广告
加载中

AI业务增速三位数 蚂蚁数科如何应对智能体规模化难题

亿邦动力 2026-09-11 13:20
亿邦动力 2026/09/11 13:20

邦小白快读

EN
全文速览

这篇是蚂蚁数科独立运营两年来首次系统性披露业务进展,核心围绕智能体经济的产业判断、布局路径与落地挑战,包含不少大众能看懂、用得上的实用信息。

核心经营数据要点

1. 蚂蚁数科过去两年整体业务年均增长近50%,AI To B业务增速达三位数,平台产品ARR增速超60%,客户净收入留存率超110%,连续6年位居中国BaaS市场第一,份额达32.4%。

可感知的实用信息

1. 普通中小商家可通过接入支付宝AI开放平台的Agentar服务,几分钟将原有小程序升级为智能体,低门槛用上AI服务。

2. 当前大模型落地产业分工具提效、智能体承担岗位职责、多智能体组织协同三个同步推进阶段,大家日常接触的智能客服、营销类智能服务已属于第二阶段应用。

3. 针对大模型“胡说八道”的幻觉问题,行业已形成模型校验、规则管控、工具兜底、确定性知识存储的系统防控方案,高合规场景会通过多重校验减少错误。

蚂蚁数科披露的智能体产业落地进展,可为品牌商布局AI能力、把握企业服务趋势提供务实参考。

智能体应用的阶段趋势

1. 当前大模型在产业落地同步经历员工工具提效、智能体承接岗位职能、多智能体组织协同三个阶段,品牌可按自身数字化节奏逐步布局,无需盲目追求全套复杂系统。

2. 智能体规模化应用的核心卡点是信任问题,未来智能体代品牌执行任务、参与交易时,必须明确身份授权、责任归属,KYA可信身份体系会成为这类应用的基础配置。

品牌可复用的落地路径

1. 可借助Agentar平台的数百项现成技能、数字专家模板,快速搭建客服、营销、风控场景的智能体,标准化场景搭建周期可缩短至小时级。

2. 中小品牌可通过支付宝AI开放平台的能力,几分钟将原有服务小程序智能体化,低成本升级用户服务体验,且正规服务商不会回流品牌敏感生产数据,仅沉淀通用业务经验,数据安全有保障。

3. 采购AI服务时可优先选择按实际业务ROI分润的合作模式,降低投入风险。

蚂蚁数科披露的智能体产业布局,为卖家释放了明确的AI工具普惠信号与新赛道机会。

可直接复用的低门槛提效工具信息

1. 中小卖家可依托支付宝AI开放平台接入的Agentar服务,一键将原有店铺小程序升级为智能体,几分钟就能生成支持AI标准协议的服务入口,无需复杂技术开发,就能为消费者提供智能咨询、服务引导类体验。

2. 平台已提供数百项现成技能模板和数字专家模板,标准化场景的智能服务配置最快小时级即可完成,大幅降低卖家使用AI的技术与时间成本。

新机会与风险提示

1. 智能体经济当前处于发展早期,智能体身份管理、可信交易相关配套服务存在大量空白,全球范围内相关标准正在制定中,提前适配规则的卖家可抢占新的流量红利。

2. 采购AI服务时要注意数据安全边界,正规服务商不会回流卖家的生产经营数据,只会沉淀通用业务规则类经验,要警惕要求上传核心经营数据的不合规服务。

3. 未来智能体代用户交易的场景会逐步普及,卖家要提前适配智能体支付、身份核验相关规则。

蚂蚁数科披露的智能体产业落地路径,可为工厂推进数字化转型、捕捉B端智能服务需求提供明确方向。

工厂数字化转型的可参考路径

1. 工厂可按照大模型落地三阶段节奏逐步推进AI应用,先从工具层面为员工提效,再逐步在客服、营销、风控、研发等岗位落地智能体,最终实现多智能体协同,把单点任务经验沉淀为工厂可复用的组织能力,无需一开始就投入重金搭建全套复杂系统。

2. 可借助专业服务商的行业大模型能力,将工厂私有化的生产数据融入模型,打造适配自身生产流程的智能中枢,不需要从零开始自研通用大模型,还可选择按实际业务价值分润的合作模式,降低转型投入风险。

