Technical details

How the Order Management Assistant works

It works in Microsoft Teams. An AI model reads the question and calls one of six read-only tools on our MCP server. The server asks every connected distributor at once and merges the answers, so one question covers all of them. Every fact comes from tested code.

Try Demo
Watch Demo
Test with your own MCP client
Tools
Architecture

// .vscode/mcp.json
{
  "servers": {
    "order-management-assistant-demo": {
      "type": "http",
      "url": "<DEMO-URL>",
      "headers": { "x-api-key": "<TEST-KEY>" }
    }
  }
}

curl <DEMO-URL> \
  -H "x-api-key: <TEST-KEY>" \
  -H "content-type: application/json" \
  -H "accept: application/json, text/event-stream" \
  -d '{"jsonrpc": "2.0", "id": 1, "method": "tools/call",
       "params": {"name": "this_week", "arguments": {}}}'

# pip install "mcp>=2.3"
import asyncio

import httpx2
from mcp.client import Client
from mcp.client.streamable_http import streamable_http_client

URL = "<DEMO-URL>"
KEY = "<TEST-KEY>"


async def main():
    http = httpx2.AsyncClient(headers={"x-api-key": KEY})
    async with Client(streamable_http_client(URL, http_client=http)) as client:
        result = await client.call_tool("this_week", {})
        print(result.content[0].text)


asyncio.run(main())

At a glance

Runs in: Microsoft Teams, as a Copilot Studio agent
Who can use it: MDS staff we invite, signed in with Microsoft Entra ID
AI model: Claude, provided by Microsoft in Copilot Studio. The live demo uses Claude on Microsoft Foundry
Tools: 6 tools on our MCP server, all read-only
Data: Our distributors’ order and deal systems
Hosting: Azure App Service, run by MDS

How it answers a question

A person asks in Teams, signed in with Microsoft Entra ID. The agent’s AI model reads the question.
The model picks a tool by its name and description, and calls it on our MCP server. MCP (Model Context Protocol) is an open standard that lets an AI model call tools on a server.
The tool asks every connected distributor at once, applies our business rules, and returns facts. Each order names its distributor, and if a distributor doesn’t answer, the result says so.

The AI model never decides a fact or a number. Those come from code that has unit tests. The model only picks the tool and words the answer.

Architecture

Figure 1. Who runs what. Steps 1–6 follow one question. Microsoft runs Copilot Studio. MDS runs the MCP server. Each distributor runs its own system, shown here with three, as in the live demo. Every call to a distributor is a round trip: the question goes out, and the data comes back.

Figure 2. One question, every distributor. The tool asks each connected distributor at once and merges what comes back into one list. Each item keeps its distributor’s name, and a distributor that doesn’t answer is named, so a partial answer never looks complete. Dashed: a distributor that didn’t answer.

Figure 3. The live demo. Steps 1–4 follow one question. The demo server runs the real agent’s code on made-up test data for a made-up reseller and three made-up distributors, and Claude on Microsoft Foundry picks the tool and writes the answer. The demo has no connection to any distributor. The icons in these figures are Microsoft’s official architecture icons.

Tools

Six tools, one for each question the team asks most. Each one is marked read-only, and the server has no tool that places, changes or cancels an order.

Tool	Ask it	Inputs	What it returns	Limits
`this_week`	“What needs my attention this week?”	None	Deals that expire within 7 days, how many lapse within 14, and how many orders look stuck	A summary: the other tools give the details
`deals_to_follow_up`	“Who do I call first? I only have time for three calls.”	`days`, optional: 1 to 90 (default 14)	The deals that expire within that many days, grouped by customer, the customers with the most first	Three calls: the top three customers, then the others by name
`aging_deals`	“Which deals have been dragging longest?”	None	The deals revised most often, most revised first, with their expiry dates	Each revision counts, so the most revised deals have dragged longest
`customer_brief`	“Brief me on Woodgrove Bank.”	`customer`: a name	Its deals by vendor, which ones lapse soon, and its recent orders at every distributor	Matches every word of the name, in any case
`find_order`	“Where is PO 700245?”	`po_number`, or `customer`: a name	Every spelling of the PO together (700245 and 700245A), with each order’s status, distributor and age	If nothing matches, it says that an older order can still exist
`stuck_orders`	“Is anything stuck?”	None	Orders that shipped weeks ago but were never marked delivered, are still processing after many days, or are on hold, back-ordered or in credit review	Biggest totals first, each with its distributor

Test it yourself

Both ways use our demo server. It runs the same code and tools on made-up data for Trey Research, a made-up reseller, and three made-up distributors. No real client or distributor data is on it.

In your browser

Ask the live demo your own questions. A real AI agent picks the tool and writes the answer.

Try Demo

With your own MCP client

We send you the demo server’s address and a test key. Put them in place of <DEMO-URL> and <TEST-KEY> in the samples at the top of this page. Any MCP client that can send a header works, for example VS Code.

Ask for a test key

Ask the questions it was made for

What needs my attention this week?
Who do I call first? I only have time for three calls.
Is anything stuck?
Where is PO 700245?
Where is the Fabrikam hardware?
Which deals have been dragging longest?
Brief me on Woodgrove Bank.

You askWhat needs my attention this week?

The model callsthis_week {}

The tool returnsDeals that expire within 7 days: 4.
- 1 Proseware deal for Northwind Traders expires tomorrow: DL-24001.
- 2 Litware deals for Woodgrove Bank expire in 3 days: DL-24014, DL-24027.
- 1 A. Datum deal for Contoso Pharmaceuticals expires in 6 days: DL-24040.
Deals that lapse within 14 days: 10, at 4 customers.
Stuck orders: 5 orders, $41,595.40.

The model then writes a short answer from this text.

	In Teams	Demo server
Code, tools and rules	Ours	The same
Data	Our distributors’ systems	Made up, for Trey Research and three made-up distributors
Who can use it	MDS staff we invite, signed in	Anyone, in the live demo. With your own client: anyone with a test key
AI model	Claude, in Copilot Studio	Claude on Microsoft Foundry in the live demo, or the model in your MCP client
Key	Private	A test key from us

Limits

It only reads. No tool can place, change or cancel an order or a deal, so a person makes every change in the distributor’s own portal.
If a distributor doesn’t answer, the result names it and leaves its data out, so a partial answer never looks complete.
The demo answers a limited number of questions each day, and we can change test keys at any time.

Want one for your team?

Tell us how your team works today, and we’ll show you what this assistant could do.

Try Demo Talk to us