Connecting your business records to an AI assistant

SynaptaGrid8 min read

#ai#mcp

In short

To connect business records to an AI assistant such as Claude, the safest common route is MCP, the Model Context Protocol: an open standard that lets an assistant call a defined set of tools in another system. Compared with pasting data into a chat or giving the assistant a database login, MCP keeps the assistant inside the system’s own rules. It acts as the signed-in person, so it can do only what that person is allowed to do, and the actions can be recorded. Decide first whether the assistant may only read or may also change records, who may connect it, and which actions should still need a person to click or approve.

Five ways to put records in front of an assistant

Teams try several routes, usually in this order. They differ most in how much of the system’s own control survives.

RouteHow it worksWhat goes wrong
Paste into the chatCopy rows into the conversationThe data is a stale copy, leaves your control, and the assistant cannot act on it
Upload an exportExport a CSV and attach itSame copy problem, and every new question needs a new export
A database loginGive the assistant credentials and let it write queriesFull access to everything the login can reach, no per-person rules, and a mistaken query can change data
A custom integrationBuild an API integration for the assistantWorks, but you build and maintain the permissions, the logging and the tool descriptions
A connector over MCPThe system offers a defined set of tools the assistant can callYou still decide who may connect and which actions need approval, but the system’s own rules apply

What an assistant can do through a connector

The assistant calls a defined tool, the system checks what the signed-in person may do, and only then does the action happen.

Questions to settle before you connect

Read only, or read and change?
Many teams start with questions only. Letting the assistant create or change records is a separate decision.
Whose permissions apply?
The assistant should act as the person who signed in, not as a shared account with wide access.
What is recorded?
Changes made through an assistant should leave the same record as changes made by hand.
Which actions need a person?
Deleting records or changing a structure are the actions to put behind a human click or an approval.
Who may connect?
Connecting an assistant to a workspace is an access decision, so decide who may do it.

The old way and the new way

The table compares feeding an assistant exports and screenshots with a connector that works on the live records.

The data the assistant sees
The old way: A copy pasted into a chat, out of date by the next day
The new way: The live records, read when the question is asked
Answering a question
The old way: Export, upload, ask, repeat
The new way: Ask; the assistant reads the records and counts them
Access
The old way: Whatever the shared login can reach
The new way: What the signed-in person is allowed to do
Making a change
The old way: The assistant writes a suggestion and a person types it in
The new way: The assistant performs the action as the person, and it is recorded
Control
The old way: None once the data is in the chat
The new way: The system’s permissions and logging still apply

How we decided

SynaptaGrid offers a connector over MCP for Data and Automation. The assistant works with the signed-in person’s own permissions, and its changes go to the audit log.

We ran it live in November, as the owner of our QA workspace, using Claude in a browser. Asked how many supplier records existed by review status, the assistant found the record type, read the records and counted four, matching the app. Given a five-row spreadsheet, it created a record type with typed fields and dropdown choices, published it and added five records in one conversation. Asked what was waiting for approval, it listed the person’s open task and recorded a decision with a reason code, and the run moved to Published. We deleted the test data afterwards through the same connector.

Four limits came out of that run, and we state them here. We ran it in Claude only, not in ChatGPT. The assistant could not find a workspace by its display name reliably, so it needed help to find the right one; we have filed that as a defect. Claude asked to be allowed to make each tool call in its manual mode, which is a setting of the client and not of ours. And we did not see a SynaptaGrid-side approval step for deletes, so this guide does not claim one: treat deleting records through an assistant as an action to confirm yourself.

We chose to run the assistant as the person, not as a service account, because it keeps the existing permissions meaningful. The cost is that an assistant can never do more than the person who connected it can.

Standards and accepted practice behind this

These are the practices the guide rests on. We align with them. This page does not claim a certification.

Model Context Protocol
An open standard for connecting an assistant to tools and data in other systems.
Source: Model Context Protocol specification (opens in a new tab)
Least privilege
The assistant gets the access of the person using it and no more.
Source: NIST glossary: least privilege (opens in a new tab)
Human oversight of consequential actions
A person confirms actions that are hard to undo.
Audit trail
Actions taken through an assistant are recorded the same way as actions taken by hand.
Source: NIST glossary: audit trail (opens in a new tab)

Where SynaptaGrid fits

SynaptaGrid Data and Automation can be connected to an AI assistant through MCP. You can ask questions about your records, create a record type and its records from a spreadsheet, and list and decide your own approval tasks, all with your own permissions.

Check the limits above before you rely on it: the client you use, how it finds your workspace, and what you want to confirm yourself before the assistant deletes anything.

Frequently asked questions

How do I connect a database to Claude?
The safest common route is a connector over MCP: the system offers a defined set of tools, and Claude calls them as the signed-in person. Avoid giving the assistant a database login, which has no per-person rules.
What is an MCP server for a database?
It is a service that exposes a defined set of actions on your data, such as listing records or creating one, to an assistant over the Model Context Protocol, so the assistant uses those actions instead of raw database access.
Is it safe to let an AI assistant change my records?
It depends on the controls. It is safer when the assistant acts as the signed-in person with that person’s permissions, when its changes are recorded, and when you confirm hard-to-undo actions such as deletes yourself.
Can an assistant approve a task for me?
In our live run, yes: the assistant listed the signed-in person’s open task and recorded their decision with a reason code, and the run moved on. The decision is recorded under the person.
Share

SynaptaGrid

See your records through an assistant

Connect an assistant to a test workspace and ask it a question about your own records.