Source-linked AI summary
Bridging Search and CRM: Productionizing AI Product Research Agents for Customer Re-Engagement
Mandar Kulkarni, Pooja A., Samir Shah
TL;DR
E-commerce platforms often separate search, recommendation, personalization, and CRM, limiting proactive re-engagement for exploratory product research. This paper deploys AI Product Research Agents that connect these workflows, achieving ~285% CTR versus earlier WhatsApp mobile campaign baselines and generating downstream purchases and GMV impact.
Problem
E-commerce systems are often optimized independently, limiting contextual, trustworthy support for exploratory product intents such as “best smartphones” or “latest 5G phones.”
Method
A production framework identifies high-intent, low-engagement users, grounds multi-agent product research in external knowledge and enterprise catalogs, and delivers personalized WhatsApp recommendations.
Results
~285% CTR compared to earlier WhatsApp mobile campaign baselines, alongside message forwarding, sharing, downstream purchases, and GMV impact.
Takeaways & Limitations
The deployment supports the practical viability of scalable multi-agent AI for personalized product discovery and CRM-driven customer re-engagement in real-world e-commerce.
Takeaways & Limitations
Read rate cannot reliably capture secondary recipients reached through forwarded messages, whereas CTR captures direct and indirect engagement through tagged product URLs.
Abstract
from arXiv · showhide
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.
1 Introduction
The paper presents a production-deployed framework that bridges search and CRM with AI-powered Product Research Agents to proactively re-engage users showing exploratory purchase intent but low engagement. Its modular research pipeline delivers personalized WhatsApp recommendations and demonstrates improved engagement, downstream purchases, and GMV impact in production.
- Motivation: Fragmented search, recommendation, personalization, and CRM systems limit sustained engagement across discovery, research, engagement, and re-engagement stages.Subjective queries such as “best smartphones” do not map cleanly to deterministic keyword-based retrieval.
- Contribution: The framework identifies high-intent exploratory users with low engagement and proactively reconnects with them through personalized WhatsApp recommendations.It bridges search and CRM workflows using AI-powered Product Research Agents rather than optimizing only in-session retrieval quality.
- Method: A centralized orchestrator coordinates Query Analysis and Discovery Agents that structure subjective intent and retrieve candidate products with supporting external evidence.Signals include product category, budget constraints, desired attributes, and latent preferences; sources include web search, expert reviews, and community discussions.
- Scalability: Asynchronous PySpark filtering selectively processes large-scale query logs using zero-click behavior, affluence indicators, business relevance, and subjective query patterns.This avoids executing computationally expensive research workflows for all user traffic.
- Production Results: ~285% click-through rates were achieved versus earlier WA mobile product campaign baselines in a 23-day deployment involving approximately 15K WhatsApp notifications.The deployment also showed forwarding and message-sharing behavior, downstream purchasing activity, and GMV impact.
2 Rule-based Search Log Filtering (PySpark)
A PySpark-based rule-based pipeline filters large-scale e-commerce search logs to identify high-potential queries for downstream product research. It combines explicit behavioral and business heuristics, including engagement, affluence, query language, and mobile-category criteria.
- Filtering framework: The PySpark framework systematically identifies high-potential queries from search interaction logs using explicitly defined heuristics derived from business insights and user behavior signals.Search logs provide user intent signals, behavioral patterns, and optimization opportunities.
- Filtering criteria: Queries with zero clicks are retained because failed downstream engagement may indicate an intent mismatch with retrieved results.The condition is implemented with a click-signal predicate using the PySpark filter() transformation.
- Filtering criteria: The pipeline retains high-affluence users through a categorical log-feature filter to prioritize queries with higher expected conversion impact.This criterion requires no external enrichment because affluence is directly available in the logs.
- Filtering criteria: Lexical predicates retain normalized queries containing subjective qualifiers such as best, latest, top, good, around, or under.These queries typically require reasoning and external knowledge grounding beyond standard retrieval.
- Filtering criteria: The dataset is restricted to the mobile phones vertical using a product-level log tag, while the pipeline remains extensible by modifying the filter predicate.Mobile phones is selected for its high query volume and rich ecosystem of specifications, reviews, and comparisons.
3 Proposed Agent Architecture
The proposed product research agent uses centralized orchestration to coordinate specialized agents through shared state, combining query analysis, broad web discovery, catalog grounding, and final validation. This hierarchical design improves instruction compliance while addressing product-matching and factual-correctness risks.
- Architecture: A centralized orchestrator coordinates specialized agents, which write outputs to shared state for downstream processing.The architecture is modular and agent-oriented, with each agent responsible for a defined functional unit.
- Architecture: The hierarchical architecture produces a significantly lower instruction-violation error rate than sequential agents invoked in a fixed predefined order.The comparison is reported as part of the ablation study in Section 6.
- Query Analysis: The Query Analysis Agent classifies discovery versus specific intent, product category, and budget constraints before returning structured signals to the orchestrator.Queries specifying a product model and variant are classified as specific.
- Discovery: Query expansion generates semantically diverse ranking-, comparison-, and feature-focused variants to improve web-retrieval coverage for candidate discovery.The Discovery Agent retrieves heterogeneous content from technology reviews, editorial blogs, and video-based sources.
- Catalog Grounding: The Flipkart Search Agent maps external candidates to available, serviceable catalog products and returns personalized pricing using account and pincode inputs.It grounds discovery results in internal catalog entries and replaces discovery-generated specifications with catalog information for consistency.
- Validation: A small possibility of incorrect product matching arises from selecting the first serviceable FSN among top-k results, so the Review Agent filters incorrect matches.The Review Agent also performs external re-validation and internal consistency checks before recommendations are presented.
