Grab's AI Agents for Automated Analytics Workflows
Grab has implemented AI agents to automate and streamline analytics workflows, significantly reducing the mechanical work for human analysts and accelerating business insights. This approach leverages a five-level autonomy model, allowing agents to handle tasks from data preparation to query execution and preliminary analysis, while humans retain oversight for critical decisions and interpretations. The system is supported by a robust data context architecture, including certified metrics, tables, and a ContextIQ system to ensure reliable agent performance.
Read original on InfoQ ArchitectureIntroduction to Grab's AI-Driven Analytics
Grab, a super app in Southeast Asia, has successfully deployed AI agents to automate significant portions of its analytics workflows. This strategic shift aims to free up human analysts from repetitive tasks, such as data preparation, alerting, and reporting, allowing them to focus on deeper analysis and self-service enablement. The initiative has led to a reduction in mechanical tickets handled by analysts from 44% to 30% and improved efficiency in answering business questions.
Five-Level Autonomy Model for AI Agents
The core of Grab's AI agent system is a five-level autonomy model that defines the scope of an agent's responsibility while ensuring human oversight. This tiered approach allows for progressive automation based on task complexity and criticality.
- Level 3: Humans frame the question and review results. Agents discover data, write/execute queries, validate results, and draft analysis.
- Level 4: Agents plan and orchestrate workflows, with human review at defined gates.
- Level 5: End-to-end autonomy, where humans set objectives, quality thresholds, and escalation rules. Human accountability remains for metric definitions, causal interpretation, business assumptions, and final decisions.
Spartan System and Data Context Architecture
Grab's Spartan system powers natural language analytics requests, supporting over 50 skills and 120 analysis frameworks to route requests to specialized workflows. This system can intelligently respond to queries, such as triggering root cause analysis or retrieving existing scorecards for experiment questions, rather than re-querying data lakes. A critical architectural decision for reliable AI agent performance is investing in comprehensive data context.
Ensuring AI Agent Reliability
For AI agents to produce reliable results, a robust and well-maintained data context is paramount. This includes certified metrics, tables, and detailed documentation. Grab's ContextIQ system manages this lifecycle, updating context as instrumentation changes and incorporating fixes from production failures.
- Over 5,000 certified tables and metrics.
- 4,000 context documents.
- 2,000 golden records.
AI Agents for Analytics Operations and Self-Service
Beyond business analytics, Grab also employs AI agents for operational tasks. Scarlet, an AI agent, handles pipeline failures by performing root cause analysis, attempting fixes, or escalating issues when necessary. For recurring analytics, agents automate metric and OKR commentary, breaking down movements and correlating them with operational changes. The BriX portal further empowers analytics workflow development, fostering significant growth in self-service capabilities and reducing the need for human involvement in routine requests.