2026 年，智能体将在企业级应用中取得哪些实质性突破？点击下载《2026 年 AI 与数据发展预测》白皮书，获悉专家一手前瞻，抢先拥抱新的工作方式！
请企业高管介绍一款 AI 智能体，你很可能会听到一个简洁的回答：“它负责总结文档”或“它处理一线客服工单。”但如果继续问，这个智能体如何接受治理、如何衡量效果、出现问题后如何纠正，你往往会听到沉默。此刻，智能体 AI 的真实处境，就藏在这份沉默之中。
我经常与全球领先企业的数据和 AI 负责人交流，探讨如何将智能体部署到生产环境，而不只是停留在试点阶段。那些走在前面的组织，胜出的原因并非模型更好、预算更多，而是它们不再把智能体当作需要精心设计提示词的工具，而是开始把它们当作需要安排到具体岗位上的“员工”。
提示词描述任务，岗位说明书定义角色
提示词告诉智能体在某个时刻要做什么。岗位说明书则更进一步：这个角色需要具备什么能力？它在组织中承担什么职责？谁对它的产出负责？
许多企业跳过了编写岗位说明书这一步。与智能体 AI 相关的首要担忧之一，是如何保持人工监督并防止智能体擅自行动。这也引出了一个问题：谁在管理这些智能体？又有谁培训过这些管理者？
我交流过的企业领导者通常会围绕以下三个审慎步骤来规划智能体项目，而每一步都对应着组织招聘新员工时会做的事。
有计划地招聘
在编写一行智能体逻辑之前，先要问：什么任务值得自动化？一家音乐科技公司通过审查工作流程，找出重复性最高、处理量最大的工作。如今，该公司的音乐检测智能体能够对 AI 生成的曲目进行分类，其处理规模是任何人工团队都难以持续承担的。
一家全球汽车制造商则在各业务部门设立 AI 负责人，通过结构化的创意征集挖掘用例，再按照优先级评估框架进行筛选，只有通过筛选的用例才会进入开发阶段。正如一位高管所说：“现在做智能体变得太容易了，所以团队往往会先动手。你们有流程吗？有阶段门槛吗？有衡量指标吗？大多数组织还没有做到这些。”
设计与优先级排序同样重要。智能体和员工一样，不会孤立地工作，信任也需要从一开始就建立。一家数字营销机构现在设置了解决方案架构师这一岗位，负责在 AI 团队与业务部门之间进行协调，确保智能体以受治理的工作流形式上线，并与实际目标紧密关联，而不是作为孤立工具被生硬地添加到现有流程中。
准备好员工手册
如果说岗位说明书是录用通知书，那么数据就是员工手册。一个无法访问准确、多样且及时的企业数据的智能体，就像一名从未完成入职培训的员工。然而，数据准备度仍是企业采用 AI 面临的首要障碍：
Snowflake 的一项研究显示，65% 的企业认为打破 AI 数据孤岛具有挑战性或非常具有挑战性，62% 的企业认为，要让数据具备 AI 应用条件，同样面临挑战或非常大的挑战。
解决办法不是增加更多数据，而是构建结构化的数据。一位数字营销机构的负责人说：“人们常常低估数据结构的重要性。他们以为智能体什么都能做，但模型并非如此运作。它们会挑选信息。要让智能体发挥作用，入职培训所用的数据必须经过精心整理。”
一家能源解决方案公司在为维修工程团队构建智能体时，就深刻体会到了这一点。一位负责人说：“搭建智能体大约只需要 20 分钟。但要从分散的数据源中收集资料，并构建它所需的嵌入，工程工作要花上好几周。”智能体的能力，取决于它拿到的“员工手册”质量。
积极管理
招聘和入职培训相对容易。项目能否成功，关键在于后续管理，因为管理智能体与管理员工截然不同，却同样需要严格的管理机制。
一家医疗科技公司的全球数据与 AI 负责人介绍了负责监督这一转变的人员正在形成的三种角色：设计者，负责构建流程和执行流程的智能体；控制者，负责监控执行情况并验证输出；审计者，则在高风险流程中独立核验智能体的工作。她告诉我：“亲自执行任务可能会越来越不再是必需的，甚至可能变成一种爱好。但人的责任不会消失。人们必须对设计中嵌入的决策、授予智能体的自主权，以及智能体产生的结果负责。”这种责任必须通过衡量来落实，而不能想当然地认为它自然存在。
一家全球汽车制造商会用 170 道问题对培训中的智能体进行五轮测试，然后以一条简单的标准决定是否部署：表现要胜过人类，而不能只是接近人类水平。一家音乐科技公司的运营团队则会在每条智能体输出交付前进行最终人工检查。我采访的每位领导者，在被问到最终由谁对智能体的行为负责时，都给出了相同的答案：人类。
这正是问题的核心。员工与新加入的智能体同事之间建立信任，需要时间。一位负责人估计，构建一个智能体需要四周，但让它真正被采用则需要八九个月，因为信任是在持续迭代中建立起来的，而不是在上线时宣布出来的。
下一张组织架构图上，会出现 AI 管理者
每个大规模部署智能体的职能部门中，都正在出现一个新岗位：AI 管理者。一位来自音乐科技公司的负责人告诉我：“下一阶段是管理多个 AI 智能体。你会成为一名管理者，负责管理这些 AI，让它们变得更高效。”
那些能够更快推进智能体应用的组织，都会明确智能体的岗位归属，审慎管理智能体所依赖的数据，并严格核查智能体的工作。技术就绪的速度，已经快于大多数管理架构的调整速度。“我们拥有智能体”和“我们有效地雇用了智能体”之间的差距，并非技术差距，而是管理差距。弥合这一差距，始于一个简单的观念转变：停止只写提示词，开始把智能体当作员工来安排和管理。
在完整电子书 《有效雇用智能体：数据高管的智能体企业蓝图》 中，我深入介绍了文中提到的每一段交流，包括一家全球信息服务提供商如何利用 175 年积累的法律与税务专业知识，培训其智能体员工队伍。
原文地址：https://www.snowflake.com/en/blog/stop-prompting-start-employing-agentic-enterprise/
