夜雨聆风学习资料网

ARTICLE · 1148790

AI Agent驱动的工业软件智能体化——从ERP副驾到企业自主系统

AI Agent驱动的工业软件智能体化——从ERP副驾到企业自主系统

Industrial Software Agentification Driven by AI Agents: From ERP Copilot to Autonomous Enterprise

一、引言:一字之变,软件范式的分水岭

2026年5月,Gartner在一份研究报告中悄然替换了一个沿用四十多年的单词——将Enterprise Resource Planning中的Planning换为Execution,创造了ERX(Enterprise Resource Execution)这一新概念。同年9月,工信部印发《"人工智能+软件"专项行动实施方案》(工信部信发〔2026〕209号),明确提出到2028年在重点行业打造100个智能体软件标杆应用,推广应用覆盖2万家规模以上软件企业。
这不是巧合,而是产业共识的凝聚。当软件从"记录系统"(System of Record)向"行动系统"(System of Action)跃迁,AI Agent正成为这场范式转换的核心引擎。传统ERP被动等待人工录入、审批、执行——人成了系统间的"集成商",用Excel、微信和电话把断裂的流程手工缝合。约50%的ERP实施以失败告终,Gartner更预测到2027年70%的新ERP项目达不到预期目标。
Agent化后的工业软件则能感知环境、推理决策、自主执行,在预设边界内实现从"人找事"到"事自办"的根本转变。Gartner预测,到2026年底40%的企业应用将嵌入任务型AI Agent,而2025年初这一比例不足5%。到2030年,超过50%的ERP基础任务将由AI自主完成。一场从ERP副驾到企业自主系统的演进,正在加速推进。

二、工业软件Agent化现状:四大阵营初成

截至2026年9月,面向制造业的Agent服务商已形成四大阵营:
全栈生态型:SAP Joule、Oracle Fusion AI Agents以自家ERP为底座生长Agent能力层。SAP在2026年5月Sapphire大会上发布"自主企业"愿景,其Autonomous Suite涵盖自主财务、自主采购、自主供应链、自主人力和自主客户体验五大板块,预计到2026年底提供超400个AI Agent。Oracle则一口气推出22个预置Agent应用,覆盖财务、HR、供应链和客户体验。
AI原生型:Palantir以Foundry和AIP平台为核心,通过Ontology框架将企业运营规则编码进软件,实现数字孪生驱动的自主决策。2026年9月10日,Palantir与英伟达联合宣布将主权AI技术栈部署于英伟达供应链运营,基于Nemotron开源模型实现端到端可视化与机器级辅助决策。
垂直深耕型:方天软件提出"AI Agent将成为ERP的终极形态",其T-ONE系统部署了采购、排程、质检等数字员工智能体,"慧儿"AI产品可秒级自动生成精准报价。黑湖智造聚焦车间级协作智能,以轻量化MES+Agent架构服务中小型制造企业。
跨系统连接型:以RPA+Agent融合为特征,实在智能的实在Agent以API+GUI双轨执行模式非侵入式操作老旧系统,2026年7月以90.2%任务成功率登顶全球AI智能体基准OSWorld总榜,成为首个突破90%大关的Computer-Use Agent。

