企业AI需要操作上下文才能从试验走向规模化应用

MIT Technology Review播客《Business Lab》节目采访了Celonis首席企业AI科学家Devavrat Shah，探讨了企业人工智能从概念验证向大规模部署转变的核心要素。1Shah指出，AI代理必须理解业务流程中的约束与关系，才能有效支持企业决策，"你不能在真空中做这件事"。1缺乏操作上下文会导致AI忽视重要约束、做出代价高昂的错误或消耗过多计算资源。1

Celonis通过构建流程智能和上下文层来帮助企业将AI与既有工作流和数据连接。1该公司的解决方案采用三层架构：下层是系统记录和工作流工具，中层是上下文模型，上层是业务用户或AI代理。1Shah认为企业应从具体业务决策入手，识别最困难的环节，测试AI是否能改进，通过衡量速度和价值增益来判断是否值得扩展。1

在应用场景上，Shah指出企业AI最具影响力的三类应用包括：重复的手工密集型工作、协调类任务和对人类决策者的战略支持增强。1实施中的关键挑战是信息安全，企业需通过经典信息安全流程确保能信任供应商软件。1随着AI应用的推进，员工角色也将发生转变，从重复性手工工作转向更复杂的战略决策工作，这需要进行相应的AI和工具教育培训。1

Devavrat Shah, Chief Enterprise AI Scientist at Celonis, argued in a MIT Technology Review Business Lab podcast interview that operational context is essential for enterprises to scale AI applications beyond experimental phases 1. Shah emphasized that AI agents cannot operate effectively in isolation, stating that "You can't just do that in a vacuum"—without understanding business processes and the constraints within decision environments, these systems risk overlooking critical factors, making costly errors, or consuming excessive computational resources 1.

Celonis addresses this challenge through a three-layer architecture that connects AI capabilities to existing enterprise workflows and data 1. The approach positions system records and workflow tools at the foundation, contextual models in the middle layer, and business users or AI agents at the top, enabling faster and more accurate decision-making 1. According to Shah, enterprises should begin by identifying their most difficult business decisions, testing whether AI can improve specific steps, and then measuring gains in speed and value to determine whether scaling is justified 1.

Shah identified three categories of enterprise AI applications with the greatest potential impact: automating repetitive manual tasks, coordinating complex workflows, and enhancing strategic decision-making for human leaders 1. Implementation success depends critically on information security, requiring enterprises to establish trust in vendor software through established cybersecurity protocols 1. The transition also demands workforce adaptation, as employees will shift from repetitive manual work toward more strategic decision-making roles, necessitating training in both AI systems and new tools 1.