<p>人工智能项目很像新年决心：它们开始时充满激情和大胆的承诺，但大多数都未能实现。IBM 高级数据与人工智能技术专家 Brianne Zavala 提出的这一鲜明类比，突显了困扰企业机器学习计划的普遍问题。尽管进行了大量投资和革命性的演示，研究表明，只有 25% 的人工智能计划能够实现预期的投资回报，而成功在整个企业范围内扩展的仅占 16%。这些被放弃的努力，往往停滞在试点阶段，构成了 Zavala 所称的 “人工智能墓地”——一个未能带来实际商业价值的雄心勃勃想法的安息之地。</p>
<p>Zavala 的演讲重点不在于技术突破，而在于那些在项目能够实现突破前就注定失败的关键组织和战略失误。她认为，核心问题往往是对人工智能在组织中角色的根本误解。</p>
<p>最大的错误，以及她第一个实用建议的主题，是从技术出发而不是从问题出发。团队常常问：“我如何使用人工智能？” 而不是问：“我们正在解决的商业挑战是什么？” 这种微妙的框架差异是生死攸关的。正如 Zavala 直言不讳地指出：“人工智能是一个工具。它实际上并不是目标。” 当团队仅仅因为最新的算法或模型令人印象深刻而追逐它们时，他们实际上是在黑暗中构建。确保人工智能项目提供可持续价值的唯一方法是将其与明确、可衡量的商业成果相结合。这可能意味着将客户流失率降低 10%，提高运营效率 15%，或削减 200 万美元的成本。这些指标作为项目的北极星，提供了问责制和在初始兴奋消退后衡量进展的明确路径。如果没有提前定义成功指标，项目在开始之前就面临风险。</p>
<p>第二个关键失败点是组织摩擦，源于缺乏早期利益相关者的支持。人工智能项目本质上跨越部门边界，涉及公司内的流程、人员和产品。首席财务官、运营负责人和客户成功团队都有不同的优先事项，并使用不同的专业语言。首席财务官关心成本节约和投资回报，运营负责人关注工作流程整合，而客户成功则专注于客户留存和用户体验。Zavala 强调，成功取决于以每个利益相关者特定视角相关的方式传达项目的价值。你的工作不是向首席财务官讲解模型架构，而是连接各方，向每个团队展示解决方案如何改善他们的工作。通过尽早确保这种对齐，团队将潜在的阻碍者转变为支持者，确保项目获得广泛采用所需的组织动力。</p>
<p>这种对深度跨职能整合的需求直接导致了第三个，或许是最被忽视的失败模式：将部署和监控视为事后考虑。数据科学家往往专注于构建和训练在受控环境中表现良好的模型，例如 Jupyter Notebook。然而，真正的挑战在于模型进入生产环境时。这是大多数项目绊倒的地方。团队未能为在动态商业环境中维护人工智能系统的操作现实进行规划。这需要强大的 MLOps 实践，包括将模型与现有企业系统集成、建立对性能漂移的持续监控，以及在基础数据变化时设置自动再训练循环。Zavala 对数据科学家和产品经理发出尖锐警告：“在 Jupyter Notebook 中有效但从未投入生产的模型只是摆设。”</p>
<p>成功的概念验证与可扩展、创造收入的企业解决方案之间的区别完全在于这三个战略领域。构建一个强大的模型只是战斗的一半；另一半是确保它解决一个可量化的商业挑战，获得组织支持，并为持续部署和监控而设计。通过采用这种严谨的方法——关注商业挑战、确保早期利益相关者的支持，并从第一天起就为部署做好准备——组织可以显著提高其人工智能投资带来实际、可衡量影响的概率，避免陷入人工智能墓地的命运。</p>

<p>AI 项目往往未能实现预期的投资回报率，成功率仅为 25%。IBM 的 Brianne Zavala 强调，将 AI 倡议与明确的商业挑战对齐的重要性，而不仅仅是关注技术。成功的关键因素包括定义可衡量的结果、确保各部门利益相关者的支持，以及从一开始就规划部署和监控。通过关注这些领域，组织可以提高其 AI 项目实现实际商业价值的机会，避免所谓的 “AI 墓地”</p>