可捕捉的新商业机会

1. 智能体规模化生产催生大量新型基础设施需求,包括适配智能体的新型协议、可信身份配件、高合规场景的幻觉防控软硬件,有相关技术储备的工厂可切入这类配套供应链。

2. 能源等垂直行业已推出专门的智能体版本,围绕生产调度、碳管理等场景的数字员工需求正在释放,工厂可结合自身业务对接相关服务,也可参与垂直场景的需求共建。

蚂蚁数科在智能体落地过程中总结的技术路径与客户痛点,可为To B AI服务商提供非常具象的实践参考。

智能体赛道的核心发展趋势

1. 当前企业客户对智能体的需求已经从零散工具需求,转向可规模化生产、管理、治理智能体的基础设施需求,智能体参与交易、执行任务时的信任问题是规模化落地的核心卡点,AI与区块链结合构建的KYA可信身份体系会成为行业必备基础。

2. 全球范围内智能体身份、支付相关的标准竞争已经展开,国内相关国标、行业自律公约正在推进中,提前适配标准的服务商将抢占生态入口位置。

客户痛点与解决方案参考

1. 当前客户有三个核心痛点:一是有自研工具的冲动,实际更关注研发范式的体系化升级而非单纯工具采购;二是金融、能源类客户对数据边界极度敏感,坚决不接受生产数据回流,仅认可规则、经验类的沉淀共享;三是高合规场景对大模型幻觉零容忍,需要模型校验、规则硬门禁、工具兜底、确定性知识存储的系统方案解决,配套风险防控产品需求明确。

2. 商业化层面,深度陪伴客户挖掘业务落点、按实际价值分润的重服务模式,虽投入较重,但客户信任感更强,最终成单概率更高。

蚂蚁数科在智能体生态上的布局动作,为各类平台布局AI生态、完善平台治理提供了可借鉴的实践样本。

平台搭建智能体生态的可参考做法

1. 要为生态内商家提供低门槛的智能体升级通道,比如蚂蚁将Agentar平台接入支付宝AI开放平台,支持千万级商家一键将原有小程序智能体化,几分钟生成符合AI标准协议的智能体,既降低商家AI接入成本,也能快速沉淀平台AI服务生态。

2. 生态建设不能只提供零散工具,要搭建类似智能体超级工厂的基础设施,支持生态参与者批量生产、管理、协同各类智能体,还要提前搭建可信身份、交易存证类的信任底座,解决智能体代客执行任务时的身份核验、责任界定问题。

平台运营与风险规避要点

1. 要严格划清数据边界,对生态内商家、用户的敏感数据做到不回流,仅沉淀通用业务规则、服务经验,避免数据合规风险。

2. 针对大模型幻觉带来的服务风险,要搭建系统级防控体系,通过硬规则门禁、关键任务引导反思、高准确率任务调用专业脚本、确定性知识单独存储等方式减少智能体出错概率,可配套风险防守类产品为生态参与者兜底。

3. 要积极参与智能体身份、支付相关的标准制定,抢占生态规则话语权。

蚂蚁数科披露的智能体落地实践、行业判断与现实挑战,为AI产业、数字经济相关研究提供了鲜活的一线产业样本。

智能体产业的最新发展动向

1. 大模型在产业落地目前同步处于三个并行阶段:员工工具化提效、智能体承接岗位职责、多智能体组织级协同,产业终局远未到来,企业需求已经从单点智能体开发,转向规模化生产、治理、协同智能体的基础设施需求。据蚂蚁披露,其AI To B业务已实现三位数增速,客户净收入留存率超110%,连续6年位居国内BaaS市场第一,To B AI商业化已经验证了正向增长路径。

2. 智能体经济的核心闭环是AI加区块链的组合,AI解决智能自动化问题,区块链解决信任自动化问题,KYA体系是智能体规模化参与交易的必备基础,目前全球已展开相关标准竞争,国内已出台智能体支付自律公约,身份管理相关国标正在推进。

