当 Agent 真正进入企业流程，哪些工作可以交给 AI ？50 + 实战案例揭晓答案！
AI 正在从回答问题的辅助工具，走向理解目标、调用系统、完成任务的工作承担者。当 Agent 真正进入企业流程，问题不再是 “如何把 AI 做出来”，而是哪些工作可以交给 AI、结果如何验收、权限、成本与风险如何收住。随着大模型向推理、工具调用、多模态和长任务延伸，开发重心正从单次调用转向上下文、工具、执行环境与反馈回路的设计，Coding Agent 也从代码补全演进为能理解代码库、运行测试、完成复杂任务的工程协作者。OpenAI、Google Cloud、Anthropic 的最新实践共同指向一个结论：模型决定 Agent 能做什么，工程决定它能否真正进入生产、创造价值。
在这一背景下，2026 年 AICon 人工智能开发与应用大会 · 北京站正式启动。本次大会将于 12 月 18–19 日，在北京新云南皇冠假日酒店举办。本届大会以探索 A 应用边界为主题，设置 13 大技术专题，覆盖大模型后训练、推理优化、Agent 工程、知识工程、智能体评测与安全、自进化系统、AI 软件交付、原生架构及组织变革等核心方向，汇聚一线技术专家，深入分享 AI 技术突破与企业规模化落地的真实实践。
帆软 CPO & AI 转型负责人张驰已确认出席 “AI 时代的组织重构” 专题，并发表题为《给 AI 业务单独划一块地：帆软怎么重新安排事、人和组织》的主题分享。过去一年多，帆软团队一边推动内部 AI 转型，一边孵化了悟帆、Dora 等 AI 产品。回头看，要把这件事做成，关键要过三关。第一，事怎么做。 小团队里，产品、设计、研发、测试依然有各自的分工，但彼此之间的交集很多。大家围绕能跑起来的东西不断迭代，过程中也踩过架构债、测试跟不上的坑。第二，人从哪来。 帆软团队没有单纯根据履历挑人，而是先把 AI 工具和使用额度发下去，看谁用得最多、具体用来做了什么。人是跑出来的，而不是选出来的。反而是做过成功项目的人，有时更容易受到过往成功经验的惯性影响。第三，组织怎么托底。 AI 转型风险大、调整多，放在原有体系里很难做成。因此，帆软团队把 AI 业务单独拉出来，实行独立核算，对失败和调整的容忍度也比老业务高得多。张驰在这次演讲中，希望把这三个层面的实践，以及过程中踩过的坑，讲清楚。
张驰，帆软 CPO & AI 转型负责人。整体负责帆软 AI 产品业务。软件工程专业出身，2014 年加入帆软，入职约一个月就参与了后来成为简道云的内部创业，从 0-1 负责简道云产品构建。2025 年起牵头帆软内部的 AI 转型，他负责的 AI 业务先后孵化了 Dora、悟帆 等产品。
他在本次会议的详细演讲内容如下：
演讲提纲
1. 事：AI 时代的产品研发协作模式重构
- 产品、设计、研发、测试如何从线性交接走向交集协作
- 坑一：复杂产品仍离不开专业研发，如何重新划分职责边界
- 坑二：编码提速后，测试与发布成为新瓶颈，如何建立新的质量保障机制
2. 人：AI 创新人才是跑出来的，不是选出来的
- 如何通过 AI 工具使用、需求征集和黑客马拉松，在实践中识别合适的人才
- 坑一：历史成功经验可能成为创新包袱，如何摆脱过去的增长惯性
- 坑二：核心 AI 人才的激励与留存，以及小团队如何补齐领域专家和业务资源
3. 组织：通过组织隔离与独立核算，为 AI 创新留出空间
- 为什么要将 AI 业务从原有组织体系中独立出来，减少传统机制和客户交付惯性的影响
- 如何通过独立预算、团队授权和阶段性里程碑管理，容忍试错、推动业务验证
- 坑：组织隔离后，如何避免前台创新与后台专家、交付、运维支持脱节
4. 总结与展望
听众收益
- 了解 AI 如何改变产品、设计、研发、测试之间的协作方式，以及如何划分职责、控制架构债与测试质量风险
- 了解如何通过 AI 工具使用、实际产出和项目实践识别合适的人才，以及如何突破过往成功经验带来的路径依赖
- 了解如何通过独立核算、预算授权和阶段性里程碑，为高风险、快迭代的 AI 业务提供试错空间，并平衡创新效率与交付保障
除此之外，本次大会还策划了Post-Training Engineering｜数据、奖励与模型能力塑造、大模型推理与算力调度实践、Knowledge Engineering｜让 Agent 理解企业、Multi-Agent Systems、智能体评测与生产可靠性探索、Agent Security｜身份、权限与行动安全、AgentOps｜企业 Agent 全生命周期管理、Self-Evolving Agents｜从经验学习到可控进化、Agentic Software Engineering｜从代码生成到软件交付、Personal Agent、AI 原生架构的演进与治理、AI 时代的组织重构、AI × Industries ｜产业智能化等 13 个专题论坛，届时将有来自不同行业、不同领域、不同企业的资深专家在现场带来前沿技术洞察和一线实践经验。
更多详情可扫码或联系票务经理 18514549229 进行咨询。
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。
When Agent actually enters the business process, which jobs can be given to AI?50 + the battle cases reveal the answer!
