09] Yes, we can discuss technology. We can [5:11] pick a model. We can pick a technology. [5:13] And remember also pick guard rates. [5:19] So let's dive a bit deep. I'm going to [5:21] go through each one of these um and and [5:24] just uh give you some pointers on on how [5:26] to think about this. First one is [5:28] people. Um and what I've seen very very [5:32] happen across uh industries and clients [5:35] is mandates, right? Everyone needs to [5:38] use Copilot. Everyone needs to use this [5:40] other tool. And what happens then is [5:43] that they're just again shoving AI into [5:46] their current workflows. People feel [5:48] disempowered. [5:50] And worst of all, they just resist [5:52] adoption. They just don't want to do it [5:54] because they don't you don't understand [5:55] their process and you're just trying to [5:57] force them to change their process [5:59] without a business outcome in mind. [6:01] And so again, I've seen a lot of [6:04] companies that have leaderboards for air [6:06] adoption and the top performers are the [6:09] ones that are producing the worst [6:10] business outcomes. [6:12] Classic example, right? You have a [6:13] software engineering team. You tell [6:15] them, hey, use Copilot or whatever other [6:17] tool, cursor, cloud code, and you [6:20] encourage your engineers to just accept [6:23] everything that the LLM is writing. [6:26] Guess what? that you're going to get a [6:27] lot of features at the door and most of [6:28] them are going to be broken and then [6:30] most of them are going to damage your [6:31] business outcomes. I've seen that across [6:33] the board. [6:35] So, how do we change that? I think this [6:38] is a bonus of transformation. [6:40] What I with the companies that I've seen [6:42] successfully adopting AI, they empower [6:44] people on the front lines to rethink [6:46] their process. They equip them with the [6:49] knowledge and the technology to think [6:50] about, oh, I have this new tool. Let's [6:52] see how we can change the way that we do [6:54] things in order to achieve business [6:56] outcomes. [6:57] This enables this process evolution. [6:59] Again, bottoms up and really encourages [7:02] skill development and adoption. You're [7:04] not just telling people, hey, go use [7:06] this tool. You're telling them, this is [7:07] the business outcome that your team has. [7:10] These are the tools that are available [7:12] to you. Go figure it out. [7:15] But, you know, it's hard to empower [7:16] people, right? [7:19] Then process. Process. second. So most [7:22] companies when they evaluate how a [7:26] process is working, they measure [7:28] outputs. And if you measure outputs and [7:31] you add AI to that, you're just going to [7:32] get a lot of outputs. And those outputs [7:34] might be bad quality and may not achieve [7:36] any business outcomes whatsoever. Again, [7:39] lines of code, if you're thinking about [7:40] software engineering, lines of code, [7:43] story points, tool usage, [7:46] these are not business outcomes. And as [7:49] I said before, AI is very very good at [7:51] creating massive amounts of outputs. You [7:53] can just go to Chad GPT and say, "Hey, [7:56] give me 10 million lines of code and [7:57] it's going to do it right away." Is that [7:59] code going to do anything good for your [8:01] business? Probably not. [8:03] So, we need to reorient process to be [8:06] focused on business outcomes, customer [8:09] satisfaction, retention, revenue, cost, [8:12] even cycle time. [8:15] That's how you empower teams to [8:18] accomplish business goals with tools as [8:20] opposed to just outputs. [8:26] There you go. And last but not least, [8:29] technology. Yes, there is obviously a [8:32] place to choose technology. Once you [8:33] have empowered teams that are empowered [8:36] to change processes that are designed to [8:38] accomplish outcomes, then you think [8:40] about technology. You think about tool [8:43] selection. You actually ideally empower [8:45] those people in the front lines to pick [8:47] their own tools. And then don't forget [8:50] create guard rails. Even the best teams [8:53] empower even the best processes with the [8:56] best tools. If you don't create guard [8:58] rails, you can get into trouble. We've [9:00] all heard horror stories, right? There [9:02] was McDonald's, I think, invested [9:04] hundreds of millions of dollars working [9:06] with IBM, I think, on redoing their [9:08] drive-thru um application and they [9:13] canned it because there were all these [9:15] viral videos of people trying to order [9:17] stuff and just fighting with the robot [9:19] because the robot kept adding chicken [9:21] nuggets. For some reason, the LM really [9:23] liked chicken nuggets. There's this [9:24] funny video, you guys, you guys should [9:26] check it out. Someone on Tik Tok just [9:28] just adding more and