Introduction
Have you ever wondered why so many visitors leave an online store without buying anything? According to research, 80% people cannot find the right product quickly. Traditional keyword searches and manual filters often yield poor results, especially when customer queries are not properly formatted or incomplete. Businesses that use an AI-powered product recommendation system have seen conversion rates go up by 20%.
According to research (McKinsey), businesses that personalize customer interactions generate up to 40% more revenue. Yet, many businesses still rely on outdated discovery methods that don’t scale or understand user intent.
We built a system that provides a smart solution using LLM (Large Language Models), Agentic AI, and vector search to give accurate, context-based recommendations. It can understand unclear queries, give personalized recommendations, support real-time conversations, helping businesses improve customer experience, reduce support efforts, and boost sales.
This case study demonstrates how we implemented a product recommendation system leveraging the Crew-AI framework to enable intelligent, scalable, and real-time product discovery, transforming the way businesses engage with customers online.
Challenges:
Businesses that sell a wide range of products often deal with large, growing catalogs. As the number of SKUs increases, it becomes harder to help customers find the right item quickly and accurately. This leads to several common problems:
- Lack of Personalized Recommendations: Traditional search tools and filters don’t remember what the customer likes, so every experience feels generic. About 7 out of 10 shoppers expect stores to give recommendations based on their preferences. Without personalization, customers can get frustrated and leave.
- Time-Consuming Search: When a store has thousands of products, the system needs to upload all the product information (like name, price, size, color, and description) and then organize it in a way that makes searching possible. The bigger the catalog, the longer this process takes.
- Unclear Queries: Many customers search using short or unclear phrases. Traditional search engines often fail to show relevant results if users don’t enter exact product names or details, leading to 78% of consumers receiving irrelevant results.
- Poor Real-Time Performance: A study found that 80% of users abandon their search and leave without buying due to poor search performance. When search tools are slow or deliver irrelevant results, customers often give up before finding what they need.
- No Memory of Past Questions: In many online chats, users ask follow-up questions or build on earlier queries. Without memory, the system starts from scratch each time, creating a poor conversational experience.
These inefficiencies directly impact revenue and customer loyalty. Poor product discovery leads to higher bounce rates, lower conversions, and a growing gap between customer expectations and business capabilities.
The Solution
The primary objective was to solve key challenges customers face when browsing online stores. Many users struggle to find the right product or remain uncertain about what to purchase until they receive meaningful recommendations. For instance, one client operating an e-commerce brand offering industrial solutions faced significant difficulties, as customers often did not know which product would address their specific needs from an extensive product catalog. This led to frustration and placed additional pressure on the client’s support team.
Standard recommendation plugins proved insufficient for this complex use case. We evaluated multiple solutions, including Q&A chatbots and RAG-based systems, but ultimately implemented an AI-powered agentic approach. This strategy provided superior scalability, higher accuracy, and seamless integration across multiple channels.
The recommendation engine we developed enables customers to find the right products instantly, even when using simple, incomplete, or ambiguous language. It transforms a slow and frustrating process into a fast, intuitive, and highly efficient shopping experience.
Turning Catalogues into Searchable Data
Our first technical step was to process the product catalogues. Many of these files were very large, unorganized, or inconsistent, which made it challenging to build a reliable system. To solve this, we designed an optimized ingestion pipeline where data were split into manageable chunks and embeddings (a way of turning descriptions into numbers so that AI can understand) were generated in batches. This approach ensured that no data was lost during processing, and the process is also fast.
We used Google Gemini because it is highly accurate in understanding natural language and can handle large amounts of data efficiently. With Gemini, we created vector embeddings. These allowed the system to understand products based on context and user intent.
Once generated, the embeddings were stored in Milvus, a specialized vector database designed for fast searches. This setup enabled the system to deliver real-time responses even when working with thousands of products.
Building an Intelligent Conversational Layer
Once we had the product data ready, we built a conversational AI agent with CrewAI to guide the workflow. We connected it with Gemini so it could understand user questions and reply naturally.
Users often typed vague requests such as “show diamond rings below $500” or “I need an engagement ring.” To make sure the system still understood them, we used fuzzy matching with Gemini’s embeddings. This way, the AI could pick up on intent even when the words weren’t exact.
Remembering Context Across Conversations
We used Redis because it is fast and designed for real-time data storage. It can quickly save and retrieve chat history, which makes it ideal for maintaining context in conversations. Unlike traditional databases, Redis keeps data in memory, so the AI can recall past interactions instantly. This ensures smooth, human-like conversations, and the user doesn’t have to repeat information.
