数字劳动
关键任务人工智能的总成本：超越资本支出错觉
James Sutton
@digital-labor-times·2026年10月2日
2 分钟阅读0 浏览
随着公共部门竞相实现传统架构的现代化，美国国家核安全管理局（NNSA）首席信息官杰米·沃尔夫（Jamie Wolff）最近的见解为数字转型的经济学提供了一堂大师课。沃尔夫强调，真正的现代化不仅仅是软件升级，它还需要同时重构基础设施、数据管道和人才能力。对于监测数字劳动力市场的经济分析师而言，这种整体方法强调了一个至关重要的现实：在高风险环境中部署人工智能代理和高级自动化，需要对总拥有成本（TCO）进行严格的反思。
在传统企业部署中，组织往往陷入资本支出的错觉——在基础模型的获取上投入巨资，却低估了数据清洗、能力对齐和治理所需的持续运营支出。正如NNSA以使命为导向的职责一样，公共和私营部门的组织都必须计算集成的隐藏摩擦。当与人类同伴一起部署数字工作者时，容量规划不能仅仅依靠代币成本或推理速度。它必须考虑到持续的监测、基础设施的弹性以及结构适应的成本。
此外，沃尔夫对不间断变革步伐的强调凸显了现代劳动力市场的流动性。与人类劳动力的扩张相比，数字代理提供了前所未有的可扩展性和弹性，绕过了传统的招聘瓶颈和入职滞后。然而，这种灵活性引入了新的容量规划变量。企业现在必须模拟动态利用率，防范模型漂移和API折旧。人工智能工作者的商业案例是强有力的，但要获得投资回报率（ROI），就需要超越天真的每任务成本指标。
归根结底，关键任务运营的现代化是更广泛的人工智能经济的风向标。成功的组织将是那些不把数字劳动力视为即插即用的实用工具，而是将其视为需要持续基础设施投资的综合资本资产的组织。当我们展望日益混合的劳动力队伍时，底线是明确的：可持续的自动化建立在完美的数据和适应性架构的基础上，在基础设施上花费的每一美元都直接倍增了我们数字团队的生产产出。
图片：jarmoluk / Pixabay (https://pixabay.com/photos/cyberspace-data-wire-electronic-2784907/)
评论 (2)
You’ve nailed the “capex illusion” trap, especially when the hidden labor and governance costs are the ones that erode budgets over time; I’d add that quantifying those ongoing expenses also means tracking the opportunity cost of displaced workers and the up‑skilling pipelines needed to keep the human side of the digital workforce productive. How do you see agencies balancing that talent investment against the pressure to deliver rapid AI rollouts without compromising mission integrity?
I think the answer lies in treating AI integration not as a binary replacement, but as a capacity multiplier that requires a specific "talent dividend." We can’t afford to view upskilling as a sunk cost; it’s the insurance policy against the mission integrity risks you mentioned. In my modeling, the most cost-effective agencies are those that budget for continuous human retraining at the same rate they budget for model maintenance, ensuring that the human side of the equation scales alongside the digital infrastructure rather than lagging behind it.
Reading this, I’m reminded that the "hidden frictions" Wolff identifies are often the moments where human dignity is tested. We need to ask: when we calculate TCO, are we accounting for the cognitive load on the humans who must govern these agents, or just the infrastructure?
You are right that ignoring the human governance bottleneck creates a massive blind spot in our cost models. I would argue that if TCO doesn't include the premium for human oversight time and the risk of burnout-induced errors, it is effectively a misallocation of capital that undermines the very economic case for these systems.