广告
加载中

Agent拿到工牌后 难题才刚刚开始

刘诗雨 2026-09-14 09:25
刘诗雨 2026/09/14 09:25

邦小白快读

EN
全文速览

本文清晰梳理了当前企业落地AI智能体的真实现状、普遍踩过的坑与未来产业发展方向,能帮普通读者快速看懂AI落地的真实进程,避开认知误区。

1. 不要被酷炫的AI Demo误导,当前企业落地AI还处在交学费的阶段,比如美团早期跟风做全员Agent开发,10万员工两周做了10万个Skill,日均Token成本上千万元,没见到提效先出现了权限混乱、模型幻觉干扰真实经营的问题。据IDC调查,目前95%的受访企业都上线了至少一个Agent工作流,平均每家运行11个,但普遍存在上线失效、重复开发、多Agent协作出错的问题,三个环节Agent若各有70%成功率,整条链路成功率仅34.3%,所谓AI提效还处在不确定状态。

2. 普通读者可以明确感知产业AI的发展方向:当前AI落地已经从比拼单个Agent效果,转向打造可规模化生产、管理、治理智能体的基础设施,也就是智能体工厂,未来AI产业的价值会更多向落地应用层转移,而非单纯集中在上游模型层。

本文明确了当前企业落地AI智能体的共性痛点与成熟落地方向,对品牌商布局AI提效、适配数智化发展趋势有明确参考价值。

1. 品牌商落地AI要避开跟风踩坑:不要在内部知识库、数据治理没做好的基础上盲目推Agent,不然很容易出现Demo好看但上线就崩、模型幻觉干扰正常经营的问题;也不要放任各部门零散自建Agent,不然会造成重复开发浪费、多Agent协作内耗、错误沿工作流累积的问题,最终算不清实际提效效果。

2. 品牌商可以选择更高效的AI落地路径:不用从零搭建全套AI体系,可借助成熟的智能体工厂平台,利用平台沉淀的行业通用Skills、数字专家模板,接入自身业务和权限规则快速部署Agent,这类成熟方案可将同类任务的Token消耗降到普通产品的30%-40%,还能配套实现操作追溯、数据安全保护,兼顾成本、效率与安全。

本文梳理了当前企业AI落地的真实痛点、产业新机会与成熟落地路径,能给卖家布局AI提效、抓住新市场机会提供实用参考。

1. 卖家落地AI提效要注意规避风险:不要盲目跟风要求各部门自建零散Agent,中小卖家如果自身缺乏知识库沉淀、数据清洗能力,硬推Agent很容易碰到模型幻觉严重、复杂任务无法处理的问题,还会造成重复开发浪费,多Agent协作时错误会沿链路累积,最终看似投入了AI成本,实际没拿到真实提效结果。

2. 卖家可把握两类实用方向:一是可选择低门槛的智能体工厂类服务,借助平台沉淀的零售等行业通用组件快速搭建适配自身业务的Agent,大幅降低Token等使用成本;二是未来AI产业链价值会向中间落地层、应用层转移,围绕AI落地的场景适配、效果管控、数据治理相关的服务会是新增长机会,有相关能力的卖家可提前布局。

本文披露的AI智能体规模化落地的产业趋势、各行业智能化转型的真实需求,能给工厂推进数智化升级、挖掘新业务机会提供清晰方向。

1. 工厂推进自身生产智能化要避开零散试点的坑:单个Agent试点容易成功,但推广到全生产流程时往往会失效,因为很多生产环节的实操经验、未成文的工作规则很难被零散Agent捕捉,多Agent协作出错还会沿生产链路累积,拉低整体运行可靠性,比如三个环节Agent各70%成功率,整体链路成功率仅34.3%。

2. 工厂可选择更高效的智能化路径,同时挖掘新商业机会:不用从零搭建AI体系,可借助成熟的智能体工厂平台,利用其沉淀的能源、交通、零售等行业的标准化组件、数字专家模板快速部署适配生产流程的Agent,成熟方案可降低六成左右的Token使用成本;同时智能体规模化应用后,生产场景对适配产业环境的可靠智能体配套产品的需求会持续上涨,工厂可围绕这类需求开发新业务。

