Bridging Search and CRM: Productionizing AI Product Research Agents for Customer Re-Engagement
This paper presents a production-deployed framework that integrates search and CRM systems using AI-powered Product Research Agents to identify users with exploratory intent and deliver personalized, grounded product recommendations via WhatsApp, resulting in significant improvements in click-through rates, engagement, and downstream revenue.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the vast digital landscape of modern shopping, a quiet friction often goes unnoticed. When a customer types a vague, exploratory question like "best smartphone for gaming" into a search bar, the system usually responds with a static list of products ranked by sales or keywords. This approach works well for specific requests, such as finding a particular model number, but it struggles with subjective needs where the user is looking for guidance, comparison, and trust. Consequently, many shoppers leave the app to consult external reviews, video channels, or forums before returning to buy. This journey is fragmented, breaking the connection between the platform and the user. To bridge this gap, researchers are increasingly turning to artificial intelligence systems that can reason, search, and synthesize information much like a human expert would. These systems, often called agents, do not just retrieve data; they understand intent, gather evidence from multiple sources, and construct a coherent answer. The question driving this new work is whether such an intelligent system can be deployed at a massive scale to not only answer these complex questions but to proactively reach out to customers who have wandered off, guiding them back with personalized, trustworthy advice.
A team of researchers at Flipkart, a major online retailer in India, has built and tested exactly such a system. They created a framework that connects the search engine, where users ask questions, with the customer relationship management tools that send messages to users. Their goal was to identify shoppers who were showing signs of confusion or hesitation—those who searched for subjective terms but did not click on any results—and then re-engage them with a helpful, AI-generated recommendation sent directly to their WhatsApp messaging app. The system they built is not a single program but a coordinated team of specialized digital workers, or agents, each with a specific job. One agent listens to the user's question and figures out what they really need, such as a budget limit or a preference for a specific brand. Another agent goes out to the wider internet to find expert reviews, comparison articles, and community discussions to identify the best candidates. A third agent checks these candidates against the retailer's own internal database to ensure the items are actually in stock, available for delivery to the user's location, and priced correctly with any applicable discounts. Finally, a fourth agent acts as a strict editor, double-checking all the facts, dates, and specifications to ensure the advice is accurate and free of errors before it is ever shown to a person.
This entire process is designed to be efficient and scalable. Instead of trying to run this complex research for every single search query, the system first filters the data to find only the most promising opportunities. It looks for specific patterns, such as users who have high purchasing power but did not click on any search results, and focuses its efforts on those cases. Once a high-potential user is identified, the team of agents works together asynchronously to build a recommendation. The output is a concise, personalized message sent via WhatsApp, containing a few product suggestions, their prices, and a clear explanation of why they were chosen. The researchers tested this system in a live environment over a period of twenty-three days. During this time, the system sent approximately fifteen thousand messages to different users, primarily focusing on mobile phones. The results were striking. The messages generated a click-through rate that was nearly three times higher than previous campaigns that used traditional recommendation methods. Even more telling was the behavior of the users who received the messages; on many days, the number of people who clicked the links was higher than the number of messages sent. This indicated that users were not just reading the messages but were actively sharing them with friends and family, who then clicked the links themselves, effectively turning the customers into a secondary distribution network.
The success of the campaign was not limited to clicks; it translated into real economic activity. The researchers tracked the users who received the messages and found that a significant number of them went on to purchase the recommended products, generating substantial revenue for the company. To ensure the system was reliable, the team conducted a rigorous manual review of thousands of the recommendations it produced. They checked the technical specifications and the launch dates of the phones against real-world data. The system proved to be remarkably accurate, with over ninety-nine percent of the specifications and launch dates being correct. This high level of factual correctness was largely due to the final "review" agent, which acted as a safety net to catch any inconsistencies or errors before the message was sent. The researchers also compared two different ways of organizing the team of agents. They found that a structure where a central manager coordinates the workers was far superior to a structure where the workers simply passed tasks down a line without oversight. The centralized approach significantly reduced errors and ensured that the system followed all the necessary rules, such as including launch dates only when the user asked for the "latest" models.
This work demonstrates that artificial intelligence can move beyond simple automation to become a proactive partner in the customer journey. By combining the ability to search the open web with the precision of a retailer's internal data, and by wrapping it all in a trusted communication channel, the system successfully re-engaged users who had previously disengaged. The study shows that when AI is used to provide grounded, explainable, and fact-checked advice, it can restore the connection between a shopper and a platform. The findings suggest that the future of e-commerce may not just be about showing more products, but about providing better, more intelligent guidance that respects the user's need for context and trust. The researchers have proven that such a system can be built, deployed, and scaled to handle the complexities of a real-world marketplace, turning a fragmented shopping experience into a cohesive and productive journey.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.