AI 时代的企业转型:我相信自由,也相信边界
看完 Pedro Franceschi 关于企业 AI 的访谈后,有一个观点一直留在我脑中:CEO 应当成为公司的 Chief AI Officer。
我认同其中的一部分,但也保留另一部分。
我认同的是,公司的管理层必须更积极地拥抱 AI,主动试探它的能力边界,并真正理解当下的技术变化。AI 已经不再只是某个部门采购的一项工具,也不只是工程团队或产品团队的事。它会影响公司的资源如何分配、团队如何协作、流程如何重构,以及企业未来会把什么能力留在内部。
如果管理层不了解 AI 的能力与局限,就很难判断哪些流程值得改造,哪些环节可以自动化,哪些岗位需要新的协作方式。只负责“大方向”和“大环境”已经不够了。因为技术变化一旦进入组织的日常运转,资源浪费往往不是来自不努力,而是来自仍在用旧结构处理新问题。
但我不完全认同“遇到大多数问题都先让 AI 自己解决”的乐观叙事。
AI 的能力边界,不等于它的授权边界
AI 能做什么,与企业应该允许它做什么,是两件不同的事。
我相信 AI 会接管大量重复、规则相对明确的工作。过去需要专门由人完成的信息整理、初步分析、分类、记录、检索和流程衔接,很多都可以交给 AI。它能够显著降低重复劳动,也能让团队把时间用在更复杂、更有价值的事情上。
但这并不意味着 AI 的输出可以被直接相信,更不意味着企业应把所有决策权交给它。
真正的商业落地,需要先界定清楚:AI 每一个步骤可以做什么、不能做什么;它可以访问哪些信息;每一次动作如何记录;结果如何审核;人需要在哪个节点介入;出现错误时由谁负责、如何追溯、如何纠正。
我更愿意把企业里的 AI 理解为一种“受控的数字执行者”。它可以获得较高的效率与一定程度的自由,但这种自由必须建立在明确授权、可追溯、可撤销的边界之内。
给 AI 更大的自由,也给企业更强的控制力
访谈中提到一个我很认同的方向:让 AI 在企业内部网络中运行,并对它在接口和 HTTP 请求之间的活动进行审计、回溯与持续优化。
我认为,这种思路在德国以及欧洲企业环境中尤其值得进一步发展。
企业可以搭建一个内部网络,在其中为 AI 提供完成内部任务所需的权限与自由度;与此同时,外部网络中的信息访问应当被更严格地限制。甚至可以部署两个不同的 AI:一个只在内部网络中处理企业知识和内部流程,另一个只负责搜集、浏览和处理外部公开信息。
两者之间不应直接混合。外部信息进入内部系统前,应被校验、记录和审核;内部数据也不应因为一个 Agent 的便利而失去边界。这样既能持续吸收外部信息,又能尽可能保护企业内部的数据安全。
在我看来,真正重要的并非“让 AI 更自由”或“让 AI 更受限”二选一,而是让自主性与控制力同步增长。AI 的权限越高,身份、数据、工具和审计的边界就越应该清楚;人对关键节点的把控也越不能缺席。
自我优化不能替代人的判断
AI 不应该只是一次性部署后就停在原地的工具。它需要从错误、例外和人工修正中不断复盘、不断优化,逐步提高处理问题的质量。
但“自我优化”不应被理解为“无需监督地自我进化”。企业仍需要把人工判断放在关键环节:检查 AI 的输出,识别它是否误解上下文、是否遗漏重要信息、是否在看似合理的表达中给出了错误结论。特别是在涉及客户、资金、招聘、绩效、合规和对外承诺时,最终判断不能只依赖模型。
我相信好的 AI 系统应该让每一次错误都留下记录,让每一次人工介入都成为下一轮改进的依据。人不必重复完成所有工作,但必须保留判断标准、责任意识和最终的控制权。
开源与私有部署的战略意义
随着模型能力不断提升,企业 AI 的商业落地已经比过去清晰得多。对于许多中型企业而言,真正的选择不再只是“要不要使用 AI”,而是“哪些能力应依赖外部服务,哪些能力应沉淀为自己的基础设施”。
我尤其关注开源模型和企业内部部署的可能性。只要企业愿意投入算力、显卡、存储、服务器、机房与散热等基础设施,就有机会在内部网络中运行更属于自己的 AI 能力。这样的投入未必会在短期内带来最直接的回报,但从长期看,它可能带来更高的数据控制力、更稳定的内部能力积累,以及更低的外部依赖。
这并不意味着闭源模型没有价值,也不意味着所有企业都应立即自建庞大的算力体系。真正重要的是,企业不能把数据安全、知识沉淀与关键业务能力完全外包给不可见的黑箱。尤其当 AI 开始进入企业内部流程时,谁拥有数据、谁控制权限、谁能审计行为,都会变成战略问题。
从 HR 角度看:专业能力会被放大
从 HR 的角度看,我认为 AI 时代最大的变化之一,是专业能力会被进一步放大。
同一个 AI 工具,交给不同的人,所产生的效率和结果并不相同。一个真正了解领域的人,更知道该问什么、该看什么、哪些结论需要怀疑、哪些例外不能忽略。他可以把问题描述得更精准,也能更快判断结果是否可靠。
以我的经验来看,面对同一个问题,专业人士可能三天就能完成的工作,缺少领域经验的人需要花一周,甚至仍无法确认 AI 的答案是否正确。AI 并没有消除专业性,反而把专业性从过去容易被忽视的背景能力,变成了能够直接决定效率和质量的核心杠杆。
因此,企业未来不应只培训员工“如何使用 AI”。更重要的是培养和识别三种能力:对业务的专业理解、对 AI 输出的验证能力,以及把复杂工作拆解为可交给 AI 执行、同时又能被人类控制的任务能力。
最后:AI 是效率问题,也是责任问题
我认为,AI 终究不只是技术或效率问题。
它也关乎人与人、人与机器、人与社会之间关系的重新定义。我们愿意交给机器哪些判断?当系统出错时,谁来承担责任?当效率被极大提高后,人保留的是决策权、价值判断,还是只剩下最终负责的义务?
这些问题没有简单答案,也会因行业、公司和个人而不同。
但有一点是明确的:企业不应因为 AI 的能力增长,就放弃对边界、责任和人的判断的重视。真正成熟的 AI 转型,不是把人从系统中移除,而是让人从重复劳动中释放出来,同时让人的专业性、责任感和判断力变得更加重要。
After watching Pedro Franceschi’s interview on enterprise AI, one idea stayed with me: the CEO should become the company’s Chief AI Officer.
