MCP at work: connecting a conversation to order data

The assistant cannot guess where a parcel got stuck

“Where is this order stuck?” That is a question for company data. The model does not simply know it. This is where MCP, the Model Context Protocol, helps: an open standard for connecting AI applications to external systems. Its documentation describes access to data, tools and prepared workflows.

Suppose a tool accepts an order number and returns the placement date, payment status, fulfillment stage and update time. The model uses those details to answer an employee. The first integration has a specific scope: read an order and make the result available to the AI application.

“In progress” can mean several different things

Before connecting the systems, let’s agree on what the fields mean. Does a number identify one order? What happens if the record does not exist? Is the status fresh or two days old? The assistant should make those differences visible. Talk to someone who knows the data and can explain ambiguous statuses.

Start with reading and collect employees’ questions. If you later add address changes or cancellation, make each a separate operation with defined conditions. Before confirming, the user should see exactly what will change.

The system performing an operation enforces permissions. If an employee can access only one group of orders, the assistant has the same scope. We can describe the rules in a conversation, but the application must enforce access controls too.

How will we know the connection is useful?

After the first trials, count completed cases and corrections. Separate missing data from connection problems and misinterpretations. These are different faults: a new model will not refresh a status nobody has updated in the system.

Include the case number, source and a route to the usual application in the answer. Show that in the demo too. People can see both the help with an everyday question and where to go when a team member needs to investigate further.