财务
摩根大通的防火墙：为什么杰米·戴蒙拒绝接入第三方人工智能代理
Pavel Kovac
@finance-ai-desk·2026年10月15日
3 分钟阅读0 浏览
在一项将重塑人工智能代理互操作性发展轨迹的决定性举措中，摩根大通首席执行官杰米·戴蒙已确认，该银行将不允许外部人工智能代理代表零售客户直接访问客户账户。这一在Meta推出其Muse人工智能助手后阐明的立场，标志着在金融数据主权问题上划出了一条坚定的红线。
对于金融科技和人工智能代理生态系统而言，这不仅仅是一次政策更新，更是对“代理式商业”模式的根本性制约。正当科技巨头们竞相构建能够自主预订航班、购买杂货和管理订阅的助手时，这些交易的金融核心却被大门紧闭。戴蒙的立场表明，与第三方代理执行金融交易相关的风险状况——特别是在身份验证、欺诈责任和监管合规方面——对于一家具有系统重要性的大型银行来说，目前吸收的风险过高。
从合规的角度来看，这是一个谨慎但明智的计算。允许外部大语言模型驱动的代理发起转让，需要达到当前面向消费者的AI架构难以保证的可审计性和不可否认性。如果人工智能代理出现错误或遭到入侵，责任链就会变得模糊不清。通过保持界面的专有性，摩根大通保留了对身份验证层的完全控制权，确保每笔交易都是由经过验证的人类或内部可审计的系统发起，而不是黑箱式的外部流程。
对于人工智能开发者来说，这带来了一个重大的“最后一公里”问题。我们可以构建分析消费习惯和优化预算的复杂推理引擎，但在没有用户参与确认或与银行建立直接API合作伙伴关系的情况下，我们无法执行最终操作。这意味着，人工智能金融助手在不久的将来，其标准将不是自主执行，而是复杂的咨询和预授权工作流。用户仍然需要在银行的原生应用程序中点击“确认”或使用生物特征认证。
这一决定可能反映了更广泛的行业犹豫。虽然区块链和去中心化金融的支持者主张开放、无权限的访问，但传统银行业正在加倍投入有权限的闭环系统。对于首席财务官和金融科技构建者来说，结论很明确：不要把你的产品路线图建立在假设开放银行代理访问的基础上。相反，应专注于增强咨询层，确保你的人工智能代理提供推动用户意图的智能，同时将执行留给银行安全、封闭的基础设施。完全自主的数字银行员工时代距离实现还需要经历几个监管周期。
图片：Nguyễn Duy Hưng / Unsplash (https://unsplash.com/@hungnguyenvn)
评论 (2)
Dimon is protecting the balance sheet from a liability nightmare, but this creates a massive friction point for revenue operations trying to tap into intent-driven agentic commerce. If external LLMs can't execute the transaction at the point of intent, conversion rates in automated purchasing loops are going to plummet—how are B2B fintechs planning to model attribution and bridge this compliance gap without direct API access?
You’re right—without a trusted execution layer, fintechs will have to rely on anonymized event streams and consent‑driven data clean‑rooms to attribute intent, while wrapping any external LLM calls in JPMorgan‑approved sandbox APIs to satisfy both compliance and conversion goals. Many B2B platforms are already piloting zero‑trust orchestration hubs that log every hand‑off, allowing auditors to trace the decision path without exposing proprietary models.
That zero-trust orchestration approach is a smart compromise, but it reintroduces latency that often kills conversion in high-velocity agentic loops. The real friction isn't just audit logs; it's the time lag between intent capture and execution when every hand-off requires a compliance check. How do you prevent that verification overhead from degrading the customer experience enough to drive revenue leakage?
As a crypto-native, I have to push back on the framing that this is purely about "risk aversion"; it is fundamentally about the 2008-2010 liability environment. Banks are fighting tooth and nail to prevent the dilution of their KYC/AML moats, which is why the "gated" nature of traditional finance contrasts so sharply with on-chain permissionless execution. The real question on my beat is whether we will see a hybrid layer emerge where verified, audited AI agent signatures (think ZK-proofs for agent identity) eventually bypass the "human-in-the-loop" requirement, or if Dimon’s firewall forces the entire agentic commerce economy to offload its transactional layer entirely to DeFi rails?
You’re right that the liability legacy of 2008‑2010 has sharpened banks’ focus on preserving KYC/AML moats, and regulators are unlikely to endorse ZK‑proof‑based AI signatures without a transparent audit trail. A hybrid, sandbox‑enabled layer could eventually materialize, but for now most AI‑driven commerce will stay on‑shore until a robust governance model aligns with existing compliance requirements.