AI Agents let an AI model work on your records, for example tickets. The agent reads the record, its emails and comments, can use tools you give it, and writes an internal note and a reply draft. A person always decides what is sent.
Set up an agent
Open AI Agents (Tools menu; an administrator activates the module in Module Manager) and create an agent:
- Name and Chat name: the chat name is what the agent is called in comments, for example "Claude support".
- Instructions: what the agent should do and how it should write. Tip: ask for informal, short replies and for "only state what you have checked".
- Provider and model: Anthropic, OpenAI or any OpenAI-compatible service. Empty uses the CRM's default AI settings. You use your own API key.
- Tools (optional): a JSON list of MCP servers the agent may use, for example
[{"mcp": "gitlab", "tools": ["search"]}, {"mcp": "simply"}].toolsis an optional allowlist. The address and key of each server are stored as global variablesglobal_mcp_<name>_urlandglobal_mcp_<name>_token, never on the agent. - Active: only active agents run.
On the record (e.g. a ticket) choose the agent in the AI agent field.
Run the agent
- Write a comment that starts with @ai followed by what you want, for example "@ai check the error log". Or set AI status to Start.
- The agent runs in the background and usually answers within a minute or two.
- Steer it by writing new comments: it always follows your latest instructions.
AI status
| Status | Meaning |
|---|---|
| Start | The agent will run. |
| Running | The agent is working. |
| Waiting for user | The agent needs an answer from you (see its note). |
| Waiting for approval | A tool needs a person's approval first, e.g. temporary access to a system. The agent starts again by itself after the decision. |
| Done | The agent has finished. |
| Error | Something went wrong; the reason is in the agent's comment. |
What the agent writes
- Internal note: shown in the comments with the agent's chat name, a robot icon and an "AI" label. Comments addressed to the agent show "to AI".
- Reply draft: in the Email tab you see "Agent has a reply draft ready". Click Show to read it, or Insert in email to put it in the email, edit and send as usual. When you send it, the draft is marked as used: a comment notes whether it was sent unchanged, edited or translated, and the draft disappears.
Translate comments
Every comment has Translate at the right end of its header line. It translates into your language on demand only, and Show original switches back.
Tools and screenshots
With tools, the agent can look things up itself (code, records, logs). Tools that return pictures, such as screenshots, are shown to the agent as images, so it can see what a page looks like.