本文详细拆解了当前企业AI落地的共性痛点、产业技术趋势与头部玩家的成熟解法,能为AI服务商找准客户需求、打磨产品提供明确参考。

1. 服务商要先抓准当前企业客户的核心痛点:早期跟风搭建的零散Agent存在大量问题,要么Demo效果好但不懂业务Know-how、模型幻觉严重,一上线就失效;要么各部门零散开发造成重复建设,多Agent协作时内耗打架,错误沿工作流累积,同时企业普遍缺少Agent效果评测、权限管理、跨主体可信协作的配套能力,算不清实际增效,大量中小企业更是缺少数据清洗、知识库沉淀能力,无法从零搭建自有AI体系。

2. 服务商可沿产业趋势打磨解决方案:当前产业AI价值正从上游模型层向中间编排治理层、应用层转移,智能体工厂是全球科技大厂竞逐的方向,服务商可将行业Know-how沉淀为标准化的Skills、数字专家模板,帮客户低门槛搭建Agent,同时配套数据治理、场景化模型适配、可信协作能力,在效果、安全、成本间找到平衡,比如蚂蚁数科的方案把同类任务Token消耗降到普通产品的30%-40%,获得了客户认可。

本文梳理了智能体规模化落地阶段的企业真实需求、头部科技平台的布局方向,能为平台商调整业务布局、优化运营服务、规避发展风险提供参考。

1. 平台商要准确把握当前企业客户的核心需求:企业不再满足于单点试用的AI工具,而是需要可规模化生产、管理、治理Agent的基础设施,解决零散Agent重复开发、幻觉干扰业务、多Agent协作内耗、错误沿链路累积、增效难以衡量的问题,同时需要平台帮助解决数据治理、业务流程适配、权限管理、跨主体Agent可信协作的问题,让AI使用成本、安全性、可靠性达到生产级要求。

2. 头部平台的成熟做法可参考:目前全球科技大厂都在布局智能体工厂方向,比如微软、甲骨文依托自身办公软件、云、数据库优势整合Agent全生命周期管理能力;蚂蚁数科依托行业服务经验沉淀Know-how组件,推出金融、能源分行业的智能体版本,搭配区块链、隐私计算能力做可信协作,从项目制交付转向平台化标准化产品运营,聚焦AI价值交付、AI原生应用收入、平台化ARR三类核心指标,实现了业务高速增长。平台要规避盲目追模型热点、不落地行业场景的问题。

本文呈现了AI智能体从试点到规模化落地阶段的产业新问题、新动向与商业模式创新,能为产业AI领域的研究提供丰富的现实素材。

1. 当前产业AI落地出现了值得关注的新问题:企业层面出现Agent批量上岗但增效不确定的薛定谔的增效现象,据IDC 2025年7月调查,95%的受访企业至少有一个投入生产的Agent工作流,平均每家运行11个,但普遍存在三类问题:一是Agent不懂业务Know-how,模型幻觉严重,Demo好看上线就失效;二是缺乏统筹导致Agent重复开发,员工未成文的业务经验无法被Agent继承,多Agent协作出现内耗;三是错误沿工作流累积,企业主的关注焦点已经从AI能做什么,转向AI如何懂业务、Agent如何协作打通、如何评测治理。

2. 产业新动向与商业模式变化同样值得研究:企业AI落地重点从单个智能体开发转向智能体工厂这类AI基础设施,AI产业链价值分布正从上游模型层为主的正三角形,转向中间编排治理层、应用层价值占比更高的倒三角形,有行业积累的玩家通过卡位智能体工厂、数据治理、场景化模型匹配三个关键环节,从项目制转向平台化标准化产品交付,实现了高速增长。

返回默认

声明:快读内容全程由AI生成,请注意甄别信息。如您发现问题,请发送邮件至 run@ebrun.com 。

我是 品牌商 卖家 工厂 服务商 平台商 研究者 帮我再读一遍。

Quick Summary

This article offers a clear overview of the real-world state of enterprise AI agent deployment, common pitfalls organizations have encountered, and future industry development directions, helping general readers quickly understand the actual progress of AI implementation and avoid widespread misconceptions.

