企业在经历盲目推动人工智能（AI）普及的狂热期后，正从单纯关注算力消耗等“投入端指标”转向严谨评估AI投资回报率（ROI）。初期管理层常追踪员工的Token消耗量或设立使用榜单，但此类投入指标无法衡量实际价值创造，且AI带来的零散时间节约往往悄然转化为员工的个人闲暇而非企业效益。因此，企业迫切需要转向以成果为导向的衡量体系，例如团队生产力、客户满意度或产品交付质量。

然而，基于成果的评估机制需要兼顾产出数量与质量，以防出现效率扭曲。萨里大学等机构的研究显示，基于某项健康与营养数据集发表的论文在2014至2021年间平均每年仅有4篇，但在2024年前九个月内激增至190篇，表明AI虽大幅拉升产出数量，但也导致低质研究泛滥。为此，谷歌等科技企业在考核工程师时综合考量速度、系统易用性以及软件质量三大维度，避免以牺牲产品标准为代价片面追求交付速度。

此外，多伦多大学等学者的研究证实，AI在制造业等传统企业的应用呈现典型的“J曲线效应”，即在员工学习新技术和管理流程重构的初期，企业生产力往往先经历短暂下滑而后才会回升。在衡量财务回报时，相较于直接裁员打击士气，更具建设性的做法是核算AI在遏制员工规模无序扩张中所节省的用工成本。鉴于AI技术演进的高度不确定性，企业在追踪投入与成果指标的同时，还应引入涵盖变革管理满意度与构建非自动化核心专业能力在内的“组织端指标”，在追求即期财务回报与长期组织学习之间维持平衡。

After an initial phase of uncritical enthusiasm focused on maximising token consumption and usage leaderboards, corporate leaders are shifting from input-driven metrics toward evaluating the actual return on investment (ROI) from artificial intelligence. While monitoring inputs like token volume may ensure engagement across engineering and finance departments, such measures fail to demonstrate whether real economic value is being created. Moreover, scattered productivity gains frequently accrue as uncaptured individual leisure rather than measurable organizational output, necessitating a transition toward outcome-based frameworks such as team efficiency, customer satisfaction, and product quality.

Implementing outcome-based evaluation requires careful calibration to prevent negative distortions where quantity eclipses rigor. Academic research led by the University of Surrey highlights this risk: papers derived from a specific health and nutrition dataset averaged four annually between 2014 and 2021, but surged to 190 in the first nine months of 2024 alone, reflecting a proliferation of AI-generated work of questionable merit. To avoid shipping subpar products faster, tech firms like Google Cloud evaluate performance across three distinct dimensions: speed of delivery, ease of operational friction, and uncompromising software quality.

Assessing returns is further complicated by the "J-curve" phenomenon identified by University of Toronto researchers, whereby productivity initially dips in established manufacturing firms due to implementation bottlenecks and management realignments before yielding net gains. Rather than achieving immediate financial returns through aggressive workforce reductions that damage morale, companies benefit more by measuring cost savings achieved through avoided headcount expansion. Given persistent technological uncertainty, enterprises must complement input and outcome metrics with organization-based measures that track change management effectiveness and internal expertise retention, balancing financial discipline with ongoing organizational adaptation.

Source: How to measure returns on AI

Subtitle: From tokenmaxxing to something more normal

Dateline: 8月 20, 2026 03:30 上午