This is the process we follow when a sales or support team wants to work with their CRM through an AI assistant.
1. Collect the real questions
Ask the team for the 20 things they look up or update most. "Which deals have not moved in two weeks?" is a good one. These questions become your test set.
2. Check for an existing server
Some CRM vendors publish their own MCP servers. Try it against your test set. If it answers well and the permissions are acceptable, you are done.
3. Design the tools
Group the questions into 5 to 15 tools. Name them in your team's language. Decide for each one: read only, write with confirmation, or not exposed at all.
4. Choose authentication
A service token is fine for a small team with the same access. If people have different permissions in the CRM, use OAuth so the AI acts as each user.
5. Build with limits
Validate every input, cap result sizes, cache common lookups and log every call. Map custom field ids to readable names.
6. Test with the team's questions
Run the test set in the AI clients your team uses. Fix tool names and descriptions until answers are right without special prompting.
7. Roll out and watch usage
Start with a few people. After two weeks, look at which tools are used and which questions fail, then iterate.
Want this done for you? See our connectors for HubSpot, Salesforce, Pipedrive and Zoho CRM.