Scaling for Speed
The next step was to make sure it could handle very large product catalogs without slowing down. To achieve this, we optimized Milvus, the vector database used for search, by applying smart indexing techniques. Indexing works like creating an organized “map” of the data, so instead of looking through every single product one by one, the system can quickly jump to the most relevant results.
With these improvements, the AI could handle catalogs that contained thousands of products, while still keeping the search experience fast.
Bringing the AI Across Channels
Once we ensured the system was running smoothly and delivering accurate results, we integrated seamlessly across multiple channels, including WhatsApp, Facebook Messenger, and a web-based chatbot. Other than this, we can implement platforms like WordPress, Shopify, and WooCommerce. This means users can interact with the AI wherever they prefer, whether they’re browsing on a site or chatting on WhatsApp.
The integration ensures that live inventory and pricing remain consistent across every touchpoint, so no matter where customers ask, they always see the most up-to-date products. We also added smart fallback messages, so even if a user asks something unclear on any channel, the AI still responds politely and helpfully. With continuous monitoring, the experience stays fast and reliable across platforms, turning the chatbot into a true virtual shopping assistant rather than a single-channel tool.
Outcomes
After the AI product recommendation system was implemented, the way users searched and discovered items completely changed. What once required endless scrolling through catalogs and manual filtering has now become an effortless experience. The entire shopping process became faster, smarter, and far more engaging.
Customers no longer had to struggle with vague searches or incomplete queries. The system understood what they meant, even if the request was not specific. This meant fewer abandoned searches, quicker results, and happier customers. With CrewAI’s agentic approach, our recommendation system not only explains why certain results appear but also fixes issues on its own. If a database query gives wrong or incomplete answers, the system automatically improves the query and runs it again until the right results are found. This makes the system more reliable, flexible, and less dependent on human support.
Some of the key improvements included:
1. Faster Product Search: Product searches that previously experienced delays now execute in under a second, even with large product catalogs. Studies from Amazon show that reducing search latency by just 100ms can increase revenue by 1%, highlighting the significant impact of speed optimizations. Customers instantly noticed how much faster things became.
2. Smarter AI Recommendations: Instead of relying on exact keyword matching, our AI understands what customers really mean by analyzing intent and context. Even if the search is vague or incomplete, it still delivers proper suggestions and helps increase conversion rates by up to 320%.
3. Personalized and Natural Conversations: The chatbot delivers smooth, human-like conversations, allowing users to ask follow-up questions, compare products, and refine searches without starting over. By remembering past interactions and preferences, recommendations become more relevant, leading to conversion increases of up to 23% while enhancing the overall customer experience.
4. Reduced Support Load with 24/7 Support: Customers now receive instant responses at any time, while support teams are relieved from handling repetitive queries. Real-world implementations of AI chatbots demonstrate the ability to manage up to 80% of routine inquiries, reducing support workload by 30% and enabling teams to focus on higher-value tasks.
As a result, customers can find products faster, businesses can serve more users without expanding staff, and support teams can concentrate on complex tasks, all while delivering a seamless and highly personalized shopping experience.
Use cases
1. Jewelry and diamonds
Customers browsing a large online jewelry store often typed incomplete or vague queries, such as “Ring for engagement” or “Show gold necklace.” As a result, search results took a long time to display the right products, and sometimes the correct items did not appear at all. This caused frustration and led many visitors to leave the site without making a purchase.
We built a product recommendation chatbot that helps users find the right jewellery based on their preferences. The chatbot uses OpenAI LLM to understand simple, natural language questions and searches the product database to suggest the best matches.
With OpenAI Function Calling, the chatbot will be able to extract metadata such as shape, cut, carat, and price from even vague customer requests.
This means there’s no need for complicated rules or manual setup, making the system smarter, faster, and able to handle different kinds of questions easily.
For example, a customer can type,
“Show me diamond rings under ₹1,00,000 for an engagement,”
“I’m looking for a gold necklace with a pendant under ₹50,000.”
The system captures essential details such as shape, cut, carat, and price, and intelligently suggests the top 5 matching items.
This solution simplified the jewellery shopping experience, making it intuitive, fast, and personalized. Customers now find exactly what they want in seconds, reducing frustration and increasing the likelihood of completing a purchase.
2. Lighting solution
A lighting brand offered a large catalog of LED bars, pods, halogen lights, and accessories. Customers often did not know which products were best suited for their vehicle. Technical specifications such as wattage, beam patterns, and IP ratings were confusing. On top of that, if a user asked a follow-up question, the system did not remember the previous context, so every interaction felt like starting over. This made the search experience frustrating and often led to lost sales.
