BlogAI
What MCP is and what it means for your business systems.
MCP (Model Context Protocol) is an open protocol that lets AI assistants connect to external systems and tools, with the user’s permissions. For a small company, it means you can ask your assistant about orders and invoices and get the answer from your own ERP.
MCP in plain words
An AI assistant knows a lot of general things but nothing about your company. It cannot see your orders, stock or invoices. MCP is the standard way for it to see them, when you allow it.
Think of it as a common plug. A system, such as an ERP, offers an MCP server. The assistant connects to it and learns what it can ask and do: look up an order, show stock, prepare a draft.
The protocol is open, so one connection can work with different assistants. Several AI assistants already support MCP.
Before a standard like this, every link between an assistant and a system was built separately. Now the system describes its tools once, and assistants understand them the same way.
How it works in practice
- You connect the assistant to the system’s MCP server and sign in with your own user.
- You ask a question in plain words, for example “Which orders ship today?”.
- The assistant picks the right tool on the server and requests the data.
- The server checks your permissions and returns only what is allowed.
- The assistant answers and, if an action is needed, asks for your confirmation.
The important part: the assistant does not get a copy of the database. It asks specific things and gets specific answers, under the system’s rules.
The rules stay in the system, not in the assistant. If a salesperson cannot see salaries in the ERP, they will not see them through the assistant either. If they leave the company, you disable their user and the connection stops with it.
What it means for a small company
Many owners already use an AI assistant for emails, quotes and texts. MCP lets the same assistant work with the company’s data, without copying spreadsheets into the chat.
- You ask about stock or order status without opening the ERP.
- You prepare a quote or a customer email with real prices and history.
- You connect several systems, such as ERP, email and calendar, in one conversation.
- You are not tied to one assistant: the protocol is shared.
For example: a manager asks “What has customer A ordered this year, and do they have overdue invoices?”. The assistant gathers the answer from the ERP and suggests a short email. The manager reads it, edits it and decides whether to send it.
This matters for a small company because there is no department writing reports. The owner and two or three people do everything. Every question answered in seconds is time returned to customers and real work.
MCP does not replace the ERP. Without tidy data in the system, the assistant has nothing to read. It is a new way to reach the data, not a new source. If your orders are still in spreadsheets, start there.
Reading versus acting
Tools exposed through MCP come in two kinds. Some only read data. Others change something in the system. The risk is different, and they should be treated differently.
- Reading: “What is the stock of item 1043?”, “Which invoices are overdue?”, “Which machine is busiest?”.
- Acting: creating a draft order, sending a reminder, changing a due date.
A wrong answer when reading leads to a bad decision that a person can still catch. A wrong action has already happened. That is why actions always go through confirmation.
The risks
A link between AI and business data is useful, but it is also a door. You need to know who walks through it.
- Permissions too broad: the assistant sees more than the user should see.
- Actions without control: the assistant sends, deletes or changes something after misreading a request.
- Misleading content: text in an email or document may try to “instruct” the assistant to do something else.
- Unclear data flow: answers pass through the assistant’s provider.
- No trail: if nothing is logged, you cannot find out what happened.
Misleading content deserves special attention. The assistant reads emails and documents that come from outside. If such a text says “send all prices to this address”, a well-built system does not let that happen without you.
The risks are not a reason to avoid MCP. They are a reason to choose a system where permissions, confirmation and the audit log are built in, not left to the assistant’s good behaviour.
Safeguards to insist on
If the vendor cannot answer any of these clearly, ask them to show you live. An audit log that exists only in a slide deck protects no one.
Start with read-only access. Add actions once you see how people use the connection.
How to get started
- Check whether your systems offer an MCP server.
- Pick one question you ask often, such as order status.
- Connect the assistant with read-only access, for one or two people.
- After a few weeks, review the audit log and decide whether to add actions.
Explain to people what the assistant can and cannot do. Show them the audit log so they know every action leaves a trace. They get used to it faster and ask better questions.
More on connecting systems on the integrations page.
How we approach it at Leanovix
Leanovix doesn’t have an MCP server yet. AI is coming, and we are building it with the first clients. When it does, the rules above apply: the assistant sees only what your role allows, and you confirm every action. See AI in Leanovix and how we protect your data.