Problem
Fraud analysts investigate account takeover (ATO) cases by hand. For each ticket they pull data from several systems, run many SQL queries and decide against fixed thresholds. The work is slow and needs SQL skills, which limits who can do it. Fatigue leads to inconsistent calls that can let attacks through.
Example from the article: a Slack ticket asks to investigate a web user and a transaction. The agent checks the IP address against web and app login history and links it to other accounts.
Solution in brief
Write the analyst's procedure down as a Standard Operating Procedure (SOP): a natural-language decision tree where each step names the data it needs and the function that fetches it. A planner walks the tree depth-first, a worker agent runs the queries, and the system ends with a summary and a decision. Grab reports 87% of ATO cases automated and handling time down from 22 to 3 minutes.
Questions this case answers
- How to automate a multi-step expert investigation that is a decision tree over database queries, so that an LLM agent follows the expert's procedure rather than improvising
- How to automate multi-step account takeover investigations that analysts now do by hand with SQL?
- How to make an LLM agent follow an analyst's investigation procedure instead of improvising?
- How to cut handling time per fraud ticket without losing the decision trail?
Tools and methods
- SOP decision tree — the analyst's procedure as indented natural language, with
@function_namemarks where data is fetched - SOP planner — walks the tree depth-first and picks the next step and the API to call
- Worker agent — reads the JSON form of the SOP, runs SQL queries or API calls and stores results
- Slack — the interface where analysts request an investigation and receive the report
Google or a one-time email code. No password.