3 hours
a day
30 requests at 6 minutes each, across the shop, the management system and the courier’s website.
Technology: MCP and AI assistants
The data sits in three different programs. With MCP an assistant queries them for you.
Orders in the online shop, picking in the management system, tracking on the courier’s website: answering one simple question means opening all three. MCP is the standard that lets an AI assistant use those programs for you, only as far as you allow.
At a glance
- Used for
- Answering questions that need data from several programs
- Uses
- The APIs and data of the programs you already have, unchanged
- Can do
- Only what you allow: it starts read-only
- Where it runs
- The MCP server on your servers, the model in the cloud or in house
A scene that repeats
A customer calls. Three programs open.
-
Tuesday, 10:05
A customer calls: the order has not arrived yet. The operator opens the online shop’s back office and looks the order up by name.
-
Tuesday, 10:07
Next, the management system, to see whether the parcel was packed. It was, yesterday afternoon, with its tracking number.
-
Tuesday, 10:09
The tracking number goes into the courier’s website: out for delivery tomorrow. The customer gets the answer.
-
Tuesday, 17:30
By the end of the day there have been twenty-eight calls like this. Each one with three programs open.
All the information is there. It sits in three different programs, and putting it together takes a person every time.
Before and after
The same call, two ways.
Today
- The customer asks where the order is
- The order is looked up in the online shop
- The management system shows whether it has shipped
- The tracking number is copied into the courier’s website
- The customer gets the answer
About six minutes and three programs open, for every call.
With an assistant connected through MCP
- The operator types the question
- The assistant queries shop, management system and courier, read-only
- It puts the answer together and says where each piece of data comes from
Where is order 2381 for Rossi?
Packed yesterday at 16:20 and handed to the courier. Tracking says: out for delivery tomorrow.
Data from: online shop, management system, courier tracking
A few seconds, without opening three programs.
In detail
MCP in four steps.
MCP stands for Model Context Protocol: a standard way to give an AI assistant tools, each with its own permissions.
- 01
Decide what it can do
You choose which data and actions to make available: look up an order, read a stock level, check a tracking number. To start, read-only.
- 02
Build the MCP server
A small program that talks to the management system, the shop or the courier, usually through their APIs, and presents those tools to the assistant in a standard format.
- 03
Connect the assistant
An AI assistant that supports MCP sees the tools and uses them when it needs them to answer. It can be one you already use or one running on your own servers.
- 04
Log and confirm
Every request goes into a log. Actions that change data, such as editing an order, ask a person to confirm.
MCP or API?
They are not alternatives: the MCP server usually uses your programs’ own APIs. What changes is who decides what to ask.
An API on its own
open orders for customer 4410
- Order 2381, open, 3 lines
- Order 2390, open, 1 line
It answers the question written into the code in advance. A new question needs new development.
The same APIs, with MCP
Does Rossi have any orders stuck? Why?
Yes, one. Order 2390 is waiting for 4 parts from the supplier, due Thursday. Order 2381 has already shipped.
Data from: open orders, stock, supplier deliveries
The assistant combines the tools you allowed, even for questions nobody planned for.
When it pays off
It pays off when one answer means opening three programs.
It makes sense if
- Answering frequent questions means opening several programs
- The data already exists: in the management system, the shop, the courier
- The questions change, and you cannot build a report for each one
- You want an AI assistant, but with clear permissions
You do not need it if
- The question is always the same: a report or a screen will do
- The process has fixed rules: rules are what you need
- Data only has to move from one program to another: a connection will do
An example with numbers
Thirty calls a day about where an order is.
Reference figures, to redo with your own.
Over 60 hours
a month
Over 21 working days: time taken from the people who look after customers.
Read-only
to start
The assistant reads and answers. Actions that change data come later, with confirmation.
How to start: a read-only MCP server on a single program, for example orders, tested with twenty or so real questions. Only then do more programs follow and, if needed, actions with confirmation.
The limits
What to know before you start.
It can misunderstand
An assistant can get a question wrong. That is why it says where each piece of data comes from, and the answer can be checked.
It only does what you allow
Every tool has its own permissions. It starts read-only, and actions that change data ask for confirmation.
A young standard
MCP appeared at the end of 2024 and is still changing. The server stays simple, so it can be updated without starting over.
The model has to be chosen
The assistant can use a cloud model or one running in house. Cost, speed and where the questions travel all change.
Frequently asked
Questions on this topic.
What is MCP, in plain words?
Model Context Protocol: an open standard, introduced at the end of 2024, that defines how an AI assistant can use the data and functions of other programs. The program offers a few tools, such as looking up an order, and the assistant uses them when it needs them to answer.
Does MCP replace APIs?
No. The MCP server usually uses the APIs of the management system or the shop. APIs connect programs with calls decided by whoever writes the code. MCP makes the same functions usable by an assistant, which chooses which ones to call.
Can an assistant do damage in the management system?
It can only do what you allow. It starts read-only. Actions that change data are added one at a time, with a person confirming, and every request stays in a log.
Does the data end up with outside services?
The MCP server runs on your servers. Where questions and answers travel depends on the model: a cloud service, or an open model running in house. It is chosen together, case by case.
Does it work with the AI assistant we already use?
MCP is supported by the main AI assistants and by several tools that run locally. If yours supports it, the same server connects to it. If not, a suitable assistant is chosen.
How is it different from RAG?
RAG finds answers in documents: procedures, manuals, contracts. MCP works on the live data of your programs: orders, stock, shipments. They are often used together.
Where do we start?
With one program and real questions, read-only. We measure how many answers are right, and only then extend.
Start with the problem
Do you open three programs to answer one question?
Tell us which ones, and which questions come up most often. We will tell you whether an assistant connected through MCP pays off, or whether a report will do.
- The first check is free and commits you to nothing
- Every message gets read, there is no call centre
- Fixed price and payment by milestone
- Italian and English, VAT invoices
- No phone call unless you want one
- If nothing should be built, we will say so
Prefer to write directly? [email protected]
Message received.