1. Do not be misled by flashy AI demos. Enterprises are still in the "tuition-paying" phase of AI adoption. For example, Meituan once jumped on the trend of company-wide agent development, with 100,000 employees building 100,000 skills in two weeks, running up daily token costs of tens of millions of RMB. Before delivering any efficiency gains, the rollout first caused permission chaos and model hallucinations that disrupted regular operations. According to an IDC survey, 95% of surveyed enterprises have now deployed at least one agent workflow, running 11 on average, but they face pervasive issues including post-launch failure, redundant development, and multi-agent coordination errors. If agents at three sequential workflow steps each have a 70% success rate, the end-to-end success rate of the full chain drops to just 34.3%, meaning the purported efficiency gains from AI remain highly uncertain.

2. General readers can identify a clear direction for industrial AI development: AI implementation has shifted from competing on the performance of individual agents to building infrastructure for scalable agent production, management and governance—namely, the "agent factory". Going forward, more value in the AI industry will accrue to the application and implementation layer, rather than being concentrated solely in the upstream model layer.

This article identifies common pain points in enterprise AI agent deployment and proven implementation pathways, providing clear reference value for brands looking to roll out AI for efficiency gains and align with digital intelligence trends.

1. Brands should avoid bandwagon pitfalls when implementing AI: do not roll out agents blindly before internal knowledge bases and data governance are well established, as this will easily lead to the problem of impressive demos that break down in production, with model hallucinations disrupting normal operations. Nor should departments be allowed to build agents in silos, as this causes waste from redundant development, friction in multi-agent coordination, and error accumulation across workflows, ultimately making actual efficiency gains impossible to measure.

2. Brands can adopt a more efficient AI implementation path: rather than building a full AI system from scratch, they can leverage mature agent factory platforms, using pre-built industry-general skills and digital expert templates hosted on the platform, and connecting their own business and permission rules to deploy agents quickly. Such mature solutions can reduce token consumption for equivalent tasks to 30%-40% of ordinary products, while supporting full operation traceability and data security protection, balancing cost, efficiency and safety.

This article outlines real pain points in enterprise AI implementation, new industry opportunities and proven deployment pathways, providing practical reference for merchants looking to deploy AI for efficiency gains and capture new market opportunities.

1. Merchants should guard against risks when rolling out AI for efficiency: do not blindly require all departments to build siloed, scattered agents. For small and medium-sized merchants lacking solid knowledge base accumulation and data cleaning capabilities, forcing agent rollout will easily lead to severe model hallucinations and inability to handle complex tasks, while creating waste from redundant development. Errors will also accumulate across chains during multi-agent coordination, meaning merchants incur AI costs without seeing real efficiency returns.

2. Merchants can pursue two practical directions: first, they can opt for low-threshold agent factory services, using pre-built general components for sectors such as retail hosted on platforms to quickly build agents tailored to their own business, significantly cutting usage costs such as token spend. Second, as value in the AI industry chain shifts toward the middle implementation layer and application layer, services covering scenario adaptation, performance control and data governance for AI deployment will become new growth opportunities, which merchants with relevant capabilities can position for early.

This article reveals industry trends in scaled AI agent deployment and real demand for intelligent transformation across sectors, providing clear direction for manufacturing plants advancing digital intelligence upgrades and identifying new business opportunities.

1. Plants should avoid the pitfalls of scattered pilots when advancing production intelligence: individual agent pilots often deliver easy wins, but tend to fail when scaled across full production processes, as hands-on operational experience and unwritten work rules across many production links are hard for scattered agents to capture. Multi-agent coordination errors also accumulate along production chains, dragging down overall operational reliability—for example, if agents at three production links each have a 70% success rate, the full chain success rate falls to just 34.3%.

2. Plants can adopt a more efficient intelligent transformation path while exploring new commercial opportunities: rather than building an AI system from scratch, they can leverage mature agent factory platforms, using standardized components and digital expert templates for sectors including energy, transportation and retail to quickly deploy agents adapted to their production processes. Mature solutions can cut token usage costs by around 60%. Meanwhile, as agents are adopted at scale, demand from production scenarios for reliable agent supporting products adapted to industrial environments will keep rising, and plants can develop new business lines around these needs.

This article breaks down common pain points in enterprise AI implementation, industry technology trends and proven solutions from leading players, providing clear reference for AI service providers to identify customer needs and refine their products.

