从充满希望到富有成效： 服务业生成式人工智能的应用成果

生成式人工智能（gen AI）可以提高运营主管们所希望的生产力，也可以成为应对成本压力的一种手段——前提是主管们能够付诸行动。麦肯锡的最新技术趋势研究发现，全球只有11%的公司大规模使用生成式人工智能。

Generative AI (gen AI) could provide the productivity boost operations leaders have hoped for, as well as a means to fight cost pressures—if only leaders could get going. McKinsey’s latest tech trends research finds that only 11 percent of companies worldwide are using gen AI at scale.¹

运营是一个重大缺口：2024 年 2 月，对北美和欧洲大型公司的 150 名高管进行的一项调查显示，只有 3% 的受访者表示他们的组织已经在运营相关领域扩展了生成式人工智能用例。2024年4月，对全球250多名企业职能部门领导进行的另一项调查发现，服务运营领域的情况仅略有改善。例如，在财务部门，目前约有45%的企业正在试用生成式人工智能解决方案，而2023年这一比例仅为11%，但只有6%的企业实现了规模化应用。

Operations is a major gap: in a February 2024 survey of 150 executives at large North American and European companies, only 3 percent of respondents said their organization has scaled a gen AI use case in an operations-related domain. A separate survey, conducted in April 2024 of more than 250 corporate-function leaders worldwide, found that service operations is faring only slightly better. In finance functions, for example, about 45 percent of organizations are now piloting gen AI solutions, compared with 11 percent in 2023—but only 6 percent have achieved scale.

调查结果反映出运营领导者对部署的众多用例中哪些能带来真正的竞争优势存在不确定性。高管们明白，从他们的生成式人工智能投资中实现全部价值不会一蹴而就：4月份调查的受访者中有三分之二的人将时间期限设定为三到五年（图1）。

The results reflect uncertainty among operations leaders about which of the many use cases they have deployed will yield real competitive advantage. Executives understand that realizing full value from their gen AI investments won’t be instantaneous: two-thirds of the April survey respondents set a three- to five-year timeline (Exhibit 1).

但许多人也表示，他们希望更加确信自己的投入能够获得回报。一位首席执行官最近告诉我们：“我们已经投入了大约1亿美元，用于数百个生成式人工智能实验；至少收获一部分价值将有助于我们确定哪些地方值得追加投资。”公司还指出，不明确的路线图、人才短缺和管理不成熟也是阻碍规模化的因素。

But many also said that they wanted to be more confident that their commitments would pay off. One CEO recently told us, “We’ve already spent about $100 million funding hundreds of gen AI experiments; harvesting at least some of the value will help us see where additional investment will be worthwhile.” Companies also cited unclear road maps, talent shortages, and immature governance as further impediments to scale.

然而，少数公司已经获得了真正的价值，他们将超过10%的息税前利润归功于使用生成式人工智能。² 早期的成功揭示了在组织中推广生成式人工智能的三个关键任务。第一是为部署生成式人工智能设计一个连贯、严谨的运营策略。这意味着要优先考虑具有长期价值的用例，关注它们不仅改变特定流程点或领域，而且重塑整个工作流程的潜力。

A few companies, however, are capturing real value already, attributing more than 10 percent of their EBIT to their use of gen AI.²  Early successes like these reveal three critical tasks in setting gen AI up for scaling across an organization. The first is to design a cohesive, disciplined operational strategy for deploying gen AI. That means prioritizing use cases for long-term value by focusing on their potential to not only transform specific process points or domains but also reimagine complete workflows.

其次，为了长期保持广泛的影响力，企业需要关注那些支持员工创造生成式人工智能的推动因素——提供必要的治理和绩效基础设施，同时投资变革管理和持续创新文化。

Second, to sustain broad impact over time, companies will need to focus on the enablers supporting the humans who make gen AI work—providing the necessary governance and performance infrastructure while also investing in change management and a continuous innovation culture.