点击链接立即报名注册：https://www.snowflake.com/events/snowflake-ascent-china/更多 Snowflake 精彩活动请关注专区
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
What substantive breakthroughs will AI agent achieve in enterprise-level applications in 2026? Click to download the White Paper " AI 2026 and Data Development Forecasts " and learn that experts look ahead and embrace new ways of working!
An AI AI agent, which you are likely to hear, is a simple answer: "It's for wrapping up the file" or "It's for processing the first-line passenger list." But if you continue to ask how the AI agent accepts governance, how to measure impact, how to correct problems when they arise, you often hear silence. At this moment, the true situation of AI agent AI is hidden in this silence.
I have regularly shared data with leading global enterprises and AI leaders on how to deploy AI agent to the production environment, rather than just at the pilot stage. The reason for those organizations that are in the front is not that they have better models and more budgets, but that they no longer see AI agent as a tool that requires careful design of prompt, but instead begin to treat them as “employees” who need to be placed in a specific position.
prompt describes tasks, job description defines roles
prompt told AI agent what to do at some point. The job description goes a step further: what does the role require? What is its role in the organization? Who is responsible for its output?
Many enterprises have gone beyond preparing job descriptions. One of the first concerns related to AI agent AI is how to maintain manual supervision and prevent unauthorized action by AI agent. It also raises the question: who is running AI agent? Who trained these managers?
The business leaders I have communicated usually plan the AI agent project around three prudent steps, each of which corresponds to what the organization would do when recruiting new staff.
Planned recruitment
Before preparing a line of AI agent logic, the question arises: What are the tasks worth automating? A music technology company reviewed the workflow process to identify the most repetitive and processed tasks. Today, the company ' s music test AI agent is able to classify the tracks generated by AI, the scale of which cannot be sustained by anyone ' s team.