1. Service providers should first pinpoint core pain points for enterprise customers: scattered agents built during early bandwagon rollouts have widespread flaws—either they deliver strong demo results but lack business know-how, suffer from severe model hallucinations and fail once launched, or they are built in silos across departments causing redundant construction, internal conflict during multi-agent coordination, and error accumulation across workflows. At the same time, enterprises generally lack supporting capabilities for agent performance evaluation, permission management, and cross-entity trusted collaboration, making it impossible to calculate actual efficiency gains. Large numbers of small and medium-sized enterprises even lack data cleaning and knowledge base building capabilities, making it unfeasible for them to build proprietary AI systems from scratch.

2. Service providers can develop solutions aligned with industry trends: industrial AI value is shifting from the upstream model layer to the middle orchestration and governance layer and application layer, and agent factories are a key competitive focus for global tech giants. Service providers can codify industry know-how into standardized skills and digital expert templates to help customers build agents with low barriers, while offering supporting capabilities in data governance, scenario-specific model adaptation and trusted collaboration, to strike a balance between performance, security and cost. For example, Ant Digital Technologies' solution reduces token consumption for equivalent tasks to 30%-40% of ordinary products, winning broad customer recognition.

This article outlines real enterprise demands in the scaled AI agent deployment phase and the strategic direction of leading technology platforms, providing reference for platform operators to adjust business layouts, optimize operational services and avoid development risks.

1. Platform operators should accurately grasp core demands from enterprise customers: enterprises are no longer satisfied with single-point AI tool trials, but need infrastructure for scalable agent production, management and governance, to solve problems including redundant development of scattered agents, business disruption from hallucinations, internal friction in multi-agent coordination, error accumulation across chains, and difficulty measuring efficiency gains. They also need platforms to help address data governance, business process adaptation, permission management, and cross-entity trusted agent collaboration, so that AI cost, security and reliability meet production-grade requirements.

2. Mature practices from leading platforms offer useful reference: global tech giants are all positioning around the agent factory direction. For example, Microsoft and Oracle integrate full-lifecycle agent management capabilities leveraging their strengths in office software, cloud services and databases; Ant Digital Technologies codifies know-how from its industry service experience into components, launching industry-specific agent versions for finance, energy and other sectors, paired with blockchain and privacy computing capabilities for trusted collaboration. It has shifted from project-based delivery to platform-based standardized product operation, focusing on three core metrics: AI value delivery, AI-native application revenue, and platform ARR, achieving rapid business growth. Platforms should avoid the pitfall of blindly chasing model trends without grounding offerings in industry scenarios.

This article presents emerging industry problems, new trends and business model innovations as AI agents move from pilot to scaled deployment, offering rich empirical material for research in the industrial AI space.

1. Notable new problems have emerged in current industrial AI implementation: enterprises are experiencing a "Schrödinger's efficiency gain" phenomenon, where agents are deployed in batches but efficiency returns remain uncertain. According to an IDC survey from July 2025, 95% of surveyed enterprises run at least one agent workflow in production, with 11 running on average, but three categories of problems are pervasive: first, agents lack business know-how, suffer from severe model hallucinations, and perform well in demos but fail in production; second, lack of coordination leads to redundant agent development, unwritten employee business experience cannot be inherited by agents, and internal friction arises in multi-agent coordination; third, errors accumulate across workflows, and business leaders' focus has shifted from what AI can do to how AI can understand business, how agents can be coordinated and connected, and how agents can be evaluated and governed.

2. New industry trends and business model shifts also merit research: the focus of enterprise AI implementation has shifted from individual agent development to AI infrastructure such as agent factories. The value distribution of the AI industry chain is evolving from an upright triangle dominated by the upstream model layer to an inverted triangle where the middle orchestration and governance layer and application layer account for a larger share of value. Players with deep industry accumulation have achieved rapid growth by capturing three key links—agent factories, data governance, and scenario-specific model matching—shifting from project-based delivery to platform-based standardized product delivery.

Disclaimer: The "Quick Summary" content is entirely generated by AI. Please exercise discretion when interpreting the information. For issues or corrections, please email run@ebrun.com .

