Using integrations from chat
Ask for work across your CRM, databases and email in plain language. How to phrase requests, how to check what the AI did, and how to stay in control when it can change things.
Once your workspace has integrations connected, you use them simply by asking. There is no menu to open and no tool to select: describe the outcome you want, and the model works out which connections to call, in what order, and what to do with the results. One sentence can search your CRM, query a database and send an email.
And this is what landed in the inbox a moment later, sent through the workspace’s Postmark connection from the default sender set on it:
How to ask
Say what you want, where the data lives, and what to do with the answer.
| Good request | Why it works |
|---|---|
| “Find the deals in Pipedrive at Proposal Made and email a summary with the contacts to sales@northwind.example” | Names the system, the filter, the output and the recipient. |
| “Top 5 customers by order value from our retail database, with their share of revenue. Table plus one takeaway.” | Names the database, the measure, the size of the answer and its format. |
| “Who is our contact at Hawthorne Goods, and what open deals do they have?” | One account, two clear questions. |
| “Draft an email to Maria Lopez about her refund. Don’t send it yet.” | Makes the review step explicit. |
| “Make it a PDF and send it to Sarah.” | Colleagues are found by name in your workspace; the file is made and attached. |
A few habits make answers sharper:
- Name the source when you have more than one: “from the retail database”, “in Pipedrive”.
- Give the filter in your own terms: the stage name, the date range, the region. The AI turns it into the right query or search.
- Say the format: a table, three bullets, one paragraph, an email.
- Continue the conversation. “Now only the ones over 10,000”, “add their phone numbers”, “send that to Anna too” all build on what the AI already found.
Check what the AI did
Every reply that used an integration starts with an actions line, such as 17 actions. Expand it to see each step in order: which connection it listed, which tables it read, which search or query it ran, and the reasoning between steps. Expand a single call to see its details, such as the SQL behind a database answer.
The calls use the same tools as the MCP server, so list_database_connections, run_database_query, search_crm_deals or send_email mean the same thing in a chat, an agent and your own code.
Also read the prose around the table. When the AI had to make an assumption - for example that a CRM stage number means “Proposal Made”, or that a contact was found by company name - it tells you, so you can correct it.
When nothing is connected
If you ask for something that needs an integration your workspace does not have yet, the AI tells you and offers a button to connect one, such as Connect a CRM. Connect it, then ask again.
Adding a workspace connection needs a workspace owner or admin. If you are neither, ask one to add it, or add a Personal connection for yourself. See How integrations work.
Stay in control of changes
Reading is safe. Writing is where care matters, because the AI acts as soon as the request is clear.
- Sending email is immediate. “Email X to Y” sends. When wording matters, ask for a draft first, read it, then say “send it”.
- CRM changes need write access. Without Allow write access on the connection, requests to add or edit contacts or create deals are refused. See CRM.
- Database writes need a higher level. A Read only database refuses anything but a read. See Databases.
- Be specific about what to change. Name the exact contact, deal or row, and read the actions afterwards to confirm what happened.
Beyond chat
The same requests work in other places:
- Agents use your connections as part of their job. See Agents.
- Workflows run a request like “email last week’s new deals to the sales team” on a schedule. See Workflows.
- Your own code can call the same connections through the Integrations API or the MCP server.