产业新问题与商业模式探索

1. 当前智能体规模化落地存在三个共性难题:客户存在工具自研冲动,核心诉求实为研发范式升级;敏感行业客户的数据边界诉求强烈,仅支持经验回流不接受生产数据回流;高合规场景下的大模型幻觉兜底需要系统工程解决。

2. 商业化层面,To B AI服务正在探索深度陪伴客户、按实际业务价值分润的交付模式,产业链价值分配正处于翻转早期,未来应用层将涌现大量机会。

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 is Ant Digital Technologies’ first systematic disclosure of its business progress since operating independently two years ago. It centers on the company’s judgment of the agent economy’s industrial trajectory, layout roadmap and implementation challenges, with plenty of accessible, practical takeaways for the general public.

Key operating metrics

1. Ant Digital Technologies has posted nearly 50% annualized overall business growth over the past two years, with triple-digit growth for its AI to-business segment, over 60% growth in annual recurring revenue (ARR) for its platform products, and a net dollar retention rate of over 110%. It has ranked first in China’s Blockchain-as-a-Service (BaaS) market for six consecutive years, holding a 32.4% market share.

Practical, consumer-relevant insights

1. Small and medium-sized merchants can upgrade their existing mini programs into AI agents in minutes by accessing the Agentar service on Alipay’s open AI platform, adopting AI capabilities at a very low threshold.

2. Industrial large language model (LLM) implementation is progressing across three parallel stages: tool-based productivity improvement, agents taking on job responsibilities, and multi-agent organizational collaboration. Common consumer-facing services such as intelligent customer service and AI marketing tools already fall into the second stage.

3. To address LLM hallucinations (factual errors generated by models), the industry has developed a systematic mitigation framework covering model verification, rule-based access control, tool-based fallback, and deterministic knowledge storage; high-compliance scenarios deploy multi-layer verification to minimize errors.

Ant Digital Technologies’ disclosure of industrial AI agent implementation progress offers practical reference for brands building AI capabilities and tracking enterprise service trends.

Stage-based trends for agent applications

1. Industrial LLM deployment is advancing across three parallel stages: employee productivity tools, agents taking on defined job functions, and multi-agent cross-organizational collaboration. Brands can roll out AI capabilities in line with their own digitalization roadmap, rather than rushing to deploy overly complex full-stack systems prematurely.

2. The core bottleneck for scaled agent application is trust. When agents act on behalf of brands to execute tasks and participate in transactions, clear identity authorization and accountability frameworks are required. The Know Your Agent (KYA) trusted identity system will become standard infrastructure for such applications.

Replicable implementation pathways for brands

1. Brands can leverage Agentar’s hundreds of pre-built skills and digital expert templates to rapidly deploy agents for customer service, marketing and risk control scenarios, cutting deployment timelines for standard use cases to just hours.

2. Small and medium-sized brands can upgrade their existing service mini programs into AI agents in minutes via Alipay’s open AI platform, improving customer experience at low cost. Qualified service providers will not extract brands’ sensitive production data, only retaining generalized business insights to ensure data security.

3. When procuring AI services, brands can prioritize cooperation models based on revenue sharing tied to actual business ROI to reduce upfront investment risk.

Ant Digital Technologies’ disclosed layout for the AI agent industry sends a clear signal of accessible AI tools for sellers, while highlighting new track opportunities.

Readily accessible, low-threshold efficiency tools

1. Small and medium-sized sellers can integrate the Agentar service via Alipay’s open AI platform to upgrade their existing store mini programs into AI agents with one click, generating service entries compatible with standard AI protocols in minutes. This eliminates the need for complex technical development while enabling intelligent customer consultation and service guidance for consumers.

2. The platform provides hundreds of pre-built skill templates and digital expert templates, allowing configuration of intelligent services for standard scenarios in as little as an hour, drastically reducing the technical and time costs for sellers to adopt AI.

New opportunities and risk reminders

1. The agent economy is in its early development stage, with major gaps in supporting services for agent identity management and trusted transactions. Relevant global standards are currently being formulated, and sellers that adapt to these rules early stand to capture new traffic dividends.

