Agentic Commerce MCP

Context

Commerce teams keep bolting chatbots onto storefronts without connecting them to the systems that matter: catalog, inventory, order status, returns. The result is a demo that gets switched off after a quarter. This project set out to build an agent that could actually reason across commerce data and take useful action.

Constraint

The agent had to work with live commerce APIs, not a static knowledge base. That meant handling real-time inventory, personalised pricing, order mutations, and graceful fallback when the AI was uncertain. Every action had to be auditable, and the system had to know when to escalate to a human.

What I built

An agentic commerce system using Azure AI for language understanding and retrieval, with Model Context Protocol servers exposing catalog search, order lookup, cart operations, and support workflows as structured tools the agent can call.

Decisions

• Used MCP rather than function calling alone so the tool surface is defined by the commerce system, not hard-coded into prompts.
• Built retrieval-augmented generation over the product catalog so the agent answers from real inventory, not training data.
• Added explicit confidence scoring: the agent flags low-confidence responses and routes them to human support.
• Separated the orchestration layer from the tool servers so each commerce API can be versioned and tested independently.
• Logged every tool call and response for post-hoc review, compliance, and prompt iteration.

Outcome

The agent handles catalog questions, guided selling, and order support queries end-to-end. Escalation rate sits below 15% for supported query types. The MCP architecture means new commerce capabilities plug in without rewriting orchestration.

Stack

• Azure AI
• Model Context Protocol
• TypeScript
• Node.js
• RAG
• Vector Store
• REST APIs
• React