I agree with part of that view, while reserving judgment on another part.
I agree that company leadership must embrace AI more actively, probe the boundaries of what it can do, and genuinely understand the technological shift now underway. AI is no longer simply a tool procured by one department, nor is it only the concern of engineering or product teams. It affects how a company allocates resources, how teams collaborate, how workflows are redesigned, and which capabilities an organization chooses to retain in-house.
If leadership does not understand both AI’s capabilities and its limitations, it will be difficult to judge which processes are worth redesigning, which steps can be automated, and which roles require new forms of collaboration. Merely taking responsibility for “the big picture” and “the broader environment” is no longer enough. Once technological change enters the day-to-day operation of an organization, wasted resources often come not from a lack of effort, but from continuing to use old structures to address new problems.
But I do not fully agree with the optimistic narrative that, when most problems arise, we should first let AI solve them on its own.
AI’s capability boundary is not the same as its authorization boundary
What AI is capable of doing and what an enterprise should permit it to do are two different things.
I believe AI will take over a large amount of repetitive work governed by relatively clear rules. Many activities that previously required dedicated human effort—organizing information, conducting preliminary analysis, classification, recordkeeping, retrieval, and handoffs between workflow stages—can increasingly be assigned to AI. It can significantly reduce repetitive labor and allow teams to spend more time on work that is more complex and more valuable.
But that does not mean AI outputs should be trusted without review, nor does it mean that enterprises should hand over all decision-making authority to it.
For AI to be deployed meaningfully in business, organizations must first define what AI may and may not do at every step; what information it may access; how each action is logged; how results are reviewed; where human intervention is required; and who is accountable, how incidents are traced, and how errors are corrected when something goes wrong.
I prefer to think of enterprise AI as a “controlled digital operator.” It can be given high efficiency and a degree of autonomy, but that autonomy must rest on boundaries that are explicitly authorized, auditable, and revocable.