more and I think it [9:29] reached like 560 chicken nuggets. No [9:32] guardra, right? So even though you may [9:35] have all the right steps in mind, you [9:37] need to finish that. [9:41] Now I want to talk for a second about a [9:43] piece of technology that is not that [9:45] sexy that nobody really talks about, but [9:47] I think it's fundamental to benefit from [9:49] AI. Most organizations have a wealth of [9:52] knowledge stuck in their legacy systems, [9:55] right? And historically it is extremely [9:59] expensive to extract that knowledge. So [10:01] we just spend a bunch of money [10:02] maintaining those systems. You know if [10:04] you think about banking, how many cobalt [10:06] applications are out there? So many. I [10:08] think Mackenzie estimates that large IT [10:10] teams in Fortune 500 companies spend 70% [10:13] of their budgets just maintaining legacy [10:16] systems. And a lot of the people that [10:18] know these systems are retiring by the [10:19] way. So all the knowledge of those [10:20] systems is is going to be gone soon. But [10:23] actually AI allows you to extract this [10:26] knowledge, put it in a new stack, and [10:28] then be able to leverage it as your [10:30] competitive advantage. Legacy systems [10:33] are not a pool. They're actually the [10:35] value that your business creates, right? [10:37] Uh all your competitors can buy all the [10:39] AI tools that you can buy. Everyone can [10:41] do the same, but they don't have the [10:42] knowledge that you have. If you're an [10:44] insurance company, you have all this [10:45] underwriting data from the past 50 years [10:48] stuck in your legacy systems. So AI [10:51] today allows you to actually trace those [10:54] legacy systems in depth more than any [10:56] human could and extract slices that you [10:59] can modernize and then you can start [11:01] using as your competitive advantage as [11:04] it is guards. You know when I talk about [11:06] guards I think everyone is embracing [11:08] this now which is great but six months [11:10] ago I was like I don't want any guards [11:12] that that allows me it's just going to [11:14] slow me down. It's bureaucracy. It's [11:16] governance. And it is not. It's actually [11:18] the only way that you can move fast if [11:20] your goal is to produce business [11:22] outcomes. Um, you know, race car guards [11:26] actually allow you to go faster because [11:28] if you don't have guard rails, it's [11:30] going to be scared. You can tip over. [11:31] Um, and when you think about guardrails [11:33] for AI, there are so many different [11:35] layers that you need to consider, right? [11:37] You have guard rails for inputs. You [11:40] know, you want to make sure that users [11:41] are not trying to jailbreak your system [11:43] or get PII or anything along those [11:45] lines. So, think about those guards. [11:48] Then you think about output guards. You [11:50] want to make sure that LLMs are not [11:52] hallucinating or providing false [11:54] information. Um, there's this story, I [11:56] think Air Canada, uh, they built a [11:58] chatbot that was basically giving [12:00] misinformation about bereavement fairs [12:02] or false things and they got a massive [12:04] bill, uh, from the regulator out of [12:06] that. So you want to make sure that even [12:08] if you think the system is working well, [12:10] there are guards that prevent bad things [12:12] from happening. There's other kinds of [12:15] guards, governance, compliance. You need [12:18] to consider all of those. And the good [12:20] thing is that AI can actually guardrail [12:23] AI better than any human can. When you [12:25] think about guardrails, don't just put a [12:27] human in the loop. If you're a software [12:28] engineering team and you're basically [12:30] moving all of all your engineers from [12:32] writing code to reviewing pull requests [12:35] from cursor, you're just putting the [12:37] bottleneck elsewhere and making a bunch [12:38] of people miserable. There are AI tools [12:40] that allow you to do that code review. [12:46] So what does this look like in in real [12:49] life? Well, again, it goes back to [12:52] changing people, process, and [12:54] technology. You go from manual reviews, [12:58] linear processes, tool thinking, [13:01] disengaged talent, people quitting, [13:03] locked in competitive advantage to a [13:06] much more fluid, scalable place where [13:09] you can have parallel workflows, [13:11] automated guardrails that produce scale [13:14] outputs, engaged talent that allows you [13:17] to scale the human cognition, [13:19] outcome focused thinking that leads you [13:22] to scale results, and then you add all [13:25] that knowledge that's stuck in your [13:26] legacy systems and that's how you get [13:28] scale business outcomes. [13:34] So how do you go from where you are [13:37] which you know some of