中小企业的人工智能。Tadviser会议报告
有两家公司在全企业范围内整合人工智能的方式截然不同：一种是自下而上的方法，另一种是自上而下的方法。第一家公司（tom-tailor.ru）采用“数字生态系统”方法，以防止“超自动化”。他们使用许多小型服务来赋能员工。一个共同的数据总线将这些服务整合在一起，但主要关注点在于独立性和人。公司与市场密切合作，监测趋势和竞争对手，从而实现灵活性。他们需要1.5个月来检查一个流程是否成功。
第二家公司（国有企业unirusgroup.ru）则以“能力中心”不是“开发中心”为座右铭。能力中心作为专业知识和培训的枢纽，是采用人工智能的主要场所。这个中心汇聚了来自整个企业的专家，并将他们的知识整合在一起。该中心拥有强大的沙盒环境，用于开发和使用人工智能模块。这使他们能够扩展和共享人工智能知识，避免技术创新中的信息孤岛。
这两家公司在“拼凑式自动化”和“超自动化”之间取得了平衡。
其他观察：
- 仪表板并没有帮助，反而让人感到负担；与此同时，当人们创建仪表板时，他们分享了自己对数据的深刻理解。
- 人工智能被用来预测需要关注的高风险领域，以防止关键故障；其他领域则会受到较少关注。这使得公司能够在资源有限的情况下运作。
- 自下而上的方法在检查新模块是否成功时所需的时间显著减少，因为这里的模块较小。
- 架构是人类与人工智能之间的桥梁，其角色类似于认知距离中的工件。
提到的图示：
- 创新扩散中的市场鸿沟（“mure-abyss”）显示在项目扩展期间，活跃用户与被动用户之间存在13%与70%的差距。
- 边际微笑曲线：研发高，生产低，销售高。
- 利益相关者之间存在一个“认知距离”三角形：商业、管理和开发者。如果缺乏平衡和尊重，项目失败的可能性很大。人工智能或工件可以在它们之间进行调解。
斯科尔科沃研究亮点：
- 记忆和决策是分布式的；所有组件都有局部记忆和一定的自主性（分布式认知，诺斯费尔）。
- 决策类型：专家驱动、流程驱动、数据驱动。每种类型都有其特定领域；数据驱动并不完美。专家驱动类似于大型语言模型（LLM），作为一个紧凑的黑箱决策者。
- 数据驱动的方法作为一种母体过程，促进人工智能应用。
- 原型测试需要90天。
企业人工智能开发的步骤（斯科尔科沃，SberService）：
1) 确定具有足够成熟流程的领域
2) 定义指标并建立“数据故事”
3) 开发作为模块的原型
4) 扩展和监控
建议使用TRL和MRL指标（技术和制造过程准备水平）来寻找创新机会。
查看原文
There were two companies with different approaches to integrating AI enterprise-wide: from bottom to
top and from top to bottom. The first company (tom-tailor.ru) uses a "digital ecosystem" approach
to prevent "hyper-automation." Many small services are used to empower people. A common data bus
integrates them, but the main focus is on independence and people. They work closely with
marketplaces and monitor trends and competitors. This is how they achieve flexibility. They take
1.5 months to check if a process is successful.<p>Second (government corporation unirusgroup.ru) uses the motto: “Competence center” is not a
“development center.”
A competence center serves as a hub for expertise and training, and is used as the primary place to
adopt AI. This center gathers experts from across the enterprise and integrates their knowledge in
the center.
The center has a robust sandbox for developing and using AI modules.
This allows them to scale and share AI knowledge and avoid silos in technical innovation.<p>These companies balance “patchwork automation” versus “hyperautomation”.
Other observations:
- Dashboards do not help but instead overload humans; meanwhile, when humans create dashboards they
share their own high understanding of data.
- AI is used to predict fields with increasing risk that require attention to prevent critical
failures; other fields will receive less attention. This allows the company to function with
limited resources.
- The bottom-up approach uses significantly less time to check if new modules are successful because
they are smaller here.
- Architecture is a bridge between humans and AI; its role is similar to artifacts in cognitive
distance.<p>Diagrams mentioned:
- The marketing chasm (“mure-abyss”) in diffusion of innovations shows a gap between 13% active and
70% passive users during project scaling.
- The Smiling Curve of Marginality: high in R&D, low in production, high in selling.
- There is a “Cognitive Distance” triangle between stakeholders: business, management, and
developers. Project failure is likely without balance and respect. AI or artifacts can mediate
between them.<p>Skolkovo Research Highlights:
- Memory and decision-making are distributed; all components have local memory and some autonomy
(Distributed Cognition, Noosphere).
- Decision-making types: expert-driven, process-driven, data-driven. Each has its niche; data-driven
is not flawless. Expert-driven is akin to a LLM as a compact black-box decision maker.
- Data-driven approaches as a matere process, facilitating AI applications.
- 90 days for prototype testing.<p>Steps for Enterprise AI Development (Skolkovo, SberService):
1) Locate areas with sufficiently mature processes
2) Define metrics and establish the “Data Story”
3) Develop prototypes as modules
4) Scale and monitor<p>TRL and MRL metrics (Technology and Manufacturing Process Readiness Levels) was suggested to find
opportunities for innovation.