Conversational Commerce Chatbot

Inspiration

We built Conversational Commerce Chatbot for the Visa challenge at LifeHack 2026. The problem statement highlighted how online shopping is often fragmented: users browse products, compare options, manage carts, and pay through separate pages, while smaller merchants may lack the resources to develop their own AI-powered shopping experiences.

As a team of three Year 1 NUS students, we wanted to explore whether these steps could be brought into one simple conversation. We chose footwear as our initial category because customers commonly need help comparing factors such as price, intended use, comfort, and style. Our goal was to create a realistic prototype that helps customers move from discovery to payment while keeping them informed and in control.

What it does

Conversational Commerce Chatbot is a category-focused shopping assistant that supports the full discover, decide, and pay journey within one chat interface.

The chatbot asks follow-up questions about the shopper’s needs, including their budget, preferred style, and intended use. It then searches products from multiple demo merchants and provides recommendations with match scores, reasons, and trade-offs. Customers can view product cards, compare options, and manage their cart through natural-language messages.

A single cart can contain products from different merchants. When the customer is ready, the chatbot presents an itemised transaction preview showing the products, quantities, merchant-specific taxes, and final total. The customer must explicitly select “Confirm & pay” before the system starts a simulated Visa payment.

The project also includes a merchant-facing flow. Smaller merchants can upload a product catalogue using CSV files, while the interface includes a mock API connection option to represent how larger retailers could integrate an existing inventory system.

How we built it

We organised the project as a JavaScript and TypeScript monorepo with separate services for the AI agent, product catalogue, payment flow, shared data contracts, and frontend.

The conversational layer uses the Anthropic API together with a controlled set of tools. These tools allow the agent to save shopper preferences, ask clarifying questions, search the catalogue, retrieve product details, generate recommendations, compare products, update the cart, and request checkout.

The catalogue and payment services are built with Node.js, Express, and PostgreSQL. Our local setup also uses Redis and Docker. The frontend is a lightweight web interface with chat messages, quick replies, recommendation cards, comparison tables, a cart grouped by merchant, and an in-conversation checkout card.

For our demonstration, we created two fictional footwear merchants with different products and tax rates. The payment service is a mock Visa integration that supports tokenisation, successful transactions, and test declines. For a multi-merchant cart, the checkout process creates one simulated charge per merchant while presenting the customer with one confirmation step.

We also separated the AI from the final payment action. The agent can prepare a checkout request, but only the application’s trust layer can initiate payment after receiving the customer’s confirmation.

Challenges we ran into

One major challenge was preventing the AI from presenting incorrect prices, availability, or totals. Although the model can suggest products, we decided that the application should remain responsible for verifying product information and calculating the final amount. We therefore revalidate recommendations against the live catalogue and calculate checkout totals on the server.

Designing a payment experience inside a conversation was another challenge. We wanted the flow to feel convenient without making it seem as though the AI could purchase items independently. We addressed this by introducing a non-skippable transaction preview and explicit confirmation step.

Supporting multiple merchants also introduced complexity. Products can have different tax rates and need to be charged separately, even though the customer sees one cart. We had to keep the user experience simple while handling this separation in the backend.

As first-year students, connecting several services was new to us. We had to learn how to define shared schemas, manage database migrations, coordinate frontend and backend development, and debug an end-to-end system within the hackathon timeline.

Accomplishments that we're proud of

We are proud that we completed a working end-to-end prototype covering the main flow required by the challenge: product discovery, comparison, cart management, transaction preview, confirmation, and simulated payment without redirecting the customer away from the conversation.

We are also proud of the safeguards built into the architecture. The AI does not have access to a direct payment tool, card details are tokenised in the mock payment service, totals are calculated by the server, and the application records an activity and checkout audit trail. Checkout is also blocked if a product’s price or availability changes after the preview.

Another accomplishment was supporting products from multiple merchants in a single conversation. Our CSV onboarding flow gives smaller merchants a relatively simple starting point, while the modular catalogue service and mock API connection demonstrate how the concept could later support merchants with existing systems.

Given that we are a team of three Year 1 students, building and connecting the conversational, catalogue, merchant, frontend, and payment components was a meaningful technical milestone for us.

What we learned

We learned that developing an AI agent involves more than connecting a language model to a chat interface. The agent needs clearly defined tools, reliable product data, server-side validation, and boundaries around actions with real consequences.

The Visa challenge also made us think more carefully about trust in agentic commerce. A smooth checkout is useful, but users must still understand what they are buying, how much they will pay, and when a transaction will occur. Consent and transparency therefore need to be part of the system’s architecture rather than added only to the interface.

From a technical perspective, we gained experience with Node.js, Express, PostgreSQL, APIs, shared schemas, prompt and tool design, and multi-service integration. As a team, we also learned to divide work according to system components while agreeing on common interfaces early.

What's next for Conversational Commerce Chatbot

Our current project is a hackathon prototype built around footwear, two fictional merchants, and simulated Visa payments. The next step would be to test the conversational flow with more users and improve how the chatbot asks questions and explains its recommendations.

We would also like to expand beyond footwear by creating reusable category configurations for areas such as electronics or food ordering. For merchants, we plan to improve the onboarding dashboard, support more catalogue formats, and develop a functional API connector for retailers with existing inventory systems.

On the payment and trust side, future work could include a Visa sandbox integration, stronger identity verification, persistent user accounts, order history, refunds, and more detailed handling of failed or partially completed multi-merchant payments.

In the longer term, we hope to turn the prototype into an embeddable commerce widget that merchants can configure for their own catalogue and brand while retaining the same confirmation, validation, and transparency safeguards.