Everything K-culture — comebacks to K-beauty, straight to your inboxGet it in your inbox

METAL MEDIA

Bridging Search and CRM: Productionizing AI Product Research Agents for Customer Re-Engagement

arXiv:2608.185432026-08-20

AI agents win back window-shopping customers by chasing them down on WhatsApp

Flipkart built a production system where multiple AI agents research products for vague search queries like "best smartphones," verify the findings, and send personalized WhatsApp recommendations to re-engage users who left without buying. In a 23-day live deployment of about 15,000 WhatsApp messages, the AI-driven campaign achieved roughly 285% higher click-through rate than earlier WhatsApp campaigns, with signs that users forwarded messages to others. It also drove measurable downstream purchases and GMV.

METAL MEDIA explanatory visual

AI agents win back window-shopping customers by chasing them down on WhatsApp

  1. 01A PySpark-based filtering pipeline scans search logs to isolate high-value, disengaged users: queries with zero clicks, high-affluence users, subjective terms like "best" or "latest," and the mobile phones category
  2. 02A Query Analysis Agent reads user intent, a Discovery Agent gathers candidate products from web search, reviews and videos, and a Flipkart Search Agent grounds those candidates in real catalog data—checking stock, delivery feasibility, and personalized pricing
  3. 03A Review Agent re-checks specs and launch dates against catalog data and external sources, pruning inaccurate recommendations to reduce hallucinations
  4. 04A centralized orchestrator that coordinates all agents produced far fewer instruction-violation errors than a simple fixed-order (sequential) pipeline, tested on 2.2K recommendations
  5. 05Over 23 days, the campaign hit about 285% higher CTR than historical WhatsApp mobile campaigns, saw visit counts exceed messages sent on several days (suggesting organic forwarding), and generated downstream purchases and GMV impact
An explanatory diagram made by METAL MEDIA, not a figure supplied by the paper's authors.

What they did

  1. A PySpark-based filtering pipeline scans search logs to isolate high-value, disengaged users: queries with zero clicks, high-affluence users, subjective terms like "best" or "latest," and the mobile phones category
  2. A Query Analysis Agent reads user intent, a Discovery Agent gathers candidate products from web search, reviews and videos, and a Flipkart Search Agent grounds those candidates in real catalog data—checking stock, delivery feasibility, and personalized pricing
  3. A Review Agent re-checks specs and launch dates against catalog data and external sources, pruning inaccurate recommendations to reduce hallucinations
  4. A centralized orchestrator that coordinates all agents produced far fewer instruction-violation errors than a simple fixed-order (sequential) pipeline, tested on 2.2K recommendations
  5. Over 23 days, the campaign hit about 285% higher CTR than historical WhatsApp mobile campaigns, saw visit counts exceed messages sent on several days (suggesting organic forwarding), and generated downstream purchases and GMV impact
Figure 1. Proposed architecture of the Research Agent for Product Recommendations.
Figure 1. Proposed architecture of the Research Agent for Product Recommendations.
Table 1. Comparison of Historical WhatsApp (WA) Metrics for mobile product campaigns and AI Agent Campaign Interaction Statistics (23-Day Average).
% WA message readsCTR
AI Agent campaign~+8%~+285%
Table 2. Quantitative manual evaluation of factual accuracy
No. of productsspec accuracylaunch date accuracy
221899.1 %99.2 %
Table 3. Effect of the choice of the agent architecture. It is observed that hierarchical supervisor-worker architecture provides superior performance.
ArchitectureInstruction Violation Rate (%)
Centralized orchestration8.5%
Sequential Agents35.4%

Why it matters

This shows a concrete, deployed example of stitching together search, AI reasoning, and CRM messaging to recover customers who leave to research elsewhere—turning exploratory browsing into an actionable re-engagement channel with real revenue impact. It offers a working blueprint for teams considering agentic AI beyond chatbots, tied to measurable business metrics like CTR and GMV.

Terms in this paper

  • PySpark · a distributed data-processing framework for filtering huge log datasets quickly
  • CTR (click-through rate) · the share of message recipients who clicked the included link
  • orchestrator · a controller agent that manages the order and information flow among specialized AI agents
  • GMV · Gross Merchandise Value, the total value of goods sold through the platform
  • UTM tags · tracking parameters added to URLs to identify which campaign or message drove a click

Original abstract (English)

Modern e-commerce platforms often operate search, recommendation, personalization, and CRM systems independently, limiting opportunities for proactive customer re-engagement. This is particularly challenging for exploratory intents such as best smartphones or latest 5G phones, where users may leave the platform for external research before purchasing. We present a scalable, production-deployed framework that bridges search and CRM workflows through AI-powered Product Research Agents. The system identifies users with exploratory purchase intent and low engagement, conducts grounded multi-agent product research using behavioral signals, external knowledge, and enterprise catalog data, and delivers personalized recommendations through WhatsApp. We evaluate the framework in a 23-day production deployment involving approximately 15K WhatsApp notifications for mobile product discovery. The campaign achieved substantial CTR improvements over traditional WhatsApp recommendation campaigns, with evidence of secondary engagement through message forwarding and sharing. The deployment also generated downstream purchases and GMV impact, demonstrating the practical effectiveness of AI Product Research Agents for proactive customer re-engagement and end-to-end customer journey optimization.

Authors · Mandar Kulkarni, Pooja A., Samir Shah

Read on arXiv

Latest papers

All papers →

Latest from METAL MEDIA

Figures: Mandar Kulkarni et al., arXiv:2608.18543, CC BY 4.0