Major Retailers Deploy Conversational and Agentic AI
Fashion retailer Mango has launched "Lisa," a proprietary conversational AI for both customer assistance and internal design, which has already co-created over 20 garments. Separately, SoundHound AI released a real-time Sales Assist Agent that provides sales associates with inventory data and customer preferences on demand. The launches demonstrate a trend toward deploying sophisticated AI agents directly into consumer and employee-facing retail environments.
- Mango's "Lisa" AI was developed in under nine months by integrating technologies from partners like Google, Microsoft, and OpenAI with their own internal platform, allowing them to train large language models with the company's specific data. The company began building its machine-learning platforms in 2018 and now has over 15, including "Midas" for pricing and "GaudÃ" for product recommendations. - The "Lisa" platform is intended for internal use to act as a "co-pilot" for employees, assisting in synthesizing large amounts of information for trend analysis, aiding in product ideation, and analyzing customer feedback. It complements another generative AI tool called "Inspire," which is used by design and photography teams to co-create prints, fabrics, and generate images for things like window dressing. - SoundHound's Sales Assist Agent operates on a multi-agent orchestration platform. This architecture uses an "Agent Coordination Engine" to manage communication and delegate tasks to specialized AI agents that connect securely to a retailer's backend systems like CRM, billing, and inventory databases. - SoundHound's system is designed for real-time, in-store environments and uses the company's proprietary Polaris automatic speech recognition (ASR) technology, which is built to function with low latency in noisy retail settings. The architecture also incorporates dynamic task distribution, which analyzes incoming requests and routes subtasks to the most appropriate agent based on its function. - The underlying infrastructure for such real-time retail AI often involves a central data platform that can integrate customer, behavioral, and transactional data. This frequently requires real-time data pipelines using event-driven architectures and streaming platforms like Apache Kafka or AWS Kinesis to feed AI models with up-to-date information. - For the generative and analytical AI models used in retail, a modern data stack often includes a cloud data platform like Snowflake for storing and processing large datasets. Tools like dbt are then used to transform this raw data into a clean, organized, and AI-ready format for machine learning models. - The management and deployment of the large language models (LLMs) these systems are based on is handled through a practice known as LLMOps, a specialized subset of MLOps. This involves managing the entire lifecycle of the model, including data preprocessing, fine-tuning, monitoring, and continuous deployment to ensure performance and reliability. - The goal of agentic AI in retail is to move beyond static, rules-based automation to a system where AI agents can sense, analyze, and act on real-time data autonomously. For a data engineer, this means a shift from batch-oriented data pipelines to building and maintaining adaptive, context-aware data flows that can respond to changes in metadata, business rules, and operational loads without failing.