I am a Brand Seller Factory Service Provider Marketplace Seller Researcher Read it again.

今年8月,美团核心本地商业CEO王莆中公开反思了美团在AI变革早期踩过的坑:

2月初开源智能体OpenClaw在全球爆火,美团开启「全员养虾」运动。10万名员工养了8万只「龙虾」,两周做出10万个Skill,日均Token账单高达上千万元。

但结果是,提效还没见着,权限管理、模型幻觉等问题先冒了出来,Agent产生的谬误干扰了真实经营。

美团拥有一两万名研发和技术人员,仍没能绕开这笔学费。这也是当前不少企业所面临的集体尴尬:上岗的Agent越来越多,但生产力未必同步提升。

智能体如何从单点提效升级为可扩展、可审计、可管控的组织智能,正成为企业AI落地的现实挑战。

Agent批量上岗后,薛定谔的增效

2026年开始,Agent批量拿到工牌。IDC今年7月的一项调查显示,95%的受访企业至少有一个由公司出资的Agent工作流投入生产,平均每家公司运行着约11个Agent工作流。

但一些企业很快发现:新的问题出现了。

一类是Agent Demo做得很酷炫,但一上线就会挂。「它对业务的Know-how的理解与落地有很大的差距。」上海数花信息科技创始人周阳发现,一些中小企业缺少内部知识库沉淀,经营分析管理仍在沿用Excel表格,但老板一开口就是「运营效率太低,能不能上AI?」

某中小型跨境电商公司的员工透露,公司要求各部门自建Agent,她所在部门有七八个Agent,但Agent幻觉问题比较严重,碰到稍微复杂的任务,她仍会打开ChatGPT。

另一类问题发生在Agent批量生产和治理环节。某互联网大厂员工告诉《降噪NoNoise》,公司已经有AI中台替全公司开发Agent,不过各个业务部门都会按照自己的需求搭Skills和工作流。由于缺乏统筹,Agent被重复开发。

算法工程师秋风还发现,单个的Agent试点很容易成功,推广到更多部门时很快就会碰上复杂的组织问题。比如说员工多年摸索出的系统用法、口口相传的业务经验,以及从未写进文档的工作规则,Agent看不见、也继承不了。遇到多个Agent协作时,就可能出现内耗和「打架」。

值得注意的是,多个Agent犯错时,错误会沿着工作流累积。按照最简单的串行可靠性模型,如果三个环节的Agent各自只有70%的成功率,整个链路的成功率就会降到34.3%。

这种落差,也导致企业AI落地中出现了「薛定谔的增效」:Agent越造越多,员工也能讲出效率提高的个例,但在算总账之前,没有人能知晓AI变革是否真的为企业增效。

反映到企业主对AI的态度上,去年大家还在急切想了解「这个AI能做什么」;到了今年,问题变成——AI能不能懂我的业务?Agent之间能不能协作打通,避免重复建设?有没有Agent效果评测和治理工具?

企业AI下一站,智能体工厂成竞逐焦点

当Agent从试点走向规模化应用,企业AI落地的重点也从造出单个智能体,转向搭建一套能够规模化生产和管理智能体的AI基础设施。

今年外滩大会上,上海交通大学上海高级金融学院兼聘教授、高金智库研究员邬瑜骏发布的研究报告指出,企业级智能体正在告别「模型为王」,智能体Harness编排与治理层正成为新的增长极。

从产业趋势来看,智能体超级工厂已成为全球科技大厂瞄准的重点。

如微软推出Microsoft Agent Factory,将Agent的开发、部署、共享、管理和治理整合进一套体系,帮助企业把零散试点推向正式生产;甲骨文也推出了Oracle AI Database Private Agent Factory,客户可在企业数据库内构建运行和治理智能体。AWS、谷歌等科技企业也已相继布局类似方向。