第三项任务也是前两项的最终目标：将生成式人工智能工具与人类能力巧妙地结合在一起，以创建最先进的解决方案，例如自主生成式人工智能agent或copilots。最成功的解决方案可以处理复杂工作流程的每个步骤。例如，一家银行的人工智能agent现在可以起草信用风险备忘录，使每位客户经理的收入提高了20%。而一家消费品制造商财务部门的copilot则减少了600万至1000万美元的财务规划和分析相关运营支出。

The third task is the culmination of the first two: thoughtfully integrating gen AI tools with human capabilities to create the most advanced solutions, such as autonomous gen AI agents or copilots. The most successful can tackle every step of a complex workflow. At one bank, for instance, a gen AI agent now drafts credit-risk memos, increasing revenue per relationship manager by 20 percent. And a copilot in the finance department of a consumer goods maker is reducing operating expenses relating to financial planning and analysis by between $6 million and $10 million.

战略性地部署生成式人工智能

Deploying operational gen AI strategically

与早期技术变革浪潮一样，生成式人工智能技术引发了试验阶段的担忧，数十次实验都未能产生持续的影响。那些已经具备部署生成式人工智能技术能力的组织，无论从短期还是长期来看，其生成式人工智能技术的投资回报率都更高。他们尤其擅长思考排序问题，注重可扩展性和可重复使用性，从而能够重新构想整个价值创造链。

As with earlier waves of technological change, gen AI raises the specter of pilot purgatory, in which dozens of experiments fail to amount to sustained impact. Organizations that have already built up their capabilities in deploying gen AI tend to see better returns on their gen AI investments over both the short and longer term. They especially excel at thinking through sequencing, with a focus on scalability and reusability so that they can reimagine entire chains of value creation.

在生成式人工智能转型中建立这种成熟度现在至关重要，留给企业的时间已经不多了。理想情况下，从生成式人工智能的低风险早期应用中汲取的经验教训可以建立关键能力，帮助高风险（高回报）的后期应用取得成功。

Building this sort of maturity in gen AI transformation is now essential, leaving companies little time to waste. Ideally, lessons from lower-risk, earlier applications of gen AI build critical capabilities that help higher-risk (and higher-reward) later applications succeed.

优先使用场景

Prioritizing use cases

一家全球性银行的经历说明了战略性部署生成式人工智能的优势。首先，在对业务影响和技术可行性进行详细评估的基础上，它将23个潜在领域缩减为两个：消费者银行业务部门的联络中心和公司及投资银行业务部门的了解客户（KYC）功能。尽管这两个领域存在明显差异，但它们不仅对生成式人工智能的影响潜力巨大，而且有一些共同点，特别是在基于生成式人工智能的知识提取和合成方面。同样的技术可以支持客户查找信息和员工查找内部文档，从而更有效地重复使用和扩展基础技术（图2）。

The experience of a global bank illustrates the benefits of deploying gen AI strategically. First, based on a detailed assessment of business impact and technical feasibility, it winnowed 23 potential domains for use of gen AI to just two: the contact center in its consumer banking unit and the know-your-customer (KYC) function for corporate and investment banking. Despite the apparent differences, the two domains not only showed high potential for gen AI impact but also shared a few commonalities, particularly for gen-AI-based knowledge extraction and synthesis. The same technologies could support customers looking for information and employees looking for internal documents, so that the underlying technology could be reused and scaled more effectively (Exhibit 2).

为了确定两个入围者中的哪一个先进行部署，公司又增加了一道筛选标准：风险。KYC功能数据的保密性使其成为风险较高的目标，因此银行转而从联络中心入手。最终的战略决策涉及在客户服务中部署哪些用例。在考虑“可扩展性”和“可重复使用性”的基础上，聊天机器人脱颖而出：它相对易于实施，可产生可衡量的结果，并有助于为类似用例提取和综合复杂数据奠定基础。

To determine which of the two finalists would go first, the company applied an additional screen: risk. The confidential nature of the KYC function’s data made it a higher-risk target, so the bank started instead with the contact center. The final strategic decision concerned which use cases to deploy within customer care. Keeping “ability to scale” and “reusability” top of mind, the chatbot came out on top: it’s comparatively easy to implement, generates measurable outcomes, and helps build a foundation for similar use cases that extract and synthesize complex data.

