A fintech backend developer developed an AI agent to automate the investigation of conversion drops in the loan issuance funnel. The system uses a knowledge base of nine directories, including funnel steps, data sources, and release history, to analyze incidents systematically.
The agent follows a structured workflow: it verifies data freshness, identifies the start time of an anomaly, segments traffic, and correlates findings with system logs, releases, and external events. It assigns statuses like 'confirmed' or 'refuted' to hypotheses to prevent redundant checks.
In a September case study, the agent analyzed a drop in virtual card issuance for new clients. It determined that a new document verification check reduced the share of accounts with low-quality images from 22% to 6%, which explained the 66% to 62% decline in total issuance.
The agent also distinguished between the impact of the verification update and a separate planned feature disablement in a second product. The system stores these investigations, allowing the agent to leverage past experiences and human-verified corrections for future incidents.
The developer plans to integrate the agent with automated alerts. When a metric deviates, the agent will initiate an investigation, retrieve relevant historical cases, and present a report with evidence for human review, creating a continuous feedback loop for system improvement.