2. When procuring AI services, pay close attention to data security boundaries. Legitimate service providers will not extract sellers’ operational production data, only retaining generalized business rule insights; be wary of non-compliant services that require uploading core operational data.

3. Scenarios where agents complete transactions on behalf of users will gradually become widespread, so sellers should adapt in advance to rules related to agent payment and identity verification.

Ant Digital Technologies’ disclosed pathway for industrial AI agent implementation provides clear direction for factories advancing digital transformation and capturing B-end intelligent service demand.

Reference pathways for factory digital transformation

1. Factories can roll out AI applications in line with the three stages of LLM implementation: start with tool-based productivity gains for frontline employees, then deploy agents for roles in customer service, marketing, risk control and R&D, and finally achieve multi-agent collaboration to convert isolated task experience into reusable organizational capabilities. There is no need to make heavy upfront investments in full-stack complex systems at the initial stage.

2. Factories can leverage industry-specific LLM capabilities from professional service providers to integrate their private production data into models, building intelligent hubs tailored to their own production processes without developing general-purpose LLMs from scratch. They can also opt for cooperation models based on revenue sharing tied to actual business value to reduce transformation investment risk.

New commercial opportunities to capture

1. Scaled agent production will create massive demand for new infrastructure, including agent-compatible protocols, trusted identity components, and hallucination mitigation software and hardware for high-compliance scenarios. Factories with relevant technical reserves can enter this supporting supply chain.

2. Vertical sectors such as energy have already launched dedicated agent versions, with growing demand for digital employees covering production scheduling, carbon management and other scenarios. Factories can connect relevant services based on their own operations, or participate in co-building demand for vertical use cases.

Ant Digital Technologies’ summary of technical pathways and customer pain points from AI agent implementation offers concrete practical reference for to-B AI service providers.

Core development trends in the agent track

1. Enterprise customer demand for agents has shifted from fragmented tool needs to requirements for infrastructure that supports scaled production, management and governance of agents. The core bottleneck for large-scale implementation is trust when agents participate in transactions and execute tasks. The KYA trusted identity system, built by combining AI and blockchain, will become essential industry infrastructure.

2. Global competition over standards for agent identity and payment is already underway, with relevant national standards and industry self-regulation conventions being developed in China. Service providers that adapt to these standards early will secure entry positions in the ecosystem.

Customer pain points and solution reference

1. Customers currently face three core pain points: first, many have an impulse to develop tools in-house, but their actual priority is systematic upgrading of R&D paradigms rather than pure tool procurement; second, clients in finance, energy and other sensitive sectors have extremely strict requirements around data boundaries, refusing to allow production data to be extracted and only accepting the sharing of generalized rules and experience; third, high-compliance scenarios have zero tolerance for LLM hallucinations, requiring a systematic solution covering model verification, hard rule access controls, tool-based fallback and deterministic knowledge storage, with clear demand for supporting risk mitigation products.

2. On the commercialization side, the high-touch service model of providing deep, ongoing support to identify high-value business use cases and sharing revenue based on actual delivered value, while requiring heavier upfront investment, builds stronger customer trust and leads to higher deal closure rates.

Ant Digital Technologies’ layout moves in the AI agent ecosystem provide an actionable reference sample for all types of platforms building AI ecosystems and improving platform governance.

Reference practices for building an agent ecosystem

1. Platforms should provide low-threshold agent upgrade channels for merchants within their ecosystem. For example, Ant integrated the Agentar platform into Alipay’s open AI platform, supporting tens of millions of merchants to upgrade their existing mini programs into agents with one click, generating agents compliant with AI standard protocols in minutes. This not only lowers merchants’ AI access costs, but also rapidly builds out the platform’s AI service ecosystem.

2. Ecosystem building cannot rely solely on fragmented tools. Platforms need to build infrastructure comparable to an “agent super factory” that supports ecosystem participants to produce, manage and coordinate various agents at scale. They also need to pre-build trust foundations covering trusted identity and transaction evidence storage, to resolve identity verification and accountability issues when agents execute tasks on behalf of users.

Key points for platform operation and risk mitigation

1. Strictly define data boundaries, avoiding extraction of sensitive data from ecosystem merchants and users, and only retaining generalized business rules and service experience to prevent data compliance risks.