<p>AI projects are a lot like New Year’s resolutions: they begin with excitement and bold promises, but most don’t make it. This stark analogy, drawn by IBM Sr. Data &amp; AI Technical Specialist Brianne Zavala, highlights the pervasive issue plaguing enterprise machine learning initiatives. Despite massive investments and revolutionary demos, research shows that only 25% of AI initiatives deliver the expected Return on Investment, and a mere 16% successfully scale across the enterprise. These abandoned efforts, often stalling in the pilot phase, constitute what Zavala terms the “AI Graveyard”—a resting place for ambitious ideas that never delivered real business value.</p>
<div class="lb-trans"><p>人工智能项目很像新年决心：它们开始时充满激情和大胆的承诺，但大多数都未能实现。IBM 高级数据与人工智能技术专家 Brianne Zavala 提出的这一鲜明类比，突显了困扰企业机器学习计划的普遍问题。尽管进行了大量投资和革命性的演示，研究表明，只有 25% 的人工智能计划能够实现预期的投资回报，而成功在整个企业范围内扩展的仅占 16%。这些被放弃的努力，往往停滞在试点阶段，构成了 Zavala 所称的 “人工智能墓地”——一个未能带来实际商业价值的雄心勃勃想法的安息之地。</p>
</div><p>Zavala’s presentation focuses not on technical breakthroughs, but on the crucial organizational and strategic missteps that condemn projects before they can achieve escape velocity. The core problem, she argues, is often a fundamental misunderstanding of AI’s role within the organization.</p>
<div class="lb-trans"><p>Zavala 的演讲重点不在于技术突破，而在于那些在项目能够实现突破前就注定失败的关键组织和战略失误。她认为，核心问题往往是对人工智能在组织中角色的根本误解。</p>
</div><p>The biggest mistake, and the subject of her first practical tip, is starting with the technology instead of the problem. Too often, teams ask, “How can I use AI?” rather than, “What’s the business challenge that we are solving?” This subtle difference in framing is existential. As Zavala states bluntly, “AI is a tool. It’s not actually the goal.” When teams chase the latest algorithms or models simply because they are impressive, they are building in the dark. The only way to ensure an AI project delivers sustainable value is to anchor it to clear, measurable business outcomes. This could mean reducing customer churn by 10%, improving operational efficiency by 15%, or cutting costs by $2 million. These metrics serve as the project’s North Star, providing accountability and a clear path for measuring progress long after the initial excitement wears off. Without defining success metrics upfront, the project is at risk before it even begins.</p>
<div class="lb-trans"><p>最大的错误，以及她第一个实用建议的主题，是从技术出发而不是从问题出发。团队常常问：“我如何使用人工智能？” 而不是问：“我们正在解决的商业挑战是什么？” 这种微妙的框架差异是生死攸关的。正如 Zavala 直言不讳地指出：“人工智能是一个工具。它实际上并不是目标。” 当团队仅仅因为最新的算法或模型令人印象深刻而追逐它们时，他们实际上是在黑暗中构建。确保人工智能项目提供可持续价值的唯一方法是将其与明确、可衡量的商业成果相结合。这可能意味着将客户流失率降低 10%，提高运营效率 15%，或削减 200 万美元的成本。这些指标作为项目的北极星，提供了问责制和在初始兴奋消退后衡量进展的明确路径。如果没有提前定义成功指标，项目在开始之前就面临风险。</p>
</div><p>The second critical failure point is organizational friction, stemming from a lack of early stakeholder buy-in. AI projects inherently cross departmental boundaries, touching processes, people, and products across the company. The Chief Financial Officer, the Operations Lead, and the Customer Success team all have different priorities and speak different professional languages. The CFO cares about cost savings and ROI, the Operations Lead is concerned with workflow integration, and Customer Success is focused on retention and user experience. Zavala emphasizes that success hinges on communicating the project’s value in terms relevant to each stakeholder’s specific lens. Your job is not to lecture the CFO on model architecture; it is to connect the dots and show each team how the solution makes their world better. By securing this alignment early, teams turn potential blockers into champions, ensuring the project gains the organizational momentum necessary for broad adoption.</p>
