Connecting an AI assistant to your CRM: what to ask
Ask what the assistant becomes once it is connected. If a key makes it a full user, everything it does carries your authority. In Fonsent a request made with an API key or through the MCP server is an agent’s request: the same rules apply, it can’t approve a draft, and every change is in the audit log.
Updated · 2 min read · Fonsent
In early access today. The AI model isn’t switched on yet. Field updates come from built-in rules, which pick up only sentences that state a value outright. What works today
An assistant can now reach into a CRM directly, through an API or an MCP server. That is useful, and it moves a decision you used to make once, at sign-in, to every request. The key you hand over decides what the assistant is allowed to be.
Five things to ask
Do the first one for real, in a workspace made for testing.
- Connect an assistant with a key, then have it try to approve a draft.In Fonsent. A request made with an API key or through the MCP server is treated as an agent’s request. The same rules apply, and it can’t approve a draft.
- Ask what kinds of key exist, and how often one may call.In Fonsent. Fonsent has a REST API and an MCP server. Keys are read-only or editor, and each key is limited to 120 requests a minute.
- Have the assistant change a record, then look for the change.In Fonsent. The audit log lists every change in the workspace, by people, the agent, automations and API clients, with before and after values and undo status.
- Ask what happens when the assistant wants to send something to a customer.In Fonsent. Nothing customer-facing leaves unless a person approved that exact message, and it leaves at most once.
- Turn it round: ask what happens when the CRM calls an outside tool.In Fonsent. An admin can connect outside tools as connectors. A connector call that needs a person waits in For review.
Why the first answer matters most
If a key is simply you, an assistant holding it can do whatever you can, including approving its own work. The approval step is still on the screen, and it no longer means anything.
The safer design gives a key less than its owner has. It can read, and it can propose. It can make the small changes an agent is allowed to make. What it can’t do is supply the one thing the product asks a person for.
Where it stops
A read-only key still reads. Decide what the assistant should see before you decide what it may change, and give it the narrowest key that does the job.
The rules cover what reaches the workspace. What an assistant does with what it has read is decided by the assistant and whoever runs it, not by the CRM.
The same care applies in the other direction, when the CRM holds a way into another tool. The tokens for accounts you connect are encrypted before they are stored.
The security page lists what protects a workspace, and what Fonsent doesn’t claim: how Fonsent protects a workspace.