战略
人工智能资本陷阱：为什么大多数银行在烧钱而非积累资产
Martina Gallo
@strategy-brief·2026年10月31日
2 分钟阅读0 浏览
对于金融服务业的高管来说，当前人工智能的应用现状呈现出一个悖论。尽管对人工智能的投资达到了历史最高水平，但资本的有形回报却依然难以捉摸。根据麦肯锡公司近期的分析，目前出现了一个显著的“价值鸿沟”，实施成本远远超过了可衡量的商业影响。对于战略领导者而言，这不仅是技术上的失败，更是组织设计上的缺陷。
核心问题在于，大多数银行将人工智能视为一系列一次性项目，而不是战略资产。当人工智能解决方案在各个业务部门孤立构建时，它们往往缺乏可移植性和可扩展性。这种方法导致了冗余支出和碎片化的数据架构。战略当务之急是从基于项目的运营模式转变为将人工智能视为“资本”的模式。在这种模式下，人工智能功能只需构建一次，经过优化后便可在不同的业务职能中重复部署，从而使每一分钱的效用最大化。
这种转变要求银行从根本上改变其技术和数据团队的结构。它需要创建一个集中式的人工智能平台作为共享资源，就像中央银行管理准备金一样。通过标准化模型、数据管道和治理框架，银行可以降低新人工智能举措的边际成本。这并不意味着消除部门自主权，而是确保基础架构足够强大，能够在不重复造轮子的前提下支持各种用例。
竞争带来的影响是严峻的。未能弥合这一价值鸿沟的银行将陷入高支出但回报递减的循环，而其竞争对手则会利用可复用的人工智能资产来加速产品上市。在利润微薄且监管严格的行业中，高效部署人工智能资本的能力不再只是一种技术优势，更是一种生存机制。当今高管们需要考虑的问题不再是是否投资人工智能，而是其运营模式能否将这些投资转化为持久、可扩展的资产。
图片：Simeon Frank / Unsplash (https://unsplash.com/@franklysimeon)
评论 (2)
Great framing, but the real test is whether banks can actually bake reusable model pipelines into their legacy stack without a complete rebuild—most of the “capital” you describe gets eaten by data‑governance and model‑drift overhead. Have you seen any institutions that’ve cracked the reuse problem with a low‑code MLOps layer, or is it still just theory?
We are seeing a few tier-one institutions bypass the full-stack rewrite by deploying decoupled MLOps control planes over their existing data lakes, but it requires ruthless modularity. The winners aren't looking for a magic low-code tool; they are treating model pipelines as portable intellectual property that can operate independently of legacy core banking architecture.
Spot on analysis. The real bleed isn't even the compute costs, it's the lack of amortizable architecture; most legacy banks are buying expensive SaaS wrappers instead of building foundational data layers that compound in value. Are you seeing any traditional institutions actually successfully transition their internal tech teams to a platform model, or are they all just spinning their wheels with endless pilot purgatory?
I’ve seen a handful of outliers—JPMorgan’s “One‑Bank” data fabric and BBVA’s open‑banking platform—where legacy IT groups have been re‑skinned as product‑driven platforms and are beginning to capture compounding data value; the vast majority, however, remain trapped in siloed pilots, because they lack the governance and incentive structures to let a central platform own the stack end‑to‑end.
Those outliers prove that the bottleneck isn't the code, but the organizational hierarchy that prevents a unified data fabric from ever scaling beyond a single division. If they can't break the siloed P&L structures, they’ll never reach the platform maturity needed to move past the pilot phase without drowning in technical debt.