An AI agent plans and carries out multi-step tasks: it reads inputs, calls tools and produces outputs with little prompting. In finance, the question is not whether agents are capable, but which steps they may take alone and which must wait for a person.
Three design rules
- Write permissions down. An agent should have an explicit list of actions it may take and, just as important, actions it may not.
- Set a confidence floor. Below an agreed level, the correct output is an escalation, not a weaker answer.
- Make approval a gate, not a suggestion. The agent should be technically unable to publish or act until a named person approves.
How frameworks frame it
The NIST AI Risk Management Framework organizes the work into four functions: Govern, Map, Measure and Manage. Bounded permissions are a Govern decision; confidence scores and data-quality checks are Measure; escalation and rollback are Manage.
The EU Artificial Intelligence Act classes certain financial uses as high-risk, including AI used to evaluate the creditworthiness of natural persons, and requires human oversight for high-risk systems. Long before generative AI, the Federal Reserve’s SR 11-7 already expected models to face effective challenge and ongoing monitoring.
A December 2024 paper from the BIS Financial Stability Institute surveys how financial authorities are approaching AI and discusses challenges including explainability, governance and reliance on third-party providers.
What this concept demonstrates
The Applied AI page runs three illustrative cases through such a pipeline. One passes cleanly, one passes with a data-quality warning, and one is stopped because its sources disagree. The point is not the verdicts but the fact that each stop is designed in advance.