2. To address service risks caused by LLM hallucinations, build a system-level mitigation framework that reduces agent error rates via hard rule access controls, reflection prompts for high-stakes tasks, professional script invocation for high-accuracy tasks, and separate storage of deterministic knowledge. Supporting risk protection products can also be offered to cover ecosystem participants.

3. Actively participate in the formulation of standards related to agent identity and payment to secure discourse power over ecosystem rules.

Ant Digital Technologies’ disclosed AI agent implementation practices, industry judgments and real-world challenges provide a vivid frontline industrial sample for research on the AI industry and digital economy.

Latest development trends in the agent industry

1. Industrial LLM implementation is currently progressing across three parallel stages: tool-based employee productivity improvement, agents taking on job responsibilities, and multi-agent organizational-level collaboration. The final industrial form is far from settled, and enterprise demand has shifted from single-point agent development to infrastructure that supports large-scale production, governance and coordination of agents. According to Ant’s disclosure, its to-B AI business has achieved triple-digit growth, with a net dollar retention rate of over 110% and a first-place ranking in China’s BaaS market for six consecutive years, verifying a positive growth path for to-B AI commercialization.

2. The core closed loop of the agent economy is the combination of AI and blockchain: AI solves intelligent automation, while blockchain solves trust automation. The KYA system is an essential foundation for agents to participate in transactions at scale. Global competition over relevant standards is already underway; China has released an industry self-regulation convention for agent payment, while national standards for identity management are in development.

Emerging industry challenges and business model exploration

1. There are currently three common bottlenecks for scaled agent implementation: customers often have an impulse to develop tools in-house, but their core demand is actually upgrading R&D paradigms; clients in sensitive sectors have strong demands around data boundaries, only permitting experience sharing and refusing production data extraction; hallucination mitigation in high-compliance scenarios requires systematic engineering solutions.

2. On the commercialization side, to-B AI services are exploring delivery models featuring deep ongoing customer support and revenue sharing tied to actual business value. The value distribution across the industrial chain is in the early stages of an inversion, with massive opportunities set to emerge at the application layer.

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.

【亿邦原创】9月10日,外滩大会现场,蚂蚁数科管理层集体露面。这是这家公司独立运营两年后,第一次系统性地向媒体披露业务进展。

蚂蚁数科CEO赵闻飙带来了几组数字:过去两年整体业务年均增长近50%,AI To B业务增速达到三位数,平台产品ARR增速超过60%,客户净收入留存率超过110%。

但比数字更值得关注的,是蚂蚁数科对Agent经济的判断,以及它试图在这个尚未定型的市场中占据的位置。

一、Agent经济:从工具到组织,从技术到信任

赵闻飙把大模型在产业中的落地分成三个阶段。

第一阶段,企业把大模型当工具用,提升员工效率;第二阶段,智能体进入业务,承担客服、营销、风控、研发等岗位职责;第三阶段,智能体进入组织——当企业拥有的智能体从几十个变成几百个,问题就不再是“怎么开发一个智能体”,而是“怎么让它们协同工作,把一次任务的经验变成组织可复用的能力”。

“这三个阶段目前仍在同时发生,”赵闻飙说,“大模型落地产业的终局还远远没有到来。”

这个判断决定了蚂蚁数科的战略方向。今年7月,他们在世界人工智能大会上发布智能体超级工厂,核心逻辑是:企业需要的不是零散的智能体工具,而是一套能够规模化生产、管理、治理智能体的基础设施。赵闻飙用工业革命做类比,“不是用一台机器代替所有机器,而是把不同的机器组成一条生产线。”

但超级工厂要解决的远不止生产效率问题,赵闻飙反复强调一个判断:“AI负责创造智能,区块链负责创造信任。”也就是说,如果智能体只是企业内部提效的工具,信任问题没那么突出;但一旦智能体开始代表企业或个人执行任务、调用系统、参与交易,就必须回答“它是谁、代表谁、被授权做什么、出了问题谁负责”。