在短短几周内，该中心设计完成的用例中就包括一个面向客户的聊天机器人。在投入使用短短七周内，新的聊天机器人就改善了客户体验，为大约20%的联络中心请求节省了等待时间。

Within just a few weeks, the center’s fully designed use cases included a customer-facing chatbot. In just seven weeks of use, the new chatbot offered an improved customer experience, eliminating wait times for about 20 percent of contact center requests.

此外，从联络中心学到的经验为银行提供了可重复利用的基础，使其能够适应KYC功能。聊天机器人现在是“智能虚拟代理”的一部分，它通过更自动化的KYC流程指导客户经理。虚拟代理可以将客户信息预先填入表格，确定所需的文件，验证数据上传，并跟进任何缺失的信息。

Moreover, lessons from the contact center have formed a reusable foundation that the bank can adapt for the KYC function. Chatbots are now a component of a “smart, virtual agent” that guides relationship managers through a far more automated KYC process. The virtual agent can prepopulate client information into forms, determine which documents are required, validate data uploads, and follow up on any missing information.

从单一解决方案到完整工作流程

From point solutions to complete workflows

然而，正如银行案例所说明的那样，生成式人工智能的核心运营问题不是“生成式人工智能如何帮助我改进当前流程？” 改进流程通常意味着只解决表面问题，而不是根本问题——例如，使用生成式人工智能自动记录会议笔记和生成行动项目，而没有考虑为什么一开始会有这么多会议。

As the bank example illustrates, however, the core operational question for generative AI isn’t “How could gen AI help me improve my current processes?” To “improve” a process often means addressing only a symptom rather than the underlying condition—for example, using gen AI to automate note-taking and action-item generation for meetings without considering why there were so many meetings in the first place.

因此，核心问题要广泛得多：“生成式人工智能如何帮助我重新思考我的运营？”回答这个问题意味着重新审视每个流程，将其作为更大工作流程的一部分，在许多情况下，也是用户或客户旅程的一部分。

The core question is therefore much broader: “How could gen AI help me rethink my operations?” Answering it means reexamining each process as part of a larger workflow—and, in many cases, as part of a user or customer journey.

打破壁垒，提供更好的服务。为了说明这种差异，让我们来看一个北美领先电信提供商的案例。该公司通过优先级排序，将重点放在客户服务上。该公司没有一开始就探索生成式人工智能工具如何改进客户服务中的特定流程步骤，而是退一步思考生成式人工智能如何与传统流程改进技术和新人才相结合，以提高客户服务职能的整体生产力。

Breaking barriers to better service. To illustrate the difference, consider the case of a leading North American telecommunications provider, whose use case prioritization exercise led it to focus on customer care. Rather than start by exploring how gen AI tools could improve particular process steps in care, the company stepped back, asking instead how gen AI could combine with traditional process improvement techniques and new talent to raise productivity within the customer care function overall.

这种视角的转变促使公司重新评估其客户体验，首先从传统绘制每个环节的流程图开始，从最初接触到最终解决。公司领导根据绘制的流程图对每个流程步骤进行审查，以确定其是否过于复杂或没有必要。他们考虑了每个步骤对客户体验的影响（例如增加复杂性或等待时间）以及消除该步骤可能带来的风险（例如增加欺诈或安全漏洞）。

That shift in perspective led the company to reevaluate its customer journeys, starting with a traditional mapping of every touchpoint, from initial contact to final resolution. With the resulting flowcharts in hand, company leaders questioned each process step to see if it was overengineered or unnecessary. It considered the step’s effect on customer experience (such as increasing complexity or wait times) versus the potential risks from its elimination (such as increased fraud or security lapses).

