An AI reviewer that checks every withdrawal but never moves money.

The challenge

Risk analysts reviewed withdrawal requests by hand. Spotting bonus abuse or linked accounts means reading a player's whole history, and slow reviews frustrate the honest majority of players.

The constraints

• Money movement must stay deterministic and auditable. An AI model may not approve a payout.
• Player-supplied text (names, comments) can carry prompt injection.
• Reading production data must carry no risk of writing to it.

Decisions & tradeoffs

• Advisory only: the platform notifies the reviewer about each withdrawal, and the reviewer returns a verdict with its reasons. The operator's payout engine, a chain of deterministic checks that fails closed, remains the only thing that can move money.
• Read-only by construction: the reviewer connects to Amazon Aurora MySQL as a user that can only run SELECT.
• Untrusted data stays data: player fields are wrapped and labelled as untrusted before they reach the model.
• Measure before trusting: the rule-based scorer was first replayed in shadow mode on 3,479 historical withdrawal requests, and a hand-labelled set of 999 requests was built to measure how often the reviewer's flags were right.

The implementation

The reviewer is a small service on Amazon EC2. For each request it gathers the player's history with read-only queries, asks Claude for a verdict with reasons, and posts the result back to the platform's review endpoint. Analysts also get a short summary in their team chat.

Outcomes

Analysts start each review with the relevant history and a reasoned recommendation instead of a blank screen. Because the reviewer is advisory and read-only, a bad verdict can slow a payout down but can never send money anywhere.

Handover & ongoing ownership

The labelled evaluation set stays with the risk team, so any change to rules or prompts can be measured against the same requests before it goes live.