这正是KYA(Know Your Agent)成为焦点的原因。

外滩大会期间,蚂蚁数科联合十余家单位推动智能体身份管理相关国家标准立项,蚂蚁集团同步发布智能体可信身份解决方案及KYA产品。蚂蚁数科技术总监魏长征解释,区块链本质是信任机制的自动化,AI是智能的自动化,两者结合才能让智能体经济完成闭环。“智能体经济要大规模发展,必须有信任基础设施。”

关于智能体经济的标准竞争已经在全球范围内展开。海外市场上,ERC-8004、Visa TAP、Mastercard Agent Pay、Google UCP等不同方案都在争夺智能体身份和支付的标准制定权,中国支付清算协会近期发布《智能体支付应用自律公约》,提出KYA机制;国家标准GB/Z 185.3-2026《人工智能 智能体互联 第3部分:身份管理》也已发布,蚂蚁科技集团是主要起草单位之一。

蚂蚁数科此时推动KYA国标立项并发布产品,意图很明确,在智能体经济的标准版图中占据基础设施的位置。赵闻飙表示:“KYA的生态位就是智能体经济的一个入口。”魏长征补充,虽然底层技术复用区块链能力,但智能体对区块链信息的需求与人不同,需要大量新型协议构建,这既是挑战也是机会。

二、蚂蚁数科的应对:三条技术线,一个工厂

蚂蚁数科的应对策略可以概括为三条技术线:智能体、企业级大模型、区块链+AI。

智能体层面,Agentar平台把模型、知识、Skill、业务流程组合起来,批量生产和管理数字员工。平台已提供数百项Skills和数字专家模板,在标准化场景中智能体交互周期可缩短至小时级。此次外滩大会上,Agentar推出金融、能源两大行业版本——金融版针对客户经营、风控等场景组织多个智能体协同工作;能源版围绕源网荷储碳等业务构建能碳数字员工。

Agentar还接入了支付宝AI开放平台,为千万级商家提供智能体服务。蚂蚁数科AI业务总裁余滨透露,中小商家可以一键式将过去的小程序智能体化,几分钟内生成支持AI标准协议的智能体。“很多中小企业对降低AI门槛有非常强烈的需求,效果比预期快很多。”

企业级大模型层面,蚂蚁数科今年初成立大模型技术创新部,推出行业大模型LingDT。赵闻飙明确表示,数科不做广义的问答式大模型,而是聚焦企业级模型,将企业私有化数据融入,打造企业智能中枢。“我们不简单做Post-Training,也做mid-train,并且有可能会做pre-train。”他透露,数科独立的开源模型将在几个月后发布。

蚂蚁数字科技算法技术部负责人章鹏进一步解释,行业模型与通用模型的争论第一天就有,但有些行业没有足够多的公开数据,通用模型难以覆盖。更重要的是,行业落地真正的评测不是公开榜单,而是in-house bench。“我们的核心竞争力不是权重,权重在不停更新。我们积累的是reward function、行业task、评价标准和评测方法。”

区块链+AI层面,蚂蚁数科从区块链起家,已有11年积累,区块链专利超过6000件。IDC报告显示,蚂蚁数科以32.4%的市场份额位居中国BaaS市场第一,连续第六年排名第一。这些能力正在被重新组合到智能体基础设施中——KYA体系、智能体身份管理、可信交易存证,都建立在区块链技术之上。

在商业化层面,蚂蚁数科成立了FDE(Forward Deployed Engineer)团队,陪伴客户从找AI价值落点到实现业务价值。赵闻飙强调深度陪伴和按价值交付:“许多单子我们讲业务层面,赢的概率很大。哪怕营销,我能够给你ROI是多少,在这个return以上我们可以分润。”这种模式虽然重,但客户信任感越来越强。

三、未解的难题:自研冲动、数据边界与幻觉兜底

战略清晰,但落地中的问题同样真实。媒体交流会上,几个提问直接指向了蚂蚁数科面临的核心挑战。

第一个挑战来自客户的自研冲动——随着Coding能力的提升,企业级Agent服务商的客户开始出现“自己搭”的想法。蚂蚁数科技术总监魏长征认为,工具研发只是起步,真正难的是研发范式的体系化升级。“我们接触到更多的客户,最关心的是如何让自己企业的整个研发范式能够升级,而不是纠结于是否一定要自研工具。”赵闻飙补充,数科有自己的AI Code产品AIWork,在几个榜单中排名靠前,“大家可以去试一下”。

