A shopper asks, "Is this size in stock and when will my order arrive?" and waits while the manager opens a stock export made that morning. The data is already outdated. Connecting an AI agent to PIM, inventory, and the order system through MCP servers closes this gap: the agent answers on stock, prices, and statuses using live data, without manual exports.
This is an open analysis based on public vendor materials, not a KT.Team case. Below is what MCP has made possible in e-commerce and retail, with links to the sources.
What MCP is and why retail needs it
Model Context Protocol is an open standard, first released by Anthropic in November 2024, that defines how an AI agent discovers and calls "tools" - functions with a predictable schema - in external systems. For the agent, it is a single adapter: instead of a dozen custom integrations, it works through one contract.
For retail, the key outcome is answers based on live data. The MCP server exposes specific tools ("check stock", "get price", "check order status"), and the agent reads the actual state of systems on demand rather than static website text or yesterday's CSV.
Catalog, pricing, stock, and orders through one layer
commercetools introduced Commerce MCP, which exposes the platform's key APIs in an agent-consumable format: carts, catalogs, prices, promotions, inventory and orders (commercetools). In the vendor's prototypes, three scenarios are shown: a shopping agent that uses real-time stock and prices for selection; a developer copilot; and a customer support agent that reads live order data to process a return or resolve an issue.
In its MCP server, Shopify exposes the catalog, stock, and orders with tools such as `check_stock` (stock check), `search catalog`, `create_cart`, and `apply_discount`. Metadata is delivered as inference-ready, prepared for model processing; a customer in chat can check real stock and see order status without human involvement (PrestaScan).
Logicbroker applies MCP to drop ship and fulfillment: the server provides tools for stock checks, order placement, price updates, and "intelligent routing" - linking catalog, orders, and partner fulfillment. Data is turned into "machine-readable schemas" that the agent queries in real time rather than parsing from pages (Logicbroker).
Why PIM comes first, not MCP
An important and underappreciated takeaway from inriver. MCP standardizes access an agent to systems and actions, but "no protocol defines the product information on which those actions depend" (inriver). In other words, MCP provides the mechanism, while the usefulness of the answer is determined by the quality and structure of the data in PIM.
inriver lists what PIM needs so an agent can answer accurately: structured product models (the agent understands exactly which product it is and how variants relate to stock), normalized data (less ambiguity), and API-first access. The practical business implication: fragmented product data across systems undermines the agent. Consolidation in PIM with governance must come before launching agent scenarios for catalogs and stock, otherwise the agent will confidently give wrong answers and manual correction will be needed.
This aligns with the engineering principle of "minimal modification of the tool core": the MCP server is deployed alongside with PIM/WMS/OMS as a thin layer of contracts, without rewriting or forking the systems themselves. Business logic lives in separate tool services, and the source systems remain decoupled and replaceable.
Governance: without it, an agent must not be given order access
All sources agree: access to sensitive operations must be governed strictly. Logicbroker describes role-based access control, an audit log for all transactions, and tool whitelisting to prevent unauthorized operations. commercetools calls this "governed execution": an administrator controls which tools are exposed and to what extent. Shopify separates Storefront MCP (for shoppers - search and checkout) from Dev MCP (internal configuration), giving external untrusted agents limited permissions.
The practical dividing line: read (stock availability, price, order status) can be opened broadly, and that is the main quick win; write (checkout, returns, cart changes) should be allowed only behind authorization and auditing. This applies a mature international standard instead of a homegrown "integration with access to everything."
What business result
- Customer and manager responses are based on current data instead of daily exports, eliminating the risk of selling what is not actually available and promising nonexistent delivery dates.
- Support reads the live order status and initiates a return directly in the conversation, not "by script."
- One MCP contract instead of a set of point integrations for each channel and each external agent.
What to change in the business process
Before: a manager exports stock and statuses to Excel once a day, responds late, and risks using outdated data. After: the agent reads stock, price, and status on demand directly from PIM, WMS, and OMS through MCP tools and responds using live data, while write operations go through RBAC and auditing. The bottleneck shifts from integrations to two things - data quality and structure in PIM, and governance over tool access. First, PIM is put in order, then read tools are exposed, and only after that comes tightly restricted write access.
