Smarter email security. Human control.
Two optional AI tools help you filter and tune: a classifier that adds one weighted signal to the spam score, and an advisor that proposes setting changes. Both are off until you turn them on.
Enterprise feature Off by default.
AI that asks first.
Default off
Nothing AI-related runs until an administrator enables it.
Administrator approval
In Overseer mode, every proposed change waits for a person to approve or reject it.
Your model, your network
Run inference on your own server. A hosted provider is strictly opt-in.
One signal. Not the verdict.
For each inbound message, the gateway asks a language model for a category, a threat score, a confidence score and a short reason. Only phishing, malware, scam and spam count as hits.
- Told explicitly that newsletters, invoices and ordinary outreach aren't spam
- Ignored below 60% confidence; counts only above a threat score of 75
- Learns from your quarantine releases, with no training or fine-tuning
- Always fails open: an outage can't stop your mail
| Setting | Default |
|---|---|
| Threat threshold | 75 |
| Minimum confidence | 60 |
| Weight on a hit | 4 (spam threshold 6) |
| Timeout | 8 seconds → no opinion |
Proposals, with guardrails.
The advisor reviews a summary of recent mail flow and your settings, then proposes bounded changes.
- Overseer (default)
Proposals are queued for an administrator. Nothing changes on its own.
- Autopilot (opt-in)
Applies only changes from a restricted catalogue (never credentials, routing or TLS) and never anything that weakens protection.
- Self-checking
At most five changes per run at 70%+ confidence. If good mail starts getting caught in the next six hours, the change rolls back automatically.
Choose where the model runs.
Self-hosted inference
Point the endpoint at any OpenAI-compatible server inside your network: a local inference server, a self-hosted gateway or a private deployment in your own cloud tenancy. No message content crosses your boundary.
Hosted provider (opt-in)
Use the Anthropic messages API or an OpenAI-compatible service. Per message, the sender, subject and up to 4,000 characters of body are sent. This is the one place anything leaves your infrastructure.
Full settings, data minimisation and a common configuration mistake are covered in the AI threat detection docs.
AI questions
Are the AI features turned on automatically?
No. The classifier and the tuning advisor are off by default. Even when you enable the advisor, it starts in Overseer mode, where every proposal waits for an administrator.
Which plan includes AI security?
Enterprise ($1,490/year). The rest of the filter stack doesn't depend on it. Community and Professional filtering work without any AI.
Does my email get sent to an AI company?
Only if you choose a hosted provider. You can point both AI features at a model inside your own network using any OpenAI-compatible endpoint, and then no message content leaves your network.
Can the AI quarantine mail by itself?
Not at default settings. A classifier hit adds a weight of 4 against a spam threshold of 6, so it needs another signal to agree.
Does AI guarantee accurate classification?
No. It's one calibrated signal among many, and it can be wrong. That's why it's weighted, fails open and stays under administrator control.
AI help, with a human in charge.
See Overseer and Autopilot in a demo, or talk to us about running inference on your own model.