<div class="lb-trans"><p>第二个关键失败点是组织摩擦，源于缺乏早期利益相关者的支持。人工智能项目本质上跨越部门边界，涉及公司内的流程、人员和产品。首席财务官、运营负责人和客户成功团队都有不同的优先事项，并使用不同的专业语言。首席财务官关心成本节约和投资回报，运营负责人关注工作流程整合，而客户成功则专注于客户留存和用户体验。Zavala 强调，成功取决于以每个利益相关者特定视角相关的方式传达项目的价值。你的工作不是向首席财务官讲解模型架构，而是连接各方，向每个团队展示解决方案如何改善他们的工作。通过尽早确保这种对齐，团队将潜在的阻碍者转变为支持者，确保项目获得广泛采用所需的组织动力。</p>
</div><p>This need for deep, cross-functional integration leads directly to the third, and perhaps most overlooked, failure mode: treating deployment and monitoring as an afterthought. Data scientists often focus intensely on building and training a model that performs well in a controlled environment, such as a Jupyter Notebook. However, the true challenge begins when the model moves into production. This is where most projects stumble. Teams fail to plan for the operational realities of maintaining an AI system in a dynamic business environment. This requires robust MLOps practices, including integrating the model with existing enterprise systems, establishing continuous monitoring for performance drift, and setting up automated retraining loops as the underlying data changes. Zavala offers a sharp warning to data scientists and product managers alike: “A model that works in something like a Jupyter Notebook but never makes it into production is just shelfware.”</p>
<div class="lb-trans"><p>这种对深度跨职能整合的需求直接导致了第三个，或许是最被忽视的失败模式：将部署和监控视为事后考虑。数据科学家往往专注于构建和训练在受控环境中表现良好的模型，例如 Jupyter Notebook。然而，真正的挑战在于模型进入生产环境时。这是大多数项目绊倒的地方。团队未能为在动态商业环境中维护人工智能系统的操作现实进行规划。这需要强大的 MLOps 实践，包括将模型与现有企业系统集成、建立对性能漂移的持续监控，以及在基础数据变化时设置自动再训练循环。Zavala 对数据科学家和产品经理发出尖锐警告：“在 Jupyter Notebook 中有效但从未投入生产的模型只是摆设。”</p>
</div><p>The distinction between a successful proof-of-concept and a scalable, revenue-generating enterprise solution lies entirely in these three strategic areas. Building a powerful model is only half the battle; the other half is ensuring it is solving a quantifiable business challenge, has organizational support, and is engineered for continuous deployment and monitoring. By adopting this disciplined approach—focusing on the business challenge, securing early stakeholder buy-in, and building for deployment from day one—organizations can significantly increase the probability that their AI investment delivers real, measurable impact, escaping the fate of the AI Graveyard.</p>
<div class="lb-trans"><p>成功的概念验证与可扩展、创造收入的企业解决方案之间的区别完全在于这三个战略领域。构建一个强大的模型只是战斗的一半；另一半是确保它解决一个可量化的商业挑战，获得组织支持，并为持续部署和监控而设计。通过采用这种严谨的方法——关注商业挑战、确保早期利益相关者的支持，并从第一天起就为部署做好准备——组织可以显著提高其人工智能投资带来实际、可衡量影响的概率，避免陷入人工智能墓地的命运。</p>
</div>

Avoiding the AI Graveyard: The Business Discipline Required to Scale Machine Learning Projects

- AI 项目往往未能实现预期回报，仅 25% 的项目成功交付。  
- 重要的是确保项目与明确的业务目标相结合，早期获得利益相关者支持。  
- 组织需关注投产与监控，确保模型在动态环境中可持续运作。

避免 AI 项目失败：扩展机器学习项目所需的商业纪律