深入了解企业 AI 成功背后的 AI 采纳阶段

每家企业都会经历相同的 AI 采纳阶段，无论是否有计划。路径从零散的工具使用到首次试点，再到实际生产，最终形成整个业务依赖的共享系统。每一步都是逐步推进的，因此公司往往在外界注意到之前就已经进入了新阶段。

区分不同公司的关键在于它们在这条路径上走了多远，市场之间的差距相当大。麦肯锡 2025 年全球调查显示，88% 的组织已在至少一个业务职能中使用 AI，约三分之一已经开始规模化，7% 表示 AI 已完全规模化。这意味着几乎所有企业都已开始部署 AI，但真正成功的却寥寥无几。

Over the last few years I have worked with leadership teams at nearly every point on this path, and one pattern has held throughout. The stage a company occupies tells you more about what its AI program will deliver than the model it picked or the vendor it signed. Each of the AI adoption stages asks for a different kind of change inside the organization, and that change is what predicts the outcome. So let’s go through them one at a time, along with the point where companies usually get stuck moving to the next one.

AI 采纳的五个阶段以及推动公司前进的决策

Across the companies we work with, these five stages come in the same order almost every time. It usually begins with a few people using AI tools on their own, and leadership often has no idea until much later. Some companies never get past that. The ones that do keep going eventually reach a point where the business itself is built around what the technology can do. The thing worth understanding is that each stage asks for something different from the organization. The decision that got a company through its pilot phase is frequently the same decision holding it back a year later, and that is where a lot of AI programs quietly come apart.

1. 零散使用且无所有者

The first stage usually looks like nothing more than curiosity.  At this point, employees are using AI tools with no central policy behind them, procurement has no visibility into any of it, and nobody is measuring what that work produces. There is real information sitting in that disorder, because curiosity at the edges of a company shows you where the friction genuinely lives, and that signal almost never reaches the executive floor through a formal channel.

The mistake leaders make here is formalizing too early, which turns genuine experimentation into a governance exercise long before anyone has worked out what actually deserves governing.

此阶段的决策: 正确的直觉是观察正在发生的情况，而不是试图管控它，因此制定一项轻量的使用政策，涵盖数据处理，其余暂时不做干预。需要关注的是重复性，因为当相同的变通方案在三个不同团队中出现且彼此未协同时，就形成了值得正式投入的用例。

2. 有资金支持但成功标准薄弱的试点

The pilot shows up with a budget, a named owner, and a demo that goes in front of the board, and then the work quietly ends there. MIT NANDA 项目的研究found that the overwhelming majority of enterprise generative AI pilots produce no measurable impact on profit and loss.

导致这种情况的原因基本可以归结为三点：第一，成功标准围绕模型准确率而非业务真正关心的指标；第二，演示下缺乏真实的数据流水线；第三，发布结束后无人负责其结果。该阶段从内部难以辨识，因为表面上看似进展顺利，演示在改进，供应商争取下一阶段，委员会仍在开会，但这些都未能转化为财务可量化的数字。

此时的行动指令: 不要批准没有业务指标且缺乏在一年后仍能对该指标负责的负责人的试点，若无法命名则直接淘汰。四个遵循此纪律的试点胜过十二个人人称赞却缺乏指标的试点。

3. 生产集成

Of all the AI adoption stages, this is the one that ends the largest number of programs. Employees and customers are now depending on the system every day, and requirements start surfacing that no pilot ever had to meet, from monitoring and human escalation paths to version control on prompts and models, incident response, and data lineage that will hold up in an audit. The NIST AI 风险管理框架documents those controls in detail, and it is worth reading before a launch date gets set.

这里更棘手的问题不是技术而是组织层面，因为该阶段更看重工程纪律而非持续实验，像研究团队那样运作的团队在此会遇到困难，而领导层则会寻找技术解释。经济模式也随之转变。试点成本一次性且易获批准，而生产则形成永久性的运营费用，未曾预算的公司往往把首次续费请求误认为投资失误的证据。

上线前必须改变的事项: 将所有权在上线前就转移，而非之后；将交付交给有生产经验的人员；并为运营成本预留三年的预算。随后指定在模型退化时接收警报的负责人，因为这唯一的姓名决定了系统是可运行的还是仅是演示。

4. 共享能力与复用

At this point, AI turns into infrastructure that several business units are drawing on. A platform team maintains the reusable components, a standard way of evaluating things, and a deployment path, so the cost of every new use case comes down because nobody is rebuilding the foundation from scratch each time.

