Jarvis Pro: The Routing System Behind AI Assistants
The technology powering Grab's AI assistant for account managers.
1Lessons from Incorrect Answers
The initial Jarvis Pro prototype could generate plausible but incorrect responses. In one case, it recommended promotions to a merchant when the real issue was operational: paused outlets and declining order fulfillment rates. This failure highlighted that fluency wasn't enough - accuracy was crucial. Tests in early 2026 showed the system achieved 99.4% routing accuracy.
2Structured Responses to Complex Queries
Jarvis Pro answers Grab account managers' complex questions like "What should I tell this merchant in tomorrow's meeting?" It first determines the question type, then loads relevant context before formulating responses. For example, if a store's order volume drops, it suggests operational fixes rather than promotions. This approach prevents misguided recommendations.
3The Critical Routing Guide
The system uses an LLM classification step constrained by a routing guide defining supported task types, permitted contexts, and response formats. Different queries like comprehensive merchant reviews or ad campaign recaps follow distinct paths. This structure ensures appropriate metrics and tools are selected for each task type.
4Balancing Memory Management
The system stores user context like country, role, and portfolio scope but prevents these from overriding explicit context. When users reference "my portfolio," stored information is used unless contradicted by specific context. This approach prevents incorrect assumptions while maintaining personalization.
5Resolving Metric Discrepancies
The system checks for differences in metric definitions and entity mappings across sources. Testing revealed that what appeared as model errors sometimes stemmed from data freshness mismatches between reporting surfaces. By normalizing source values and comparing approved metric paths, the system identifies such issues.
6Multi-Layered Evaluation Approach
Rather than relying on a single aggregate score, performance is measured across multiple layers using 351 labeled prompts. Evaluation assesses routing accuracy, context selection, and final response quality separately. This granular approach clearly identifies system weaknesses for targeted improvements.
7Usability-Security Tradeoffs
The system sequentially executes routing classification, memory checks, and context selection. To maintain usability, memory verification precedes expensive data retrieval, and warehouse queries are consolidated. While this slows initial response time, it prevents potentially unsafe recommendations from being generated quickly.
8Impact on Business Decisions
In testing, Jarvis Pro improved brand answer quality scores from 78.5 to 91.0. However, these results come from offline evaluations and don't prove business impact. In real-world use, the system could enhance operational efficiency by enabling account managers to make more informed decisions.
Original article: Grab Tech — Jarvis Pro: AI Asistanın Doğru Yönlendirme Sistemi