Why the inbound AI classifier is deliberately not run on outbound

A reasonable question: if MxGuard has a powerful hybrid-AI spam filter, why doesn't it apply it to the mail you send out through the smarthost? The answer is that it would do more harm than good.

The models are trained on mail sent to you, not by you. The inbound classifier and the Claude layer are tuned to judge whether an incoming message is trying to manipulate the recipient — urgency, impersonation, unfamiliar senders. Your own outbound mail legitimately contains invoices, payment requests, urgent notices and messages to people you've never emailed before. Running the inbound models backwards over that traffic would flag your genuine business mail as suspicious.

So MxGuard applies the right tools for outbound instead: ClamAV virus scanning, per-account rate limiting, optional DKIM signing and full logging — measures that protect deliverability and contain abuse without second-guessing the content of your legitimate mail.

The result is a relay that improves delivery without introducing false positives on the mail you actually meant to send.