2026年2月25日，据报导 Deutsche Bank AG 与 Goldman Sachs Group Inc. 正在推进交易监控中的 agentic AI，这表明合规工具正从基于规则的方式，转向可自主规划分析并升级风险的系统。Deutsche Bank 正与 Alphabet Inc. 的 Google Cloud 建立大型语言模型工作流程，而其已表明的设计是由人工合规主管作为最终决策者，而非授予完全自动化。
Deutsche Bank 的计划目标是针对委托、交易与市场变动进行异常侦测；另一次于 2026 年稍晚推出的 LLM 部署，旨在监控交易员与销售人员的通讯行为，例如将机密资料转寄到个人电子邮件。Bernd Leukert 表示，合规现代化已淘汰旧有架构、关闭 200 台内部监控伺服器，并将误报降低超过 25%；目前监控涵盖超过 40 个内外部通道，且每天处理 1 terabyte per day（约 1000 gigabytes/day）的电子通讯。Goldman Sachs 也在评估 agentic 交易监控用例，而 Nomura Holdings Inc. 正在讨论具备受保护 IP 边界的跨银行模型训练，以及可能由监管机构支持的协作。
据报导，效率面的案例影响重大：Nomura 估计 AI 可将误报降低 30% 到 40%，并在合规成本上每年最多节省 $5 million；而 ThetaRay 等公司正把 agentic 方法延伸到反洗钱控制，服务对象包括 Banco Santander SA 在内的银行。但高层一再强调 human-in-the-loop 治理，且安全主管警告，控制不佳的代理会带来新的攻击面，包括意外资料外泄、未授权的营运行动，以及决策可解释性薄弱，因此导入正以分阶段部署推进，而非立即采取端到端自主化。
On February 25, 2026, Deutsche Bank AG and Goldman Sachs Group Inc. were reported to be advancing agentic AI in trading surveillance, signaling a broader shift from rule-based compliance tools toward systems that can autonomously plan analyses and escalate risks. Deutsche Bank is building large language model workflows with Alphabet Inc.’s Google Cloud, and the stated design keeps a human compliance officer as final decision-maker rather than granting full automation.
Deutsche Bank’s program targets anomaly detection across orders, trades, and market moves, and a separate LLM rollout later in 2026 is intended to monitor trader and salesperson communications for behaviors such as forwarding confidential data to personal email. Bernd Leukert said compliance modernization already retired legacy architecture, shut 200 internal surveillance servers, and reduced false positives by more than 25%, while current monitoring spans more than 40 internal/external channels and processes 1 terabyte per day (about 1000 gigabytes/day) of electronic communications. Goldman Sachs is also evaluating agentic trade-surveillance use cases, while Nomura Holdings Inc. is discussing cross-bank model training with protected IP boundaries and possible regulator-backed collaboration.
The reported efficiency case is material: Nomura estimates AI could cut false positives by 30% to 40% and save up to $5 million annually in compliance cost, and firms such as ThetaRay are extending agentic methods into anti-money-laundering controls for banks including Banco Santander SA. But executives repeatedly stressed human-in-the-loop governance, and security leaders warned that poorly controlled agents can create new attack surfaces, including accidental data exposure, unauthorized operational actions, and weak decision explainability, so adoption is proceeding in phased deployments rather than immediate end-to-end autonomy.