AI News AI资讯 12h ago Updated 2h ago 更新于 2小时前 49

Dialogue with Ant Digital Technologies: Building a Super Factory for Commercial Agents, Co-building the Chinese Industry Version of Harness Standards 对话蚂蚁数科:打造商业智能体超级工厂,生态共建中国行业版Harness标准

Ant Group Digital Tech launched the "Commercial Agent Super Factory" at WAIC 2026, shifting focus from general model capabilities to industry-specific vertical models and engineering落地 (implementation). The platform, Agentar2.0, provides pre-built templates for nearly 200 job roles and hundreds of "Skills" tools, enabling enterprises to deploy digital experts without starting from scratch. Key differentiators include solving industry-specific challenges through data governance, engineering frame 蚂蚁数科发布“商业智能体超级工厂”及Agentar2.0平台,预置近200个岗位级数字专家模板,推动AI从通用问答向垂直行业任务解决转型。 提出打造符合中国行业特点的“行业版Harness”,通过意图识别、策划、执行、表达的“四车间”逻辑实现模型能力的工程化落地。 针对金融等高合规要求场景,建立以回答准确率(90%-95%+)和胜合率(85%+)为核心的价值交付与效果付费验证体系。 战略重心转向数据底层治理、多行业安全适配及评测体系完善,旨在通过标准化产品输出助力传统行业智能化升级。

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Ant Group Digital Tech launched the "Commercial Agent Super Factory" at WAIC 2026, shifting focus from general model capabilities to industry-specific vertical models and engineering落地 (implementation).
  • The platform, Agentar2.0, provides pre-built templates for nearly 200 job roles and hundreds of "Skills" tools, enabling enterprises to deploy digital experts without starting from scratch.
  • Key differentiators include solving industry-specific challenges through data governance, engineering frameworks (a "Chinese version of Harness"), and strict compliance/safety measures tailored for regulated sectors like finance.
  • Performance metrics in financial scenarios show answer accuracy of 90%-95% and an "agreement rate" with human professionals exceeding 85%, supporting a "value delivery, pay-for-effect" business model.
  • Future strategies involve expanding vertical models into healthcare, energy, and transportation, while standardizing engineering methodologies and evaluation systems across industries.

Why It Matters

This article highlights the critical industry transition from generic Large Language Model (LLM) competition to practical, value-driven AI agent deployment in enterprise settings. For AI practitioners and CTOs, it underscores that success now depends less on raw model parameters and more on engineering rigor, data governance, and domain-specific adaptation. The introduction of a standardized "Harness" framework for Chinese industries offers a replicable blueprint for integrating AI into complex, regulated business workflows, making it a vital reference for organizations planning their next phase of digital transformation.

Technical Details

  • Agentar2.0 Platform: Serves as the infrastructure for the Commercial Agent Super Factory, featuring nearly 200 pre-configured job-level digital expert templates and hundreds of subscription-based "Skills" tools for immediate deployment.
  • Vertical Domain Models: Utilizes industry-specific large models trained on combined external data and internal enterprise data, addressing the hallucination and lack of professional depth common in general-purpose models.
  • "Four-Workshop" Engineering Framework: A proprietary process for financial scenarios involving Intent Recognition, Planning, Execution, and Expression. This modular approach ensures precise handling of complex user needs (e.g., distinguishing between insurance queries and wealth management requests).
  • Evaluation Metrics: Employs rigorous testing including "Answer Accuracy" (90-95% in finance) and "Agreement Rate" (consistency with human experts >85%), validated through multiple rounds of testing (e.g., 28 rounds in a major city bank project).
  • Data Governance & Safety: Integrates tools for underlying data cleaning and alignment, ensuring compliance with regulatory standards (especially in finance) and adapting safety protocols for other sectors like manufacturing (production safety) and energy (prediction accuracy).

Industry Insight

  • Shift to "Value-Based" AI Procurement: Enterprises should move away from evaluating AI solely on technical benchmarks and adopt "pay-for-effect" models where ROI is tied to specific business outcomes like conversion rates or risk reduction.
  • Importance of Engineering Over Models: The "Chinese version of Harness" concept suggests that the competitive moat for AI vendors will be their ability to engineer reliable, scalable workflows rather than just training larger models. Companies must invest in process re-engineering to support AI integration.
  • Cross-Industry Standardization: As AI expands beyond finance into healthcare and energy, the need for standardized data governance and evaluation frameworks becomes paramount. Early adopters should prioritize partners who offer robust, industry-adapted engineering tools rather than generic API access.

TL;DR

  • 蚂蚁数科发布“商业智能体超级工厂”及Agentar2.0平台,预置近200个岗位级数字专家模板,推动AI从通用问答向垂直行业任务解决转型。
  • 提出打造符合中国行业特点的“行业版Harness”,通过意图识别、策划、执行、表达的“四车间”逻辑实现模型能力的工程化落地。
  • 针对金融等高合规要求场景,建立以回答准确率(90%-95%+)和胜合率(85%+)为核心的价值交付与效果付费验证体系。
  • 战略重心转向数据底层治理、多行业安全适配及评测体系完善,旨在通过标准化产品输出助力传统行业智能化升级。

为什么值得看

本文揭示了AI产业从单纯追求模型参数规模向注重业务场景价值和工程化落地的关键转折,为理解企业级AI应用提供了具体的方法论参考。蚂蚁数科提出的“行业版Harness”概念及“四车间”工作流,为解决大模型在复杂业务流程中的幻觉和控制难题提供了可借鉴的架构思路。

技术解析

  • Agentar2.0平台架构:作为“商业智能体超级工厂”的核心载体,该平台提供数百个可订阅的Skills级开箱即用工具,支持企业快速启用具备专业知识、业务流程及工具调用能力的数字专家,降低开发门槛。
  • 行业版Harness工程化方法:借鉴海外Harness概念并结合中国市场特点,在金融场景中确立了“四车间”逻辑(意图识别、策划、执行、表达),强调对用户真实意图的深度解析而非表面响应,以适配不同行业的颗粒度流程设计。
  • 垂直大模型与数据治理:强调行业垂直大模型的重要性,通过融合外部数据与企业内部业务数据进行治理,解决通用模型在专业度和业务适配上的不足,特别是在制造、医疗等数据基础复杂的行业。
  • 效果验证指标体系:建立了量化的价值交付标准,包括专业场景下的回答准确率(90%-95%以上)以及专业人员与智能体结果一致性的“胜合率”(85%以上),并通过多轮评测(如头部城商行28轮测试)确保模型可靠性。

行业启示

  • AI落地进入深水区:行业关注点已从模型能力竞争转向应用价值创造,企业需构建具备行业知识、业务理解和工程化能力的智能体系统,才能实现从个人效率提升到组织协同升级的转变。
  • 工程化能力成为核心壁垒:单纯拥有大模型已不足以应对复杂业务需求,如何将业务语言转化为技术能力并形成可复制的工程化解决方案(如“四车间”模式),将成为区分AI供应商竞争力的关键。
  • 生态共建与标准化输出:未来AI发展依赖于行业生态的开放与合作,通过API化组件和标准化方法论(如评测体系、行业版Harness),加速AI在金融、医疗、能源等多行业的渗透与价值验证。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent LLM 大模型 Finance AI 金融AI Deployment 部署