Use a role account
Before anything else: connect support@ or hello@, not you@. Connections are authorised per account rather than per task, so an employee
connected to your personal mailbox can reach all of it, forever, for every job it ever does.
A role mailbox also means replies come from an address someone else can pick up, which matters the first time you are away and a customer replies.
Authorise it
Install the mail plugin the employee needs, then ask the employee to do something requiring it. It will tell you what is missing and hand you the connection flow. Authorise there.
A password typed into a conversation is a password stored in a conversation. The connection flow exists so the credential never enters the transcript, and there is no case where pasting one is the right move.
Decide how mail arrives
Two ways, and the difference matters more than it looks.
| Trigger | Behaviour | Use when |
|---|---|---|
watch | The plugin streams changes as they land | Anything where response time matters |
schedule | Checks on a cron expression | Digests, or a mailbox with no streaming support |
A watch trigger picks up a message at 2am at 2am. A schedule picks it up at the next run, which is fine for a morning digest and wrong for a customer waiting on a reply.
Teach it your voice before it writes anything
Paste three real threads you handled well, with your actual replies. Then have it answer a fourth you already know the answer to, and compare. Examples beat description: three things you wrote will do more than a paragraph about your tone.
Put the durable rules in the persona, since they apply to every future reply:
Restating the question back, and em dashes. Both read as machine-written even when the rest is good. The em dash has a second problem: in a subject header it can arrive as mojibake in some mail clients, so it is worth banning rather than merely discouraging.
Set the gate before it can send
Sending mail is contacting someone outside your company, so it is gated by default. In the desktop app, open the employee's settings and find Approval controls.
- Approval while you are still reading its drafts. This is the default, and it is the right starting point.
- Always allow only once you have read enough drafts to trust them.
- Blocked for anything it should never do, such as sending on a thread that has escalated.
Check it
- Send yourself a message and confirm it is picked up
- Reply to its reply and confirm the thread holds together rather than starting a new one
- Check the subject line renders correctly, especially any punctuation
- Send something abusive and confirm it stays professional and escalates rather than engaging
When it goes wrong
Replies start a new thread each time
Threading breaks when a reply invents its own message identifier or rewrites the subject. The reply must carry the original subject unchanged and reference the original message id exactly as received. Test by replying twice.
The subject arrives as garbled characters
Non-ASCII in a header that was not encoded, usually an em dash or a curly quote. Ban em dashes in the persona, which fixes the common case.
It read the mailbox but did not reply
Often the employee judged the message not worth answering. If every message deserves a reply, say so as a hard rule rather than a preference, because prose guidance loses to the model's own judgment.