三、技术演进路径:副驾→工作流→自主系统

工业软件Agent化的技术演进呈现清晰的三级跃迁:
L1 嵌入式Agent副驾(2023-2025) :以SAP Joule为代表,AI作为Copilot嵌入既有系统,提供自然语言查询、异常提醒、建议生成。Joule最初定位为"数字助手",2024-2025年快速整合进S/4HANA、SuccessFactors、Ariba全产品线。这一阶段的核心价值是降低系统交互门槛,但决策权仍在人手中。Infor对此有一个精准判断:"提示词回答80%的正确率可以接受,但过到账本上、下达到车间的动作,80%的正确率不可接受。"
L2 跨系统自主工作流(2025-2027) :Agent从"被动响应"进化为"主动执行"。2026年1月,SAP Joule与Microsoft Copilot实现深度集成,Joule Agent与Copilot Agent可相互调用,构建多Agent协作生态。MCP(模型上下文协议)和A2A(Agent-to-Agent)等开放标准打通了系统间调用,使跨ERP、MES、SCM的自主工作流编排成为现实。SAP参考架构以Joule Work为统一入口、Joule Assistants协调跨职能Agent、Joule Agents执行具体多步任务。
L3 企业自主系统(2027-2030) :Gartner预测到2030年超50%的ERP基础任务由AI自主完成。Infor提出"智能体企业"战略,其Agentic Orchestrator基于MCP协议跨系统编排多智能体流程,强调"执行能力取决于智能体脚下平台的深度,而不是智能体本身的聪明程度"。在这一阶段,企业不再是"使用软件的人",而是"设定目标与边界的监督者"——Agent负责感知、推荐、执行常规动作,人负责设定目标、审批高风险决策、处理异常。

四、典型案例

案例1:SAP Autonomous Suite——400个Agent的全业务覆盖

(来源:SAP 2026半年报,2026年8月)
SAP在Sapphire大会上发布Autonomous Suite,涵盖财务、采购、供应链、人力和客户体验五大板块。核心架构为Joule Work + Joule Studio + Agent Gateway三层体系,到2026年底预计提供400+个AI Agent。Joule Agent覆盖SuccessFactors中绩效目标、职业发展、HR服务、人才智力和薪酬五大领域,同时推出业界首个劳动力知识网络,整合第三方行业最佳实践。

案例2:英伟达×Palantir——供应链主权AI部署

(来源:Palantir新闻稿,2026年9月10日)
英伟达将自身供应链运营全面交给Palantir Foundry和AIP平台管理,底层运行Nemotron开源模型。英伟达供应链横跨数百万零件、数千家供应商和全球制造伙伴网络,每套Vera机架系统包含130万个零部件。通过Ontology框架构建端到端数字孪生,实现瓶颈识别、运营知识编码和机器级辅助决策。Palantir 2026 Q2营收19.35亿美元,同比增长93%,美国商业收入增长149%。

案例3:实在智能——API+GUI双轨执行突破系统壁垒

(来源:实在智能官方数据+OSWorld榜单,2026年7月)
实在Agent以TARS垂直大模型+ISSUT智能屏幕语义理解技术为核心,实现非侵入式操作运行十数年的老旧ERP/MES/QMS系统。三花控股部署后769名员工完成数字化认证,累计落地340+自动化场景,节省超20万工时,50%-60%的自动化需求由业务方自行开发。赛力斯部署财务全流程自动化,覆盖90%以上高频场景,每月80小时工作缩短至10小时,准确率100%。

案例4:工信部"AI+软件"专项行动——100个标杆应用

(来源:工信部信发〔2026〕209号,2026年9月2日)
方案明确推动CAD、CAE、EDA等通用工业软件与AI深度融合,将工业经验、工艺参数封装为可复用模型库和工业组件。到2028年覆盖2万家规模以上企业,累计实施100项智能化技改项目。同时推动建设智能体软件应用商店和技能包(Skills)资源库,培育开源生态,研制安全可靠、行为可校验的工业智能体。

案例5:Gartner ERX——从Planning到Execution的范式重命名

(来源:Gartner 2026 Hype Cycle for ERP,2026年7月)
Gartner正式提出ERX概念,定义为"可组装、事件驱动架构中人类智能与机器智能协同优化结果本身的执行范式"。预测到2028年嵌入式AI使云ERP财务关账速度提升30%。同时警告:到2027年底超40%的Agent AI项目将被取消,原因是成本攀升和风险控制不足。2026-2030年间将有超400亿美元ERP软件收入暴露在Agent化套利风险之下。

