AI-powered sales assistant for a retail leader: Cutting customer request processing time by 90%

Yalantis partnered with a US electronics retailer that was experiencing increasing interest in their products year over year, resulting in an increasing workload for their sales department.

The retail company’s VP of Sales and Head of Innovation contacted Yalantis to validate the feasibility of implementing an AI assistant for the sales department. Sales specialists needed help quickly answering customer questions and offering suitable electronic devices. To efficiently close deals, the assistant also needed to be able to transfer more personalized interactions, such as discussions of pricing and discounts, to human sales representatives.

As soon as our AI team gathered the retail company’s requirements, analyzed the sales team’s pain points, evaluated the ability of an AI assistant to meet the client’s needs, and identified basic functionality, we began work on a PoC version of the AI sales assistant to help our partner decide whether such technology was worth investing in.

Delayed communications and a wide range of customer inquiries brought the following issues for the sales team:

• Lack of time for thorough responses. The sales team spent lots of time manually preparing exhaustive answers to customer inquiries. This slowed down their communication and potentially led to sharing incomplete information.
• Decreased customer lifetime value (CLV). Slow responses and inconsistent engagement from sales representatives due to the high workload impacted customer trust and satisfaction. Frustrated customers were less likely to become repeat buyers or recommend the store to others.
• Employee burnout and high turnover. Dealing with unsatisfied customers and doing lots of manual work caused sales specialists to lose motivation and eventually leave the company. Hiring new employees didn’t help, as training them slowed down the sale department’s performance even further.

To develop a high-functioning PoC that the sales team could test right after its release, the Yalantis project team took the following steps:

Step #1. Set up a knowledge base

Our business analyst and ML specialist collected and developed a specialized knowledge base encompassing detailed information about the retail company’s products, including their characteristics. The knowledge base is currently limited but will be extended when we move to the next project development stage.

Step #2. Selected an AI model

A Yalantis ML engineer researched and compared various open-source large languages models such as google-flan-t5-large, google-flan-t5-xl, mistral, and ChatGPT 3.5 Turbo. The final choice fell on ChatGPT 3.5 Turbo, as it:

• offered satisfactory natural language processing quality
• provided accurate responses
• had lower latency when processing customers’ requests than other models

Step #3. Performed prompt engineering and testing

As the next step, our project team created prompts to guide the AI model in having meaningful interactions with customers. Eventually, the model could detect and elicit essential requirements during a conversation and ask additional questions to determine which products would meet a customer’s needs.

To ensure the AI sales assistant performs as expected, a Yalatis QA engineer thoroughly tested the AI model by going through multiple made-up scenarios resembling real-life communication between sales representatives and customers.

Implementing Yalantis AI Virtual Assistant for streamlined development

For this PoC, the Yalantis team focused on selecting, training, and testing the most suitable AI model. But apart from a high-functioning AI model, the PoC AI assistant also required proper infrastructure, data pipelines, and a user interface. To expedite the development process, our team leveraged our internally developed Yalantis AI Virtual Assistant.

The reasoning behind our decision

AI Virtual Assistant streamlined setup of backend infrastructure and data pipelines. It incorporates industry-standard technologies like AWS Cloud, Kubernetes, and PostgreSQL to ensure a solid foundation for building AI assistants.

The core strength of AI Assistant Accelerator lies in its user-centric simplicity. Once the model was prepared, the Accelerator seamlessly integrated it with existing infrastructure and the front end, minimizing the development time and resources needed.

Further, our team customized the front end to align with the retail company’s requirements. This efficient integration process required just one combined day of work from backend and DevOps engineers, translating to significant cost and time savings for this project.

The main aim of the AI assistant is to:

• gather customers’ requirements by asking questions
• suggest suitable electronic devices
• provide detailed device characteristics

For personalized assistance with pricing and purchase options, and to discuss potential discounts, conversations are transferred to a dedicated human sales representative. Representatives can also provide additional resources like photos and videos showcasing device performance, or recommend alternative options based on a customer’s needs.

The PoC version of the AI sales assistant has already proven useful to our partner’s sales department. But we’re already planning further enhancements that will involve:

• increasing the accuracy of equipment selections
• enabling the AI assistant to recommend several products and process more complex cases
• extending the knowledge base to cover more products
• automating knowledge base updates and enabling continuous learning
• fine-tuning model/prompt engineering for new use cases and scenarios
• providing links to products during conversations with potential customers
• adding multi-language support to open up international shopping opportunities
• including voice recognition to further speed up customer request processing