Many teams already use Claude or ChatGPT every day. For work that involves company data, though, the process is still manual. Staff copy data out of a system, paste it into a chat, then copy the result back. It is slow, and company data moves around with no record of where it went.
The Model Context Protocol (MCP) offers a cleaner route. MCP is an open standard introduced by Anthropic in 2024 for connecting AI assistants to tools and data. With MCP, you can connect Claude to company systems such as your CRM, order system or document storage, with controlled permissions and every access logged. This article covers what it takes, which systems to start with, and how to roll it out in stages.
How MCP works in brief
MCP has two sides. One is the client, the AI application people use, such as Claude. The other is the MCP server, a program you run in front of your company systems. The server exposes a set of tools, for example "find customer", "check order status" or "create draft quotation".
When a user asks something, the AI picks the relevant tool and calls it through the server. The server checks permissions, runs the request against your system and returns the result to the AI. Because MCP is an open standard, the same server can be used by Claude, ChatGPT and other assistants that support MCP. You do not need a separate integration for each AI app.
In practice, users simply ask in plain language. An admin at an event organiser, for example, could write "show decoration vendors whose contracts end next month". The AI calls the contract search tool, reads the result and writes a short answer. The admin does not need to know table names or system menus, and no data is copied out by hand.
What it takes to connect Claude to company systems
Technically, the requirements are modest. The business decisions around them matter more, especially who is allowed to do what.
- A system with an API or database that can be reached securely. Older systems without an API usually need a connecting layer first.
- A clear list of tools. Each tool should do one thing, with a name and description the AI can easily understand.
- A login method tied to each user's identity, usually OAuth.
- Permission rules per tool and per role.
- An audit log recording who called which tool, when, and with what result.
- A place to run the server, either in the cloud or on an internal server.
Which systems to connect first
Choose systems that people ask about often and that carry low risk. As a first step, read access is almost always safer than access that changes data.
Take a building materials distributor. The sales team often calls the warehouse just to ask about stock and delivery status. Read-only tools for stock and order status already save a lot of time with no risk of changing data. At a consulting firm, a good starting point might be searching contract documents and project history. Finance, payroll and anything that sends messages directly to customers can wait for a later stage.
One simple question helps you choose. Which tasks make staff message a colleague in another division every day just to get one number or one status? Those are usually the best candidates for the first tools, because the benefit is felt right away and is easy to measure.
Permissions, OAuth and audit
Signing in as the user
With OAuth, each user signs in with their own account. The MCP server then acts on that user's behalf, so a salesperson only sees customers in their own region, following the access rights already set in the system. Avoid a single shared admin account for everyone, because every access restriction disappears with it.
Permissions per tool
Decide which tools each role may use. Keep read tools and write tools separate. For actions with real consequences, such as changing prices, cancelling orders or emailing customers, add a human approval step before the action runs.
Audit log
Record every tool call with the user, the parameters and the result. The log helps when management has questions, when you trace an error, and when you need to show that personal data is handled in line with the principles of Indonesia's Personal Data Protection Law (UU PDP). Folders, Averoa's document product, follows this approach. Claude and ChatGPT get per-folder access through MCP, and every request is logged.
Rolling out in stages
A safe rollout usually goes through a few stages:
- Choose one small team and two or three read-only tools.
- Run it for a few weeks and read the audit log regularly. Look for questions that failed and tools that were called by mistake.
- Improve tool names and descriptions. The AI picks tools based on their descriptions, so clear wording makes a real difference.
- Add write tools one at a time, with human approval for risky actions.
- Extend to other teams once the permission rules and usage patterns are settled.
Prepare a short guide for users as well, explaining what they can ask, what is not supported yet, and where to report results that look wrong. One page is usually enough.
Next steps
MCP lets AI assistants work with company data without copying and pasting, as long as permissions and logging are designed in from the start. Averoa builds custom MCP servers with OAuth, per-tool permissions, audit logs and human approval steps through its MCP & AI Agents service.