A global car manufacturer has set up AI directors in various business units, which are screened in accordance with the priority assessment framework through structured creative call-for-excavations, and will only enter the development phase through the screening of examples. As one executive said: “It's too easy to do AI agent now, so teams tend to do it first. Do you have a process? Do you have a stage threshold? Do you have a measure? Most organizations have not done so.”
Design is as important as prioritization. AI agent, like its staff, does not work in isolation and trust needs to be built from the outset. A digital marketing agency now has the position of Solutions Architect, who coordinates between the AI team and the business sector to ensure that AI agent is online in the form of the governed workflow and closely linked to actual objectives, rather than being added to existing processes as an isolated tool.
Prepare the employee manual.
If the job description is an entry letter, the data is the employee manual. A AI agent, which does not have access to accurate, diverse and timely business data, is like an employee who has never completed his induction training. However, data readiness remains the primary obstacle to the adoption of AI by enterprises:
A study by Snowflake shows that 65% of businesses consider breaking the AI data island to be challenging or very challenging, and 62% consider that the same or very challenging challenge is to make the data available for AI applications.
The solution is not to add more data, but to construct structured data. “The importance of data structures is often underestimated. They thought AI agent could do anything, but the model didn't work. They select information. To make AI agent work, the data used for induction training must be carefully collated.”
This was deeply felt by an energy solution company when it built AI agent for the maintenance engineering team. One of the officials said, "It will take about 20 minutes to build AI agent. However, to collect information from a dispersed data source and build the embedded that it requires, engineering work takes weeks.” The ability of AI agent depends on the quality of the “staff manual” it receives.
Active management
Recruitment and induction training are relatively easy. The key to the success of the project lay in follow-up management, since management of AI agent was distinct from that of management staff, but also required rigorous management mechanisms.
The head of Global Data and AI, a medical technology company, described three roles that were being developed by the person responsible for overseeing the change: the designer, who was responsible for constructing and implementing the process AI agent; the controller, who was responsible for monitoring implementation and validating the output; and the auditor, who was independent in verifying AI agent in a high-risk process. She told me: “The task in person may become less and less necessary and may even become a hobby. But human responsibility does not disappear. People must be held accountable for the decision-making embedded in the design, the granting of autonomy to AI agent and the results produced by AI agent.” That responsibility must be implemented by measurement, and it cannot be assumed to exist naturally.
A global car manufacturer would use 170 questions to test AI agent for five rounds of training, and then decide whether to deploy by a simple criterion: performance is better than humans, not just close to humans. The operating team of a music technology company conducts a final manual check before each AI agent output is delivered. Every leader I interviewed gave the same answer when asked who was ultimately responsible for the actions of AI agent: Human.
That is at the heart of the problem. Building trust between staff and new AI agent colleagues takes time. One of the principals estimated that it would take four weeks to construct a AI agent, but that it would take eight or nine months for it to actually be adopted, because trust was built in a continuum, not announced on the line.
The next organizational chart will see an AI manager
A new post is being created in each of the functional units where AI agent is deployed on a large scale: AI Manager. "The next stage is to manage multiple AI AI agent. You'll be a manager, run these AIs, make them more efficient."
Organizations that are able to move forward with AI agent applications more quickly will identify the job titles of AI agent, carefully manage the data on which AI agent relies and rigorously verify the work of AI agent. Technical readiness has been faster than most management structures. The gap between “we have AI agent” and “we have effectively employed AI agent” is not a technology gap, but a management gap. Bridging this gap begins with a simple conceptual shift: stop writing only prompt and begin to organize and manage AI agent as an employee.
In the complete e-book, AI agent: A Blueprint for Effective Employment of AI agent Enterprises for Data Executives, I provided an in-depth account of each of the exchanges mentioned, including how a global information service provider could train its AI agent workforce using the legal and tax expertise accumulated in 175 years.
Original address: https://www.snowflake.com/en/blog/stop-prompting-start-emplying-agent-entry/
Click on the link to register immediately: https://www.snowflake.com/events/snowflake-ascent-china/ More Snowflake
Auto-fetched and translated by AI Atlas. All rights belong to the original author.