例如，在该公司绘制了更改电话号码的流程图后，发现其中一个步骤非常复杂和痛苦，因此该公司让客户选择委托员工完成该步骤，并收取一定费用。但由于客户不愿意付费，员工往往会引导客户完成这一步骤--这对联络中心来说是个昂贵的选择。一旦公司了解了客户卡壳的原因，就能设计出自助服务解决方案。结合其他技术，生成式人工智能能够提供详细的自动指导，这意味着该公司可以缩短平均通话时长（和成本），同时完全取消收费，改善客户体验（见图 3）。

For example, after the company mapped out the journey of changing a phone number, one particular step surfaced as so complex and painful that the company gave customers the option to delegate it to staff in exchange for a fee. But because customers were reluctant to pay, staff would often guide them through the step—a costly alternative for the contact center. Once the company understood the reasons customers got stuck, it could design a self-service solution. In combination with other technologies, gen AI’s capabilities to provide detailed, automated guidance meant that the company could reduce average call length (and cost) while eliminating the fee entirely, improving customer experience (Exhibit 3).

对客户痛点根源的深入分析还揭示了公司在生成式人工智能提供解决方案之前需要解决的内部失调问题，例如营销团队设定的价格变动导致客户来电激增，护理团队无法处理。由于没有意识到这些变化，代理商会将客户转给其他部门，往往是多次循环，导致护理团队提供大幅折扣，希望能留住沮丧的客户。

Deeper analysis into the root causes of customer pain points also revealed internal misalignments that the company needed to address before gen AI could provide a solution—such as when price changes set by the marketing team led to surges in customer calls that the care team couldn’t handle. Unaware of the changes, agents would transfer customers to other departments, often in multiple loops, leading the care team to provide deep discounts in hopes of retaining the frustrated customers.

因此，公司改造了跨职能工作流程，使客户服务团队能够与市场营销部门合作，预测潜在的客户关注点，并提前制定适当的应对措施。领导者还重新审视了服务团队所需的技能，开发了新的人才配置文件（以及相关的能力建设模块），可与工作流程共同发展。改变内部协作模式为日后基于生成式人工智能的自助服务选项奠定了基础，而生成式人工智能分析工具则可以优化人员分配，以提供额外的呼叫中心覆盖范围。

Accordingly, the company revamped its cross-functional workflows so that the customer care team could work with marketing to anticipate potential customer concerns and develop appropriate responses in advance. Leaders also reexamined the skills that service teams would need, developing new talent profiles (and associated capability building modules) that could evolve with the workflows. Changing the internal collaboration model set the foundation for a later, gen-AI-based self-service option, while analytic AI tools could optimize staff allocation to provide additional call center coverage.

释放员工的能力。员工旅程是最后一块拼图。公司分析了座席体验的每一步，从登录到解决客户咨询和完成任务。这项分析包括简化流程，降低座席人员必须与之交互的技术系统的复杂性。这家电信供应商还发现了座席人员激励机制与客户需求之间可能存在的偏差，确保对座席人员进行激励，使其优先考虑客户满意度和客户问题的解决，而不是简单地处理大量电话。

Freeing up employees’ capacity. Employee journeys were the final piece of the puzzle. The company analyzed every step of the agent experience, from logging in to resolving customer inquiries and completing tasks. This analysis involved streamlining processes and reducing the complexity of technology systems that agents had to interact with. The telecommunications provider also identified potential misalignments between agent incentives and customer needs, ensuring that agents were given incentives to prioritize customer satisfaction and resolution rather than simply handle a high volume of calls.

通过将工作流程优化作为关键旅程的一部分，这家电信供应商利用生成式人工智能在一系列模拟和技术改进的基础上，显著而持久地改善了客户服务功能。总呼叫量减少了约 30%，平均处理时间减少了四分之一以上，同时服务质量也得到了改善：首次呼叫解决率提高了 10 到 20 个百分点。

By taking an integrated approach to workflow optimization as a part of critical journeys, the telecommunications provider achieved significant and lasting improvement in its customer care function, with gen AI building on a range of analog and tech-based improvements. Total call volume fell by about 30 percent, and average handle time by more than one-quarter, even as service quality improved: first-call resolution rates rose by ten to 20 percentage points.