公司进入该阶段的最明显信号根本与技术无关。预算权限从创新团队转移至业务线负责人，当某部门负责人用运营预算资助 AI 工作并将回报计入自身业绩时，组织真正跨越了这一阶段。我们看到，这一转变对用例质量的提升，超过同一年内任何技术决策的影响。

资金应放置的位置: 将平台团队视为基础设施进行资助，而非项目，并以复用率而非上线数量来衡量。随后将 AI 预算从创新部门转移至各业务部门，因为一旦业务负责人出资，用例的针对性就会提升，无需额外干预。

5. AI 融入运营模型

Companies at this stage are designing products, processes, and decisions around AI capability from the start, and nothing gets retrofitted later. New offerings assume the capability is already there, roles change to reflect what the systems are handling, and planning cycles shorten because an idea can be tested against real data in a matter of days.

真正达到此阶段的组织寥寥无几，声称已达此水平的数量远高于实际符合条件的。测试方法非常简洁，只需在一次会议中提出:如果模型明天消失，公司会停止哪些工作。坦诚的答案往往让人不舒服，因为它指向收入而非便利。每次我们向领导团队提出此问题，得到的真实答案都比其战略稿所宣称的阶段低一至两个层级。

董事会层面的提问: 将 AI 能力视为战略资产，并以同等严格的标准审视。审查供应商集中度、模型依赖性和数据所有权，然后询问每个没有模型就无法运行的流程的备选方案是什么，并在需要之前为该备选方案提供资金。

公司跳过某阶段会发生什么：真实案例

We have seen enough companies, and the sequence is almost always the same: where a pilot goes well, the board likes it, everyone wants to show progress, and the company goes from demo to full rollout without stopping at production integration in between. Nobody wants to spend a quarter on monitoring and ownership when the demo already works, so they skip it. One of our clients did exactly that, and it cost them more than they expected.

出错之处

The client was a services business running a support workload, and the model handled the first-pass responses that used to sit with their team. It went live across the whole operation on the back of a pilot that had worked well, and for four months nothing looked wrong. When we came in and looked at what was actually running, there was no monitoring on the model, so nothing would tell them if the answers started going off. There was no escalation route for the cases a person should have checked. The team that built it had moved on to the next project and handed ownership to nobody, and because the pilot had been a one-off cost, no one had put an operating budget behind something that now ran every day.

造成的代价

The answers did start going off, and anyone watching would have caught it early, but nobody was watching, so a customer found it first. Fixing the model turned out to be the easy part and took about two weeks. Winning back the people who had relied on that output took the rest of the quarter, and leadership spent most of that same quarter arguing about whether the AI investment had been a mistake, when the mistake was how they launched it.

从中得到的教训

Monitoring and ownership look like overhead until you go without them, and leaving them out makes a launch feel cheaper than it is. The cost does not go away, it just shows up later as lost trust instead of a number on a budget.

如何识别贵公司当前的 AI 采纳阶段

Four questions place a company on this path faster than any formal assessment. Ask them with your operating leaders, and answer each one with evidence.

• 您能指出每个已部署模型推动的业务指标吗？
•  当模型出现退化时，谁会收到警报？
• 从需求提出到投入生产，一个新用例需要多长时间？
•  您已有的 AI 工作中，有多少是对已构建内容的复用？

Clear answers confirm the stage a company has reached. Uncertain ones point straight at the layer worth funding next, which makes them just as valuable to have.

The companies that pull ahead are rarely the ones with the best models, but the ones that read their position honestly and put money into what carries them to the next stage. That is harder than it sounds, because every team reports on its own progress and nobody wants to be the one saying their part is behind. The leaders who get this right ask the questions directly, and they check the answers against what is actually running.

Every organization already belongs to one of these stages, whether anyone has named it or not. The advantage belongs to the leadership teams who know which one, and who fund the next step deliberately.