How Grab Leverages Large‑Model Agents to Transform Risk Analysis

Introduction

Grab, a leading super‑app in Southeast Asia covering ride‑hailing, food delivery, and fintech, faces intense risk‑control demands due to high‑frequency monetary incentives and diverse financial services. Traditional risk‑control methods struggle with data silos and limited knowledge reuse, prompting Grab to develop a large‑model‑driven intelligent analysis platform that combines "knowledge injection" and "process reconstruction".

Key Topics Covered

Challenges faced by risk analysts

How large models and agents reshape risk‑analysis paradigms

Technical implementation of an agent‑driven enhanced analysis platform

Business practice and case studies

Future outlook

1. Grab's Risk Business Background

Grab's rapid expansion and incentive mechanisms attract fraudsters, while its fragmented Southeast Asian payment ecosystem creates fragmented, cross‑platform risk scenarios. Data sources are heterogeneous, including transaction logs, device info, biometric data, location traces, Wi‑Fi status, and credit‑card checks. Current risk decisions rely on a rule engine plus machine‑learning models.

2. Role of Risk Analysts

Risk analysts must quickly respond to incidents, perform deep data mining for business insights, and design or optimize rule engines. The high‑pressure, knowledge‑intensive nature of the role makes scaling difficult.

3. Limitations of Traditional Analysis

Growth in business scale leads to exponential demand for analysts, causing knowledge silos, over‑reliance on a few experts, and human errors that bottleneck efficiency and decision quality.

4. Large Models and Agents Reshaping the Paradigm

Traditional analysis follows a hypothesis‑validation‑decision loop heavily dependent on analyst intuition. Large models offer massive memory, multimodal understanding, and logical reasoning but suffer from knowledge hallucination and performance drop on complex tasks. Agents inherit these strengths and limits. Practical, scalable entry points are identified as knowledge injection and process injection.

5. Core Technologies of the Enhanced Analysis Platform

Architecture

The platform centers on a large model that orchestrates diverse plugins, some of which are agents themselves. A modular, loosely‑coupled design ensures easy upgrades.

Data Layer + UI Layer

The data layer consolidates multi‑source heterogeneous data, while the UI layer provides real‑time conversational interfaces, reducing context‑switching for analysts.

Knowledge Linking via RAG

Retrieval‑Augmented Generation (RAG) and model fine‑tuning connect business knowledge to the model. Two knowledge types are injected:

Unstructured business documents (risk reports, chat logs, assessment docs) are semantically indexed by RAG to fill model knowledge gaps.

Structured metadata (data dictionaries, high‑frequency SQL queries) is retrieved via RAG to ground the model in accurate data context.

These injections markedly improve reliability and decision accuracy in complex risk scenarios.

Complex Task Execution with SOP Trees

Standard Operating Procedures (SOP) are represented as hierarchical trees rather than linear code or static workflows. The large model dynamically generates prompts based on the current node, decides next actions using tool feedback, and traverses the tree, enabling flexible, explainable, and high‑quality risk analysis.

6. Business Practice Cases

Case 1: Analytics Copilot‑RAG

A self‑service analysis bot that empowers analysts and non‑technical users to perform risk queries and data visualizations via natural language, leveraging a Data‑Arks middleware and the SpellVault agent framework.

Case 2: RiskOps Autopilot‑SOP Agent

An AI‑driven automation agent that encodes SOPs as plain text, parses them into tree structures via depth‑first algorithms, and executes tasks such as credit‑card anomaly checks and risk report generation, achieving higher consistency and lower maintenance cost than code‑based solutions.

7. Future Outlook

Analysts and AI will collaborate deeply: AI handles roughly 90% of repetitive analysis (data checks, rule matching), while analysts focus on the remaining 10% of strategic decisions. Over time, agents will autonomously perform complex multi‑dimensional risk mining and dynamic strategy generation, shifting the analyst role from execution to decision hub.

Thank you for reading.