规模化的（人类）秘诀

The (human) secret to scale

与以前的技术一样，生成式人工智能的全部潜力取决于其在整个组织中的规模。达到这一点的公司寥寥无几。他们的经验强调了四个要素的重要性，所有这些要素都以人为中心，而不是以技术为中心。前两个要素提供了关键的指导；后两个要素则更直接地改变了人们的工作方式，尤其侧重于变革管理。

As with previous technologies, gen AI’s full potential depends on its reaching scale throughout an organization. Few companies have reached this point. Their experience underscores the importance of four elements, all of which center on humans rather than technology. The first two elements provide critical guidance; the second two more directly change the way people work, with a particular focus on change management.

治理

Governance

生成式人工智能的成功部署不可能是临时性的。这不仅是因为生成式人工智能的风险已广为人知--从基于生成式人工智能的工具训练数据不准确到产生错误结果的 “幻觉”，还因为最先进的组织（从人工智能中获得的息税前利润超过 10% 的组织）倾向于集中管理其生成式人工智能计划。在这些表现出色的企业中，几乎有一半的企业都报告了集中管理的情况，而在其他企业中，这一比例仅为 35%。

Successful deployment of gen AI can’t be ad hoc. This is due to not only gen AI’s well-publicized risks—from inaccurate training data for gen-AI-based tools to “hallucinations” that produce incorrect results—but also the tendency of the most advanced organizations (the ones generating more than 10 percent of EBIT from gen AI) to centralize their gen AI initiatives. Almost half of these high performers report centralizing compared with only 35 percent of other companies.

治理结构的组成部分有助于支持快速实施和共同标准（见图 4）。明确的决策权对于评估生成式人工智能提案尤为重要，透明的审核流程为每个阶段关口提供了明确的标准。

The components of the governance structure help support rapid implementation and common standards (Exhibit 4). Clear decision rights are especially important for assessing gen AI proposals, supported by a transparent vetting process with well-articulated standards for each stage gate.

绩效基础设施、数据和分析

Performance infrastructure, data, and analytics

绩效基础设施的现代化对于适应生成式人工智能对工作格局的改变至关重要。第一步是重新定义指标，以反映公司新的运营战略，并让领导者了解生成式人工智能在整个组织中的进展。这些指标可以帮助组织产生并维持积极成果。接下来，通过一个有明确取舍标准的、有章可循的阶段性审查流程，将那些只是有前景的部署与最有成效的部署区分开来。最后，通过对生产力提升、客户体验改善和相关产出的更精确衡量，企业可以针对人类员工制定辅导和培训计划，并在生成式人工智能表现不佳时进行干预。

Modernizing performance infrastructure is crucial to accommodate gen AI’s changes to the work landscape. The first step is redefining metrics to reflect the company’s new operational strategy—and to allow leaders to see how gen AI itself is progressing across the organization. Such metrics can help the organization generate and sustain positive results. Next, a disciplined, stage-gated review process with clear go/no-go criteria separates the merely promising deployments from the ones most likely to be productive. Finally, with better measurement of productivity gains, customer experience improvements, and related outputs, companies can tailor coaching and training programs for human workers and interventions when gen AI’s performance lags.

变革管理

Change management

众所周知，改变技术并不是组织转型中最困难的部分——改变人们的工作方式才是最困难的。早期的经验似乎表明，对于生成式人工智能而言，这一点尤其正确。对于生成式人工智能，一个很好的经验法则是“公司每花一美元开发模型，就应该计划花三美元进行变革管理。”

It’s a truism that changing technology isn’t the hard part of transforming an organization—it’s changing how people work that’s hard. Early experience seems to show that this is even more true for gen AI, for which a good rule of thumb is “for every dollar spent on model development, a company should plan to spend three dollars on change management.”