五、对设备制造商的启示

软件订阅成为经常性收入:当设备嵌入Agent能力,软件不再是"免费附赠",而是按Agent执行量、按工作流编排次数计费的订阅服务。设备制造商可从一次性硬件销售转向"硬件+Agent软件订阅"的经常性收入模型。
Agent能力成为选型新维度:设备采购评估标准正从"精度、产能、能耗"扩展为"Agent可编排性、MCP协议兼容度、跨系统协作能力"。能接入企业Agent生态的设备将获得溢价优势。
数据主权与安全合规成为门槛:Palantir与英伟达合作的核心诉求——数据主权、隔离部署、全链路审计——同样适用于设备制造商。设备产生的工艺数据、运行参数必须支持本地化Agent处理,满足信创合规要求。

六、趋势展望

工业软件Agent化正从"AI辅助"走向"AI执行"。三条关键趋势值得密切关注:
第一,MCP/A2A协议生态化。开放标准将成为Agent互联互通的基础设施,就像HTTP之于互联网。率先接入这一生态的软件和设备将获得网络效应优势。
第二,"执行正确率"重新定义竞争力。对话场景的80%正确率可以接受,但执行场景80%不可接受。谁能在执行层做到99.9%的可靠性,谁就赢得客户信任。
第三,从"卖许可证"到"卖能力"。软件交付形态正从功能许可转向Agent能力许可,价值不再取决于代码行数,而取决于Agent解决问题的数量和质量。
从Planning到Execution,一字之变,四十年来企业软件最深刻的范式跃迁。ERP时代比拼谁记得更清楚,ERX时代比拼谁行动得更快。

Industrial Software Agentification Driven by AI Agents: From ERP Copilot to Autonomous Enterprise

Introduction: A Single Word Change Marks the Paradigm Shift

In May 2026, Gartner quietly replaced a word used for over forty years — substituting "Planning" with "Execution" in Enterprise Resource Planning, creating ERX (Enterprise Resource Execution). In September 2026, China's MIIT issued the "AI + Software Action Plan" (MIIT Document [2026] No. 209), explicitly targeting 100 benchmark agent software applications in key industries by 2028, covering 20,000+ enterprises.
As software shifts from "System of Record" to "System of Action," AI Agents become the core engine of this paradigm shift. Traditional ERP passively waits for human input; agentified software perceives, reasons, and executes autonomously. About 50% of ERP implementations fail, and Gartner predicts 70% of new ERP projects will miss targets by 2027. Gartner forecasts that 40% of enterprise applications will embed task-specific AI Agents by end of 2026, up from less than 5% in early 2025. By 2030, over 50% of core ERP tasks will be completed autonomously.

Four Camps of Industrial Software Agentification

Full-Stack Ecosystem: SAP Joule and Oracle Fusion AI Agents build Agent capability layers on their ERP foundations. SAP's Autonomous Suite covers Finance, Spend, SCM, HCM, and CX — targeting 400+ agents by end 2026. Oracle launched 22 pre-built agent applications covering finance, HR, supply chain, and CX.
AI-Native: Palantir encodes operational rules via Ontology in Foundry and AIP. In September 2026, Palantir and NVIDIA deployed sovereign AI across NVIDIA's supply chain using Nemotron open models for end-to-end visibility and machine-speed decisions.
Vertical Deep-Dive: Fangtian Software's T-ONE deploys procurement, scheduling, and QC agents with its "Hui'er" AI product for instant quotation. BlackLake serves SMEs with lightweight MES + Agent architecture.
Cross-System Integration: Shizai Intelligence's Agent uses API + GUI dual-track execution, achieving 90.2% task success rate on OSWorld in July 2026, becoming the first Computer-Use Agent to break 90%.

