A financial institution may need to block suspicious transactions in milliseconds while giving investigators enough context to review difficult cases. Defense in depth addresses both needs by layering distinct controls rather than expecting one model to handle every risk. Each layer has a different job and can cover a limitation of another. section 7.4.3
In fraud detection, separate the time-critical decision from the deeper investigation:
- Use a low-latency model such as XGBoost or LightGBM to block transactions in real time using structured data. section 7.4.3
- Run LLMs asynchronously behind that shield to process unstructured text and summarize dense alerts for human investigators. This preserves the fast layer’s speed while putting deeper analysis where it adds value. section 7.4.3
For a customer-facing generative AI response, protection continues after generation. Model routing can send simpler requests to a smaller, cost-effective model and complex or sensitive requests to a more powerful model on secure infrastructure. Before any response reaches a user, output guardrails check factual accuracy and brand tone and append required legal disclaimers. High-risk responses, such as financial advice, may require approval by a qualified employee. section 3.2.1
The layers also have operational costs. A fraud workflow must weigh higher infrastructure costs and integration time against risk appetite. Retaining LLM outputs for investigations can create substantial storage costs; requiring conclusions to cite the database rows or text snippets behind them helps create a defensible audit trail. Layering controls is therefore a design and governance decision, not a substitute for evaluating cost, speed, and explainability. section 7.4.3