还有数据边界的难题,蚂蚁数科服务了大量金融、能源客户,这些客户的数据极为敏感。余滨明确表示,数科不会做数据回流,“这是行业的行规性问题”。但哪些可以沉淀?余滨的答案是“经验回流”——通过本体论技术,把业务规则、特征、关系沉淀下来,而非客户的生产数据。“这不是数据的回流和共享,而是经验的回流和共享。”

当然,大模型的幻觉和可靠性也是智能体经济的关键影响因素。在金融、政务等高合规场景中,智能体一旦胡说八道,后果可能是灾难性的。章鹏介绍了蚂蚁数科在模型层面的努力:在后训练阶段加入引用、参考文献和思考过程double check的数据;与清华大学联合发布data Agent Bench,在数据处理和引用上做优化;在工具层用Harness做规则性管控。魏长征补充了四个手段:硬门禁、软引导、工具、记忆。“对于一些规范性的问题、授权的问题、数据边界的问题,一定要有Harness做规则性管控。软引导解决的是幻觉问题,在关键任务上做引导性反思。工具层面,报表类准确率要求高的内容不会让模型直接运算,而是调动脚本化工具。记忆系统则把确定性知识和行业背景以embedding方式存储。”

赵闻飙最后补了一句:“要防止这个问题,是个系统工程。”他顺势给自家产品打了广告——蚁天鉴,做防守;龙虾卫士,已获信通院认证。“假如模型内核、智能体内核还不能防止,我们还有这些产品可以提供给客户。”

海外市场是另一个变量。蚂蚁数科已与马来西亚Ryt Bank、Touch‘n Go、菲律宾GCash、哈萨克斯坦BCC银行、越南FPT等开展合作。蚂蚁数科国际业务总裁李世炜分析,AI产业链价值分配正处于从正三角形向倒三角形翻转的早期阶段,“目前最赚钱的是芯片和大模型公司,但未来应用层将出现大量机会。我们聚焦前台的AI To B深水区。”

“独立运营两年后,我们的方向比两年前更加聚焦。”赵闻飙在交流会结尾说。

从金融起步,走向能源、医疗、消费、制造,蚂蚁数科正在用“智能体超级工厂+区块链信任底座”的组合,试图在全球Agent经济的标准争夺中占据一席之地。

但这个市场才刚刚开始,标准尚未定型,客户需求仍在演化,竞争格局远未清晰。蚂蚁数科的赌局,也才刚刚开局。

本文首发于 亿邦动力 官方网站

文章来源:亿邦动力

广告
微信
朋友圈

FAQ回顾

大模型在产业场景落地会经历哪些发展阶段?

大模型在产业落地共分三个阶段:第一阶段作为工具提升员工个人效率;第二阶段作为智能体承担客服、营销、风控、研发等岗位职责;第三阶段实现多智能体组织协同,将任务经验沉淀为组织可复用能力,目前三个阶段仍在同步推进。

KYA机制是什么,对智能体经济有什么作用?

KYA即Know Your Agent(了解你的智能体),核心是明确智能体的身份归属、授权范围、责任主体,解决智能体代表主体执行任务、调用系统、参与交易时的信任问题,是智能体经济规模化发展必需的核心信任基础设施。

企业规模化部署智能体有什么可落地的解决方案?

企业可选择蚂蚁数科推出的智能体超级工厂方案:基于Agentar平台可批量生产、管理数字员工,目前已推出金融、能源行业版本,搭配KYA可信身份体系、区块链信任底座,可实现智能体从生产到治理的全生命周期规模化管控。

智能体落地时如何防控大模型幻觉带来的风险?

防控大模型幻觉需要系统性方案:模型层在后训练阶段加入引用校验、双重思考核查机制;工具层通过硬门禁管控高风险操作,调用专用脚本工具处理高准确率要求任务;搭配确定性知识记忆系统,辅以蚁天鉴、龙虾卫士等专项风控产品兜底。

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