AI is moving from a question-and-answer support tool to an understanding of the goal, to a system, and to a duty bearer. When Agent actually enters the business process, the question is no longer “how to make AI”, but what work can be given to AI, how results are acceptance check, permission, how costs and risks are collected. With the extension of LLM to reasoning, tool call, multimodular and long missions, the development of the design of a shift in focus from a single call to context, tools, implementation environment and feedback loops, Coding Agent also evolved from a code completion to an engineering collaborator who understands the code library, runs tests, and completes complex tasks. Open AI, Google Cloud, Anthropic ' s latest practice points to one conclusion: the model determines what Agent can do, and the project determines whether it can actually enter production and create value.
In this context, the AICon Congress on Artificial Intelligence Development and Application, Beijing Station, was officially launched in 2026. The Congress will be held at the New Yunnan Crown Holiday Hotel in Beijing from 18 to 19 December. The Congress, which focused on exploring the A application boundary, set up 13 major technical topics covering the core directions of LLM post-training, reasoning optimization, Agent engineering, intellectual engineering, AI agent assessment and security, self-evolution systems, AI software delivery, native architecture and organizational change, bringing together first-line technical experts to share in depth the real practices of AI technological breakthroughs and enterprise size.
Zhang Zhang Zhiqi, Head of the CPO & AI Transition, has confirmed his presence on the theme “The reorganization of the organization of the AI era” and has published a theme-sharing entitled “Single space for AI operations: how to rearrange, people and organizations”. Over the past year, the swagger team has been pushing for internal AI transformation while hatching swabs, Dora and other AI products. Looking back, it's three things to do. First, how to do it. Within small teams, products, designs, research and development, and testing continue to have separate divisions of labour, but there is a great deal of overlap. You're constantly evolving around things that can run, and in the process you've stepped on architecture debts and tested the pits that can't keep up. Second, where people come from. Instead of simply selecting people on the basis of their résumés, the team sent out the AI tools and the use of quota to see who used them most and specifically for what. People run away, not elected. Rather, those who have done successful projects are sometimes more vulnerable to the inertia of previous successful experiences. Third, how to get to the bottom of Division. AI transition risks are high and adjusted and are difficult to achieve in the old system. As a result, the team pulls out AI operations alone and is self-accounting, and tolerates failures and adjustments much more than old ones. In this presentation, Zhang Zhang wishes to make clear the practice of these three levels, as well as the pits that have been trampled through.
Zhang Zhi, Navigator CPO & AI Transition Manager. I'm in charge of all of this. The software engineering professional, who joined the swab in 2014, became involved in the internal business of becoming a simple cloud for about a month, from 0-1 to build a simple cloud product. Since 2025, leading the internal transformation of the sails, his AI business has hatched products such as Dora and Gospel.
His detailed presentation at this meeting reads as follows:
Outline of the presentation
1. Matters: Reconstruction of the production collaboration model of the AI era
- Products, design, research and development, testing how to move from linear intercourse to intersectional collaboration
- Pipe One: complex products still have to be developed professionally and how to redefine the boundaries of responsibility
- Pipe II: After coding speed, testing and publish became new bottlenecks and how to establish new quality assurance mechanisms
2. People: AI Innovators are out, not selected
- How to identify the right talent in practice through the use of AI tools, demand collection and hacking of marathons
- Catalysing one: history ' s success experiences can become a burden of innovation and how to move away from past growth inertia
- Pipe II: Core AI motivation and retention of talent, and how small teams can complement expert and operational resources in the field
3. Organization: space for AI innovation through organizational isolation and independent accounting
- Why separate AI operations from the old organizational system and reduce the impact of traditional mechanisms and customer inertia
- How to tolerate mistakes and facilitate operational validation through independent budgeting, team delegation and milestones management
- Pits: how to avoid a disconnect between front-office innovation and back-office experts, delivery, and back-office support after organizational isolation
4. Summary and outlook
Audience benefits
- Understanding AI how to change the way products, designs, research and development, and testing work together, and how to divide responsibilities, control architecture debt and test quality risks
- Understanding how to identify the right talent through the use of the AI tool, actual outputs and project practices, and how to break the path-based dependence of past success stories
- Understanding how to provide trial error space for high-risk, fast-temporal AI operations through independent accounting, budget delegation and milestones, and balance innovation efficiency with delivery assurance
In addition to this, the Congress planned Post-Training Engineering data, awards and modelling capability development, LLM reasoning and arithmetical scheduling practices, Knowridge Engineering to Agent to understand enterprises, Multi-Agent Systems, AI agent to assess and explore the reliability of production, Agent Security, permission and operational security, AgentOps| Enterprise Agent management, Self-Evolving Agents| From experience learning to manageable evolution, Agenic Software Engineering from code generation to software delivery, Personal Agent, the evolution and governance of the AI original architecture, the organization of the AI era, AI x Industries, industrial intelligence, etc., 13 thematic forums will bring forward technological insights and first-line practical experience from different industries, fields and firms.
More details can be cleared or consulted with the ticket manager 18514492229.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.