We built a product recommendation system using an Agentic AI approach with Crew AI to help users quickly find the right products for their vehicle. The system understands user questions, guides them step by step, and gives accurate recommendations based on vehicle type, specifications, and user needs. It also remembers past conversations, so if a user talked about their specific vehicle days ago, the system will recall that information and suggest relevant products when they return later.
“Show me 6-inch LED light bars under ₹15,000 for a Jeep.”
“I need a 10-inch light in white for a Ford Ranger.”
The chatbot remembers past interactions during the conversation, so if users refine their preferences, it dynamically updates results while keeping prior details intact.
For instance, if the user first asks,
“What is the price of a 50W Spot Beam?”
Then follow-up questions like,
“Does it include a wiring harness?”,
“What is the power rating?”
This solution transformed the product search experience, making it fast and personalized. Customers can now find lighting products suited to their exact vehicle quickly and engage in a natural conversation, leading to higher engagement and increased conversions.
3. Weighing scale
One of our clients sells different types of weighing machines for various purposes. Customers often didn’t know which product fit their needs. Many were unfamiliar with important specifications like capacity, accuracy, portability, and measurement units, which made choosing the right product difficult.
On top of that, the search system was slow and often didn’t show relevant results. This caused frustration and confusion, and many customers left the site without buying anything.
Initially, the recommendation accuracy was around 60%. After improvements, we increased it to around 85-90%, helping customers find the right products more easily.
We developed a product recommendation chatbot for weighing solutions that helps users quickly find the right product for their specific use case. Customers can type queries such as:
“I need something for my bakery that measures up to 100 kg.”
“I need a heavy-duty scale for warehouse use.”
“Show me something within a budget of $200.”
The chatbot understands the natural language input, extracts key attributes like capacity, portability, and price, and instantly suggests the most relevant products.
Now, customers find suitable weighing machines based on their needs without technical expertise, enjoy a natural conversational flow, and experience instant, relevant suggestions. This has led to fewer abandoned searches and improved sales for our client.
Conclusion
In conclusion, our AI-powered Product Recommendation Chatbot transforms the traditional e-commerce experience into an intelligent, efficient, and highly personalized journey for customers. By leveraging advanced technologies such as Large Language Models, Agentic AI, vector search, and real-time memory management, we address critical pain points in product discovery, eliminating slow search, irrelevant results, and poor conversational continuity.
For business leaders and CXOs, this solution delivers measurable impact: faster search speeds improving customer retention, smarter recommendations driving higher conversions, and significant reductions in support costs through automated, 24/7 assistance. Most importantly, it enables scalable, data-driven personalization across channels, helping your organization stay competitive, increase customer satisfaction, and grow revenue without proportionally increasing operational overhead.
Investing in intelligence like this positions your business to exceed rising customer expectations, improve decision-making, and accelerate digital transformation in an increasingly competitive market.
FAQs
1. How does the chatbot know which product to recommend?
The chatbot uses AI to understand customer requirements (like features, budget, or usage scenario) and matches them with the most suitable options from the catalog.
2. Do customers need to know technical specifications before asking?
No. They can describe their needs in simple language, such as “I want something for my bakery” or “I need a solution for shipping parcels,” and the chatbot will guide them.
3. Can customers compare multiple products?
Yes. The chatbot can show top recommendations side by side, highlighting key differences in features, performance, and pricing.
4. Does the chatbot remember previous questions?
Yes. It retains conversation context, so customers can refine their queries without repeating details (for example, “Show me the same option but with higher capacity”).
5. Is the system customizable for different industries?
Absolutely. Whether it’s retail, food, logistics, or manufacturing, the chatbot adapts to any product catalog and helps customers find the right fit.
6. How long does it take to implement this system?
For a typical product catalog with a few thousand items, initial setup and integration can take 2–3 weeks. This includes data preparation, AI model setup, and chatbot integration with your website.
7. Can this system scale with a growing product catalog?
Yes. The system is designed to handle large catalogs. As new products are added, embeddings are updated automatically, ensuring recommendations stay accurate and up to date.
8. What are the infrastructure requirements?
It can be deployed on cloud platforms (AWS, GCP, Azure) or on-premise. We use scalable components like vector databases, memory stores, and AI models, so the system runs smoothly even with heavy traffic.
9. How does this reduce support team workload?
By automating repetitive product-related queries, the chatbot handles up to 70–80% of common customer questions. This reduces operational costs and frees human agents to focus on complex or high-value tasks.
10. Can the chatbot be branded and customized?
Yes. Businesses can fully customize the chatbot’s tone, design, and recommendation logic to align with their brand and customer experience goals.