More freedom for AI, and stronger control for the enterprise
The interview raised an approach I strongly agree with: allowing AI to operate within an enterprise network while auditing, tracing, and continuously improving its activity across APIs and HTTP requests.
I believe this approach is particularly worth developing further in German and wider European business environments.
An enterprise can build an internal network that gives AI the permissions and latitude required to complete internal tasks. At the same time, access to information on external networks should be more tightly restricted. It may even make sense to deploy two separate AI systems: one that handles enterprise knowledge and internal processes only within the internal network, and another that is responsible only for collecting, browsing, and processing publicly available external information.
The two should not be mixed directly. Before external information enters an internal system, it should be validated, logged, and reviewed. Internal data should not lose its boundaries merely for the convenience of an agent. This design makes it possible to continuously absorb external information while protecting enterprise data as much as possible.
To me, the important point is not choosing between “giving AI more freedom” and “restricting AI more tightly.” It is allowing autonomy and control to grow together. The more authority AI has, the clearer the boundaries around identity, data, tools, and auditing must be. Human control over critical decision points cannot be absent.
Self-improvement cannot replace human judgment
AI should not be a tool that is deployed once and then left unchanged. It needs to learn from errors, exceptions, and human corrections, continuously review its performance, and gradually improve the quality of its work.
But “self-improvement” should not be understood as unsupervised self-evolution. Enterprises still need human judgment at critical points: to inspect AI outputs, identify whether the system has misunderstood context, missed important information, or reached an incorrect conclusion through language that only appears plausible. Especially in matters involving customers, funds, hiring, performance, compliance, and external commitments, the final judgment cannot rest on a model alone.
I believe a good AI system should leave a record of every error and turn every human intervention into input for the next round of improvement. People do not need to repeat every task themselves, but they must retain the standards of judgment, the sense of responsibility, and ultimate control.
The strategic significance of open source and private deployment
As model capabilities continue to improve, the commercial deployment of enterprise AI has become much clearer than it was in the past. For many mid-sized companies, the real question is no longer simply whether to use AI, but which capabilities should rely on external services and which should become part of their own infrastructure.
I am particularly interested in the potential of open-source models and on-premises enterprise deployment. If a company is willing to invest in computing capacity, GPUs, storage, servers, data-center facilities, and cooling, it can create AI capabilities that are more its own and operate them within its internal network. Such investment may not deliver the most immediate return in the short term, but over time it can provide greater control over data, more stable internal capability building, and lower dependence on external providers.
This does not mean closed-source models lack value, nor does it mean every company should immediately build a large computing estate of its own. What matters is that enterprises should not outsource data security, accumulated knowledge, and critical business capabilities entirely to opaque black boxes. Once AI begins to enter internal workflows, who owns the data, who controls permissions, and who can audit behavior all become strategic questions.
From an HR perspective: expertise will be amplified
From an HR perspective, I believe one of the biggest changes in the AI era is that professional expertise will be amplified further.
The same AI tool will not produce the same efficiency or results in the hands of different people. Someone who truly understands a field knows better what to ask, what to examine, which conclusions deserve skepticism, and which exceptions cannot be ignored. They can describe a problem more precisely and judge the reliability of an output more quickly.
In my experience, for the same problem, work that a professional can complete in three days may take someone without domain experience a week—or may still leave that person unable to verify whether the AI’s answer is correct. AI has not eliminated expertise. Instead, it has turned expertise from a background capability that was often overlooked into a core lever that directly determines efficiency and quality.
For this reason, enterprises should not only train employees in “how to use AI.” More importantly, they should develop and identify three capabilities: professional understanding of the business, the ability to validate AI outputs, and the ability to decompose complex work into tasks that AI can execute while humans still retain control.
Finally: AI is a question of efficiency, but also of responsibility
I believe AI is ultimately not only a question of technology or efficiency.
It is also about redefining the relationships between people, between humans and machines, and between people and society. Which judgments are we willing to delegate to machines? When a system fails, who bears responsibility? When efficiency rises dramatically, do people retain decision-making authority and value judgment, or only the obligation to take responsibility in the end?
There are no simple answers to these questions, and they will differ across industries, companies, and individuals.
But one point is clear: enterprises should not abandon their attention to boundaries, responsibility, and human judgment simply because AI’s capabilities are growing. A mature AI transformation does not remove people from the system. It frees them from repetitive work while making their expertise, sense of responsibility, and judgment more important.