Connect an AI assistant to your CRM safely
Connecting an assistant to a CRM is a permissions problem before it is an AI problem. Five steps, in order: a read-only task, a scoped key, approval before writes, a log and a measure after two weeks.
- 5 min
The CRM holds contacts, quotes and payment history, so the order of the steps matters more than the choice of assistant. The method below is the one AINS follows. Each step ends with something you can check.
Start with one read-only task
Pick a single task where the assistant only reads: show the open quotes for a contact, or summarise the last calls on a deal. Choose one that people repeat often enough to notice the saving, and where a wrong answer is easy to spot. Nothing in the CRM can change, so the worst case is a wrong answer that a person spots. This first task also shows the state of the data. Duplicate contacts, phone numbers in three formats and empty fields all surface here, and fixing them first makes every later step safer.
Give it a scoped key
Create a dedicated CRM user or API key for the assistant, with read access to the modules the task needs and nothing broader. Never use an administrator key. The key lives in the MCP server, outside the chat, and can be revoked without touching any person’s login. Write down which modules and fields the key can reach, so the next person can read the list in a minute. If a later task needs another module, add that module on purpose and write down why.
Put a person in front of every write
When the assistant starts to write, such as a call note or a pipeline stage, each change goes to a named person to approve first. Emails to clients stay behind approval for good. Some low-risk writes can run without approval later, once the team has watched the approvals and found nothing to correct. That decision belongs to the client, and it can be reversed at any time.
Log every call
Every call the assistant makes is written to a log in the client’s own account: the time, who asked, the action and, for a write, the old and the new value. The log answers what the assistant did on a given day, and it makes a wrong update easy to undo. Without it, the first mistake becomes a long investigation. The client keeps the log for as long as the assistant runs and can read it without asking AINS.
Measure after two weeks
Write the baseline before the start: how many minutes the task takes by hand and how often it comes up. After two weeks of real use, compare these four figures against it.
- Tasks the assistant completed
- Minutes per task, compared with the baseline
- Corrections people made
- Approvals that were rejected
Then keep the task, widen it or stop. A task that saves no time or needs constant correction is stopped, and that is a useful result.
Check before writes go live
Before the first write is switched on, confirm five things.
- The key reaches only the modules the task needs.
- A named approver exists and can be reached during working hours.
- A test call appears in the log from start to finish.
- Someone knows how to revoke the key and has done it once in a test.
- The baseline time is written down.
What to do next
The free 30-minute review picks the first read-only task with you and says what the key needs to reach. What an MCP server is explains the idea behind the server, and AI agent or fixed workflow helps when a task may not need an assistant at all. The scope and price come in writing within 48 hours of the review. See the n8n automation agency page, the services or book the review.