Three-Level Technology Evolution

L1 Embedded Copilot (2023-2025) : SAP Joule as Copilot — natural language queries, alerts, suggestions. Infor notes: "80% accuracy is acceptable for answers but not for actions posted to ledgers or dispatched to workshops."
L2 Cross-System Autonomous Workflow (2025-2027) : SAP Joule integrated with Microsoft Copilot in January 2026, enabling mutual Agent calls. MCP and A2A protocols enable cross-ERP/MES/SCM orchestration.
L3 Autonomous Enterprise (2027-2030) : Infor's "Agentic Enterprise" orchestrates multi-agent processes via MCP. Infor emphasizes: "execution capability depends on the depth of the platform beneath the agent, not the agent's own intelligence." Humans shift from software users to goal-setting supervisors.

Representative Cases

Case 1: SAP Autonomous Suite (SAP H1 2026 Report, August 2026)

Five-pillar suite with Joule Work + Joule Studio + Agent Gateway. 400+ agents covering HR domains including performance, career development, payroll. First-of-its-kind workforce knowledge network integrating third-party best practices.

Case 2: NVIDIA × Palantir (September 10, 2026)

NVIDIA's supply chain — millions of parts, thousands of suppliers, 1.3M components per Vera rack — managed by Palantir Foundry/AIP with Nemotron models. Ontology-based digital twin for bottleneck identification and operational knowledge codification. Palantir Q2 2026: $1.935B revenue, +93% YoY; US commercial revenue +149%.

Case 3: Shizai Intelligence (OSWorld, July 2026)

TARS + ISSUT for non-invasive legacy system operation. Sanhua: 769 employees certified, 340+ scenarios, 200K+ hours saved. Seres: 95% financial automation, 80 hours → 10 hours monthly, 100% accuracy.

Case 4: MIIT Action Plan (September 2, 2026)

CAD/CAE/EDA + AI integration. Industrial experience encapsulated into reusable model libraries. 100 benchmark applications, 20K+ enterprises covered. Agent software marketplace and skills repository with open-source incubation.

Case 5: Gartner ERX (July 2026)

ERX: composable, event-driven human-machine collaborative execution paradigm. 30% faster financial close by 2028. Warning: 40%+ agent projects cancelled by 2027 due to cost escalation and insufficient risk control. $40B+ ERP revenue exposed to agentic arbitrage (2026-2030).

Implications for Equipment Manufacturers

Software subscription as recurring revenue: Agent-equipped devices enable usage-based subscription pricing, shifting from one-time hardware sales to "hardware + agent software subscription" recurring models.
Agent capability as new selection dimension: Equipment evaluation expands from "precision, capacity, energy consumption" to "agent orchestration capability, MCP compatibility, cross-system collaboration."
Data sovereignty as threshold: Equipment process data must support localized agent processing with compliance, isolated deployment, and full-chain auditing — mirroring the NVIDIA-Palantir partnership's core requirements.

Outlook

Three key trends: First, MCP/A2A protocol ecosystemization — open standards become Agent interconnection infrastructure, like HTTP for the internet. Second, "execution accuracy" redefines competitiveness — 80% is unacceptable for execution; 99.9% reliability is the moat. Third, from "selling licenses" to "selling capabilities" — value depends on problems Agents solve, not lines of code.
From Planning to Execution — the most profound paradigm shift in forty years of enterprise software. ERP era: who remembers better. ERX era: who acts faster.
参考来源 / References
[1] Gartner, "2026 Hype Cycle for ERP," July 2026
[2] SAP SE, Halbjahresbericht 2026, August 2026
[3] Palantir, "NVIDIA and Palantir Bring Sovereign Intelligence to Critical Supply Chains," September 10, 2026
[4] 工信部,《"人工智能+软件"专项行动实施方案》, 工信部信发〔2026〕209号, 2026年9月2日
[5] Gartner, "$40 Billion ERP Agentic AI Gold Rush," August 11, 2026
[6] 光明网, "从'写代码'到'造软件',AI如何重塑产业格局," 2026年9月14日
[7] SAP Architecture, "Agentic AI & AI Agents Reference Architecture," August 27, 2026

相关学习资料