4 Templated WhatsApp message
The system delivers concise, personalized product recommendations through templated WhatsApp messages containing product names, prices, agent-generated reasoning, and direct links. It shortens tracked product URLs through a sequential pipeline to preserve readability and support campaign-day click attribution.
- 4 Templated WhatsApp message: WhatsApp is selected as the primary notification channel for delivering product recommendations.Message generation is optimized for brevity because of WhatsApp’s strict length constraints.
- 4 Templated WhatsApp message: Each templated message includes the recommended product name, product price, agent-generated reasoning, and a direct product link.Direct links facilitate seamless navigation and faster purchase decisions.
- 4.1 Product URL Shortening and UTM Tags for Click Analysis: Full-length product URLs are often impractical in WhatsApp templates because they increase message length and reduce readability and user experience.The system addresses this constraint with an in-house URL shortening service.
- 4.1 Product URL Shortening and UTM Tags for Click Analysis: The in-house service converts original product URLs into compact redirect links while preserving their destination semantics.Compact links are embedded in outgoing WhatsApp messages.
- 4.1 Product URL Shortening and UTM Tags for Click Analysis: Campaign-specific UTM parameters are appended to original product URLs before shortening to enable downstream analytics and click attribution.Day-wise UTM tags are consistently applied to all product URLs distributed on a given day.
- 4.1 Product URL Shortening and UTM Tags for Click Analysis: The URL pipeline sequentially generates the original URL, adds campaign-level UTM parameters, and transforms it into a compact shortened URL.This ordering embeds tracking information before redirection links are created.
- 4.1 Product URL Shortening and UTM Tags for Click Analysis: When clicked, the shortened URL transparently redirects users to the corresponding original URL containing the embedded UTM parameters.The redirect mechanism supports accurate measurement of engagement.
5 Experimental Results
A 23-day production WhatsApp campaign using AI-generated product recommendations increased engagement through higher CTR and organic forwarding, while also showing downstream GMV impact. Manual evaluation further found high factual precision from review-agent filtering, and hierarchical supervisor-worker architecture performed best.
- Campaign deployment: 23 days: The production WhatsApp campaign automated daily recommendation generation and message delivery for mobile product discovery.The campaign delivered approximately 15K WhatsApp notifications during the evaluation period.
- Engagement results: Organic message forwarding caused visits to exceed delivered WhatsApp messages on multiple days, amplifying the recommendations’ effective reach.Forwarded recommendations generated additional clicks on embedded product links.
- Engagement results: ~8%: The relative increase in message read rate was modest, whereas the AI-agent-driven campaign achieved substantially higher CTR than historical WhatsApp mobile campaigns.CTR captures direct and indirect engagement through UTM-tagged product URLs, while read rate excludes reliably untracked secondary recipients.
- Business impact: 15 days: Downstream order-log analysis found substantial overlap among targeted users, WhatsApp-recommended products, and products later purchased, demonstrating GMV impact.The analysis covered user behavior after WhatsApp message delivery.
- Factual accuracy: 2,218 recommendations across 730 queries were manually evaluated for specification and launch-date accuracy, and review-agent filtering produced high factual precision.Any specification mismatch or launch-date discrepancy resulted in an irrelevant label.
- Agent architecture: Hierarchical supervisor-worker architecture provided superior performance among the evaluated agent architectures.This comparison is reported in Table 3.
6 Ablation Study
The ablation study compares centralized orchestration with a decentralized sequential-agent architecture. The sequential pipeline underperforms because intermediate transformations attenuate or lose critical instruction signals without centralized control, consistent with prior work favoring hierarchical architectures for complex tasks.
- 6 Ablation Study: The decentralized design invokes agents in a fixed sequence: query analysis, discovery, Flipkart search, then review.Each agent consumes the output generated by the preceding agent.
- 6 Ablation Study: Sequential processing degrades performance as intermediate representations progressively transform and attenuate or lose critical instruction signals.The degradation is attributed to the absence of centralized control.
- 6 Ablation Study: The findings align with prior work showing sequential agent pipelines underperform hierarchical architectures on complex tasks.The comparison supports centralized orchestration for preserving control across multi-agent workflows.
7 Related works
Recent LLM advances have enabled agentic systems with reasoning, planning, and tool-use capabilities for complex tasks. Related work applies LLM-driven agents to subjective, exploratory e-commerce queries, including gifting scenarios.
- 7 Related works: LLM advances have produced agentic systems that augment models with reasoning, planning, and tool-use capabilities for complex tasks.The passage cites several prior works supporting this trend.
- 7 Related works: Dammu et al. explored LLM-driven agents for subjective and exploratory e-commerce queries.Their work addresses e-commerce situations where user needs are subjective.
- 7 Related works: Gifting scenarios pose challenges because user needs are subjective and information-oriented.The related system leverages reviews, conversations, and web browsing.
8 Conclusion
The paper presents a scalable, end-to-end enterprise framework that bridges search and CRM through AI-driven Product Research Agents for exploratory product discovery and customer re-engagement. Production findings show higher CTR than prior campaign baselines, secondary sharing engagement, and downstream purchasing and GMV impact.
- 8 Conclusion: The framework unifies search and CRM workflows through AI-driven Product Research Agents in large-scale e-commerce environments.It is designed for exploratory product discovery and customer re-engagement.
- 8 Conclusion: The production pipeline integrates behavioral analytics, grounded multi-agent reasoning, recommendation validation, and proactive communication channels.These components form a unified mechanism for personalized product discovery and CRM-driven re-engagement.
- 8 Conclusion: The campaign achieved higher CTR than prior recommendation campaign baselines, with organic message forwarding and sharing indicating secondary engagement.It also generated downstream purchasing activity and GMV impact.