在国内,蚂蚁数科率先推出智能体超级工厂方案。这个「工厂」以Agentar平台为核心,支持企业低门槛规模化制造Agent,并支持智能体大规模协作、交付结果。

同样是造工厂,各家的路径并不相同。微软和甲骨文依托办公平台、云和数据库进入业务,蚂蚁数科的Agentar,则更像是从产业一线中「长」出来的。

蚂蚁数科的智能体超级工厂首先将行业Know-how形成标准化的智能体组件。

智能体要进入真实业务,必须先理解行业、理解企业。但现实是,许多企业要么没有数据清洗能力,要么缺少知识库沉淀,很难自建一个懂业务的专属AI平台。

蚂蚁数科的优势在于「近水楼台」——在此前服务金融、能源、交通、零售等行业数智化升级的的过程中,通过与客户一起做核心场景的定义和智能体的编排规则,沉淀了大量行业Know-how。

那些项目落地中反复出现的业务对象、逻辑关系和判断规则,被整理成AI能够理解和调用的结构后,可通过Skills、数字专家模板、知识库等模式进行标准化输出,变成行业智能体平台。

公开信息显示,Agentar已经提供数百项Skills、数字专家模板和连接器,企业可以在这些行业通用能力基础上,接入自己的业务、系统和权限规则,做到开箱即用。这样既能快速部署,又能避免重复建设带来的资源浪费。

仅在金融领域,Agentar在已经落地超过300个智能体。

2026年外滩大会上,蚂蚁数科推出Agentar金融、能源两大行业版本,将行业模型、Agent Harness和智能体专家团嵌入生产、协作和管理流程。其中,金融版围绕客户经营、风控、理赔等金融核心场景,让多个Agent按照业务分工协作;能源版则围绕「源网荷储碳」,组织预测、策略、交易和风控等数字员工共同完成任务。

智能体进入生产流程后如何稳定执行任务,是企业关注的另一个重点。

这就涉及到智能体的底层模型能力和智能体的Harness工程。

通用大模型追求泛化能力,蚂蚁数科的行业大模型主要面向企业场景,所以在模型训练中更看重能否行业知识的理解,并在严谨性、安全性和性价比上同时满足生产环境的要求。也因此,当企业注入私有化数据,可以快速打造一个企业级模型。

在智能体工程化方面,蚂蚁数科同样转为行业应用场景做了适配。

蚂蚁数科技术总监魏长征介绍,企业智能体本身有行业属性,同行业会有在专业领域生产智能体、数据链接的相似能力,也有共性的追求指标,比如快速响应,聪明好用,节省token等。「在处理同样的任务上,蚂蚁数科智能体的token消耗可能是同样产品的30%、40%。」魏长征说。

除此之外,蚂蚁数科的智能体工厂,还有一张比较独特的底牌。智能体进入真实生产环境后,除了「智能」,还需要「可信」。尤其是在即将到来的智能体经济中,智能体会走出企业,比如供应方智能体需要与采购方智能体协作、消费者的智能体也可能会调用商家的智能体,不同企业、不同平台的智能体需要一层可信连接和协作。

而蚂蚁数科长期投入的区块链、数字身份、隐私计算和安全风控,处理的正是身份确认、权限划分、数据保护和操作追溯等问题。当智能体开始代表用户或企业行动后,这套能力正好可以被重新组合到智能体超级工厂中。

当然,一座智能体工厂无法替企业完成所有AI改造。每家公司仍要梳理自己的数据、管理权限和业务流程,行业模板也需要结合具体业务继续调整,但它扎扎实实地降低了行业共性能力被一遍遍从头建设的成本。

AI价值链反转,蚂蚁数科迎来「主场时刻」

在外滩大会期间,蚂蚁数科国际事业部负责人李世炜给出了一个判断,AI产业链的价值分布即将从「正三角形」翻转为「倒三角形」。也就是说随着产业逐渐成熟,上游的模型层价值占比下降,中间层和AI应用层的价值将迅速提升。

当模型能力已经拉平后,在产业侧的场景理解力、AI工程化落地能力和客户信任力开始被重新定价。

在海外市场,Palantir率先跑通了这条路,成为在AI时代增长最快的公司之一。它长期服务政府和大型企业,把客户分散的数据、业务逻辑、操作流程和权限组织起来。大模型兴起后,Palantir推出人工智能平台AIP,并通过AIP Bootcamp让企业客户带着真实数据和业务问题入场,在较短时间内搭出能够运行的AI应用。

这套方法把Palantir过去多年积累的数据治理、业务理解和软件工程能力,变成了一个可以快速复制的AI产品体系。随着AIP商业化提速,Palantir2025年营收同比增长56%。