沟通是起点。通过提供最新进展并解决潜在焦虑，企业可以促进未来采用并创造一种员工之间相互理解的文化。但比与员工交谈更好的是倾听：他们的专业知识和知识可以带来基于生成式人工智能的稳健、经济高效的解决方案，而不是影响甚微的生成式人工智能噱头。同时，技能提升和再培训计划可以帮助平稳过渡。

Communication is the starting point. By providing updates on what to expect and addressing potential anxieties, organizations can promote future adoption and create a culture of understanding among employees. But even better than just speaking to staff is to listen: the expertise and knowledge they contribute can make the difference between robust, cost-effective gen-AI-based solutions and gen AI gimmicks with little impact. In parallel, upskilling and reskilling initiatives can help smooth the transition.

持续创新文化

Continuous innovation culture

对于像生成式人工智能这样的新技术，庆祝成功和分享最佳实践尤为重要，因为这种技术的创新周期很短。仅仅跟上最新的机会就需要付出努力并保持开放的心态：这不是“购买还是构建”的问题，而是“购买和构建”的问题，需要不断审视市场提供的机会。

Celebrating successes and sharing best practices is especially vital with a new technology such as gen AI, where innovation cycles are short. Simply keeping abreast of the latest opportunity requires both effort and openness: it’s not a question of “buy versus build” but “buy and build,” continually reviewing what the market offers.

企业可以营造一种环境，让一线员工感到有权力提出想法，无论这些想法来自何处，并让他们自由地重新审视关于合作伙伴和供应商在创新采购中的潜在作用的假设。通过鼓励通过反馈和创新不断改进，企业可以优化代理商和客户的体验，同时最大限度地提高生成式人工智能的价值。

Organizations can foster an environment where frontline workers feel empowered to contribute ideas, whatever their source—and where they feel free to reexamine assumptions about the potential role of partners and vendors in sourcing innovation. By encouraging continuous improvement through feedback and innovation, organizations can optimize the agent and customer experience while maximizing the value of gen AI.

以一家欧洲领先的媒体和电信公司为例进行说明。该公司的使命是在 2024 年前实现生成式人工智能的产业化和规模化，并有望在一年内取得实际效益。该公司的做法不仅仅是追赶最新的技术潮流，而是要增强员工的能力，改变客户体验。为了将愿景变为现实，该公司确定了一个极具影响力的用例：由生成式人工智能驱动的副驾驶，旨在让客服人员在通话过程中更快、更有效地检索知识。

To illustrate, consider the case of a leading European media and telecommunications company. This organization embarked on a mission to industrialize and scale gen AI by 2024, with tangible benefits expected within another year. The company’s approach was not merely about chasing the latest tech trend; it was about empowering its workforce and transforming the customer experience. To bring its vision to reality, the company identified a high-impact use case: a gen-AI-powered copilot designed to equip customer service agents with faster and more effective knowledge retrieval during calls.

公司的首要任务是让座席人员了解情况并参与其中，公司每周都会组织工作组收集有关可用性和设计的定性反馈。此外，还通过代理对生成式人工智能生成的回复进行评分来收集定量反馈。“办公时间 "提供了一个提问和项目更新的论坛，培养了代理的主人翁意识。这种透明度有助于减少潜在的挫折感，确保代理人对副驾驶员的成功感到投入，并促成设计的实质性改变。

Keeping agents informed and engaged was a top priority for the company, which hosted weekly working groups to gather qualitative feedback on usability and design. Additionally, quantitative feedback was collected through agent ratings of the AI-generated responses. “Office hours” provided a forum for questions and project updates, fostering a sense of ownership among agents. This transparency helped mitigate potential frustrations and ensured that agents felt invested in the success of the copilot—and led to substantial changes in design.

事实证明，以用户为中心的方法不仅有助于完善协同功能，还能促进成功扩展。通过让虚拟代理尽早参与这一过程，该公司确保解决方案能够解决当前流程中的实际问题，并改善客户服务和代理体验。最终，代理在查找相关知识方面的平均处理时间缩短了 65%。

The user-centric approach proved instrumental not only in refining the copilot but also in encouraging successful scaling. By including frontline agents early in the process, the company made sure that the solution solved real problems in current processes and improved customer service and agent experience. The end result was a 65 percent reduction in average handle time for agents in finding relevant knowledge。

文章来源：知识搬运局

感恩遇见，共赢未来！我的微信，欢迎添加！！