AI connected to your systems. On a separate account.
Claude, ChatGPT or your own agent connected to your CRM, Drive, spreadsheets and database through MCP or an API. The assistant answers from your data instead of general knowledge. The condition: it gets only the permissions it needs, and everything is logged.
One sentence that explains everything
AI sees only what the account it runs on sees. And it can do everything that account can do.
An MCP server or an API connector does not add any permissions. It uses the ones held by the account used to connect. If an employee connects an assistant with their own login, the assistant sees everything that employee sees. If an administrator does it, the assistant sees everything. If the account can delete, the assistant can delete, and it will not hesitate the way a person would.
Hence the order: first a separate account for AI, scoped to one task, read-only wherever reading is enough, with sensitive fields excluded. Then the connection. Then the log and a review. Only after that, the conversation about how clever the model is.
What I connect
CRM
Salesforce through its MCP server or API, HubSpot and other systems with an API. For Salesforce: a separate user with a permission set built for the task. A separate site about AI in Salesforce.
Drive and documents
Google Drive, SharePoint, OneDrive. Named folders, not the whole drive. The assistant reads procedures and proposals, not the payroll.
Spreadsheets
Google Sheets and Excel as the source for “how many”, “who” and “when” questions. Protected against the assistant changing cells.
Databases
A read-only database user, views instead of tables, limits on rows and query time. The assistant writes the query; the database enforces what is allowed.
Email and calendar
Microsoft 365 and Google Workspace: search, summaries, preparation for a meeting. Sending and calendar changes only with confirmation.
A custom MCP server
When your system has no ready-made server, I write one: tools matched to your tasks, with a log and scope control. I run my own server with around 75 tools every day.
Skills and instructions
Sets of instructions that teach the assistant your offer, your customers and your tone. Written down, versioned, editable without a developer.
Access management
Who sees and can do what: accounts, scopes, tokens, expiry dates. A quarterly review, a list of connected tools, unused ones switched off.
How it goes
The task and the questions
Which questions the assistant should answer, and for whom. From that follows what it needs access to, and what it does not.
An account for AI
A separate account in each system. Read-only where that is enough. Sensitive fields excluded. An expiry date.
Connection and tests
MCP or API. Tests with awkward prompts: “delete the old records”, “show me the ID numbers”, “change the price”. The assistant should refuse because the account does not allow it.
Go-live for a few people
First two or three people who know the data and will notice a mistake. We read the log together after two weeks.
Review and care
The log, tokens, new versions of models and servers. Once a quarter a list: what is connected, who uses it, what to switch off.
MCP, API or an integration platform
MCP when an assistant (Claude, ChatGPT) is to answer people’s questions from data and use several systems at once. A standard supported by a growing number of tools; Salesforce and Google run their own servers.
An API and custom code when the connection has to run without a person (a report, an automation), or when no MCP server exists for the system.
An integration platform (Workato, Make, n8n) when there are several connections, the logic changes, and someone on your side wants to edit it without a developer.
I choose for the task and for what you already have. Day to day I work with MCP and OAuth 2.1, the Salesforce and Google APIs, and Workato.
Questions about connecting AI
What is MCP?
A standard that lets an AI assistant use tools: read a record in the CRM, search a drive, run a database query. The MCP server exposes the tools, the assistant calls them, and the system enforces the permissions of the account the server runs on.
Our staff are already connecting their own ChatGPT accounts to our systems.
That is the most common starting point. I replace the personal connections with a separate AI account on a limited scope, switch on admin approval for new connections, and make a list of what is already connected. More in the article on the read-only account.
Can AI write data too, not only read it?
It can, but only once reading works flawlessly and writing is needed for the task. Writing starts with a person confirming and with a narrow scope: one field, one record type.
What goes in the log?
Who asked, which tool was called, with what parameters, what came back (size, not content, if the data is sensitive), and when. The log lives with you and we read it together once a month.
Does this need a developer permanently?
To set up, yes. For daily use, no. For care: a few hours a month or a quarter, depending on how many connections there are. I leave documentation so the next person does not start from zero.
We are on Salesforce.
Then two things: the separate site about AI in Salesforce, and the free permissions audit, worth doing before you connect anything.
On Salesforce and planning AI in it? First find out who can do everything in your org.
A free audit of permissions, access and AI settings in Salesforce. Report within a week, no access to your system needed.
Describe what you want automated
One sentence about the task is enough to start. You do not need to know the names of any tools. I reply within one business day.