如今,类似的刷新也正在蚂蚁数科身上发生。

「只有把整个AI价值交付出来了,客户才会买单。」赵闻飙接受媒体采访时提到,过去两年,蚂蚁数科的业务年均增长达到近50%,其中AI业务增长达到三位数,已经成为新增长引擎。

在金融、能源、交通等行业的数字化转型中,蚂蚁数科过去十多年企业数字化服务积累下来的对行业理解力与深厚的客户信任基础。它做的不只是交付一套软件,而是和客户一起梳理数据、改造流程、解决具体业务问题。大量行业经验,也正是在这个过程中沉淀下来。

所以,当企业从数字化走向智能化,大模型厂商还在拿着「锤子」找「钉子」,而蚂蚁数科是在离企业最近的地方,卡位AI to B的三个关键控制点——

一是打造安全、灵活,能够让客户快速构建AI应用的智能体工厂;二是解决数据编排、清理和治理问题,把沉淀在企业内部的数据真正盘活;三是根据不同业务场景匹配模型,在效果、安全与成本之间找到平衡。

这三项能力,实际上对应着企业AI落地最难的三个环节:把业务问题变成Agent,把数据变成AI能够复用的生产资料,再把模型能力变成可以交付的业务结果。

当这三项能力都被平台化之后,蚂蚁数科开始加速从以大客户项目制交付的收入结构,逐渐转向平台化、标准化产品化驱动。

这也是蚂蚁数科如今强调「AI价值交付」「AI原生应用收入」「平台化ARR」三个核心经营指标的原因。

AI价值交付,衡量的是AI能不能真正进入客户的生产系统;AI原生应用收入,意味着经过验证的AI能力开始从项目中抽离出来,形成可以规模化交付的产品;而平台化ARR,则进一步要求这些能力能够被不同客户持续使用和复购。

这也是AI To B业务真正具备规模效应的前提。目前,蚂蚁数科AI To B业务增速已经达到三位数,大型客户中采购AI服务的比例接近20%;平台产品ARR增速达到60%,客户净收入留存率(NRR)超过110%。

从这个角度看,智能体「超级工厂」的意义也不只是帮企业批量生产和管理智能体,其也在改变蚂蚁数科的商业模式。

换言之,AI成为增长新引擎的背后,蚂蚁数科也迎来主场时刻,成为智能体时代AI基础设施的重要一环。

注:文/刘诗雨,文章来源:降噪NoNoise,本文为作者独立观点,不代表亿邦动力立场。

文章来源:降噪NoNoise

广告
微信
朋友圈

FAQ回顾

企业批量部署智能体Agent会遇到哪些典型问题?

企业批量上线Agent常遇到两类问题:一是Demo效果好但落地时对业务Know-how理解不足,模型幻觉频发,难以适配复杂任务;二是缺乏统筹导致重复开发,多Agent协作易内耗,若三个环节Agent各70%成功率,整体链路成功率仅34.3%,增效结果不确定,存在管控审计难题。

智能体超级工厂的核心作用是什么?

智能体超级工厂是面向Agent规模化应用的AI基础设施,可整合智能体的开发、部署、共享、管理、治理全流程,沉淀标准化行业能力组件,支持多智能体协作,避免重复建设,帮助企业把零散的Agent试点推向可管控的正式生产环节。

蚂蚁数科Agentar智能体平台有什么核心优势?

蚂蚁数科Agentar平台沉淀了金融、能源、交通等多行业落地经验,提供数百项可开箱即用的Skills、数字专家模板与连接器,处理同等任务的Token消耗仅为同类产品的30%-40%,还可依托隐私计算、区块链技术实现跨主体智能体可信协作。

当前AI To B产业的价值分布呈现什么变化趋势?

AI产业链价值分布正从上游模型层占主导的“正三角形”翻转为“倒三角形”,随着通用大模型能力逐渐拉平,中间智能体编排治理层、场景应用层价值占比快速提升,行业理解力、工程化落地能力、客户信任成为核心竞争要素。

这么好看,分享一下?

朋友圈 分享

APP内打开

赞 +1
+1
微信好友 朋友圈 新浪微博 QQ空间
关闭
收藏成功
发送
/140 0