the stuff that I [13:39] discuss about mandates and things like [13:40] that may resonate with you to a place [13:42] where you can actually generate an ROI [13:44] from AI. Well, first you need to [13:47] baseline where you are, right? You need [13:49] to understand where are your people, [13:52] where are the different processes, the [13:54] different teams, do they have um output, [13:57] do are they are you measuring outputs or [13:59] are you measuring business outcomes? Um [14:01] and things along those lines. And if [14:03] you're measuring business outputs, then [14:05] think about okay, what is this team [14:06] actually producing for my business? It's [14:08] funny, I work with a lot of clients [14:09] where I ask them, hey, what value does [14:11] this team provide to your business? and [14:13] they say story point like that's not [14:17] value right they have a really hard time [14:19] thinking about how software engineering [14:21] teams are producing value so the first [14:22] step is think about what is that [14:24] business value once you have that [14:26] baseline then design some beacon teams [14:29] depending on how big your organization [14:31] is design a few that can represent a [14:34] majority of your teams and have them [14:36] empower them say okay beacon team number [14:38] one this is the business outcome that [14:41] you're supposed to produce [14:43] This is a menu of tools. This is [14:45] training. Figure out how to incorporate [14:47] this technology, this AI into your [14:50] processes. Change the process so that [14:52] you can produce business value. [14:55] Empower them. Empowerment is so [14:57] important. You need to let people in the [14:58] front line change their processes. It's [15:01] not going to work otherwise. [15:04] And finally, as you progress and you [15:06] have several beacon teams, establish [15:08] knowledge sharing systems, incentives, [15:11] you know, there I've seen like things [15:13] like leaderboards and I was talking to [15:15] someone uh earlier this morning whose [15:17] company established like, you know, the [15:19] best AI initiative of the month and they [15:21] got some uh cash incentive. So basically [15:24] create those. [15:27] Finally, once you have that success in [15:29] those beacon teams, this is the [15:31] traditional agile transformation path, [15:33] right? Split and sit. You have a bunch [15:34] of people that have done this, have [15:36] produced business outcomes at scale [15:38] using AI, put them in other places in [15:40] your in your um company and have them [15:43] spin up other teams. [15:46] So, just to summarize, and I know I keep [15:48] going back to the same things, but this [15:50] is actually pretty simple. It's a people [15:53] problem first, then a process problem, [15:54] and then a technology problem. Empower [15:57] your people, give them knowledge, also [16:00] give them accountability. You know, with [16:01] great power comes great responsibility. [16:03] If you're telling folks, hey, you know, [16:06] you're you have all these tools [16:08] available. You can do whatever you want [16:09] to achieve this business outcome, you [16:11] should also hold them accountable. [16:12] That's the counterpart to that. But do [16:14] that first. Then change your process. [16:19] leverage that unlockable business logic [16:22] that lays in your legacy systems so you [16:24] can actually use it to your advantage [16:28] and create a foundation. You know, there [16:30] there are times where um you really need [16:33] to make sure that all your data is in [16:35] one place and and have all these kind of [16:37] platform things that allow you to kind [16:39] of redesign the the process. [16:42] And last, guardrail your technology. You [16:45] know, create systems that are ideally [16:47] upgrade friendly. You know, models are [16:49] changing constantly. Every day we have a [16:51] new model. We have a new agentic system. [16:53] So, make sure that you're designing [16:54] something that can upgrade as the models [16:56] upgrade. But guardrail guardrail and go [16:59] over. [17:04] So, just kind of want to leave you all [17:06] with with this last this this is a a [17:08] Twitter thing that everyone is talking [17:09] about this in a different context, but [17:11] you know, you have 36 months to make it. [17:13] And it's not because it takes 36 months [17:15] to adopt AI effectively. is because in [17:17] 36 months everyone thinks the business [17:20] world is going to change so dramatically [17:22] that if you are not doing this [17:24] effectively you're going to be out of [17:26] business. [17:28] So we have 36 months to do it. The good [17:31] news is that it's a pretty simple recipe [17:33] if you think about it in the right [17:34] steps. People process then technology. [17:39] Thank you so much. I appreciate your [17:40] time.