Financial organizations are moving away from asking a person to approve every step an AI agent takes. At AI Engineer New York, speakers from Two Sigma, Anthropic and Bloomberg argued that constant prompts create approval fatigue without guaranteeing that reviewers notice the decisions that matter.
Two Sigma described composable controls that each handle one job. Deterministic rules can process clear cases, while classifiers route ambiguous or consequential requests to people. This preserves human judgment without turning routine work into a queue of repetitive confirmations.
Anthropic framed the same boundary around identity, permissions and action. Its example trade-break agent can investigate and propose an order amendment, but a person must approve the actual change. The agent’s authority is therefore narrower than its analytical capability.
Bloomberg emphasized controls that run outside the model: scoped credentials, limits on tool calls and outbound traffic, audit trails, spending budgets and kill switches. External enforcement matters because a manipulated or mistaken agent should not be able to disable its own guardrails. The presentations did not establish that financial agents are ready to trade autonomously; predictions about broader capability in 2027 remain forecasts. Their practical message is narrower: define who an agent represents, what it may propose and exactly where acting requires approval.