# Prosus's Journey with Production AI Agents ## Background and Context Prosus, a global technology group focusing on e-commerce, has implemented AI agents at scale across their organization. With operations in about 100 countries and platforms connecting buyers, sellers, students, teachers, restaurants, and diners, Prosus has been investing in artificial intelligence and machine learning since 2018. Their AI ecosystem includes approximately 1,000 experts and data scientists, with several hundred models in production generating hundreds of millions of dollars in impact. ## Toan: Enterprise AI Assistant ### Development and Architecture - Built in-house starting in 2020 with early experiments using GPT-3 - Implemented as a Slack-based and web-accessible assistant - Agent-based architecture with LLM-agnostic design - Connects to internal databases with strong security and privacy measures - Includes analytics and insights capabilities while respecting privacy - Composable design allowing companies to integrate specific components ### Technical Implementation - Stack components include: - Uses a variety of LLMs - Implements analytics model for usage statistics and continuous learning - Features data exploration capabilities with natural language queries - Includes query creation and execution tools - Visualization capabilities ### Performance Metrics and Adoption - Used by 24 companies - Over 15,000 active users - Nearly 1 million requests per month - Reduced hallucinations from 10% to 1% - Average time savings of 48 minutes per user per day - Cost per interaction stabilized at around 25 cents ### Cost Management - Token costs decreased 98% since initial implementation - Agent interactions showed 50% cost reduction between May-August 2023 - Increased complexity of queries led to higher token usage per question - Implemented various optimization strategies: ### Impact Assessment - Primary benefits beyond time savings: - Case study with iFood showed: ## OLX Magic: E-commerce AI Assistant ### Features and Implementation - Built on similar agent technology as Toan - Integrates with OLX marketplace platform - Key features: ### Technical Architecture - Specialized e-commerce-focused agent framework - Components include: ### Development Learnings - Evolution from chat-based to traditional e-commerce UI patterns - Importance of specialized e-commerce-specific agent design - Critical role of underlying search infrastructure - Need for custom embedding pipelines for listings - Integration of multiple data sources (titles, images, descriptions) ### Implementation Challenges - Required significant search infrastructure updates - Needed specialized prompt engineering for e-commerce context - Development of custom evaluation metrics for e-commerce use cases - Integration of personalization capabilities ## Overall Development Insights - Development team size: ~15 people for Toan - Development timeline: 2+ years for Toan, several months for OLX Magic - Continuous iteration and improvement based on user feedback - Focus on practical adoption and usability - Strong emphasis on cost optimization and efficiency - Balance between automation and human interaction - Importance of domain-specific customization