Trust
What the AI features send, and where
There is no MailDock inference service. That is not a policy we could change quietly — it is the architecture, and it is why the rest of this page is short.
The short version
| What | Where it goes |
|---|---|
| A thread you ask to summarise, explain or suggest replies to | To your provider, on your key: the messages, with their From, To, Date and Subject |
| A draft you ask to rewrite, translate or check | The draft, to your provider |
| A reply you ask AI to draft | The message you are replying to, and your display name, to your provider |
| Questions to the admin copilot | Delivery metadata it looks up (senders, recipients, remote replies), for domains with AI on only |
| Attachments | Excluded unless allowed per domain — and no feature reads one yet |
| Mail nobody has asked the AI about | Never sent by an inbox feature |
| Anything at all, with localOnly set | Never leaves the machine |
| Your provider key | Encrypted at rest, sent only to that provider |
| What the audit log records | Metadata — who, when, which feature. Not content. The copilot's tool log keeps the arguments it ran with (a domain, a recipient searched for), on your server |
Bring your own inference
Every AI feature runs against a provider you configure and pay: Anthropic, OpenAI, Gemini, xAI, Meta Llama or Groq. Your relationship with that provider — including their data retention and whether they train on your inputs — is between you and them, and their enterprise terms usually say no. We are not in the middle of it, and we could not be: there is no endpoint here to route through.
Or configure nothing, and the features are simply absent. Or point them at Ollama or vLLM on your own hardware, set localOnly, and no request leaves the machine at all.
Two switches, not one
An administrator sets a per-domain policy for what is available. A user opts in for their own mailbox. Neither alone is enough — an administrator cannot switch on AI for somebody else’s mailbox, and a user cannot enable something the domain’s policy forbids. The user’s switch is in webmail, underSettings → AI features, and it only appears when the policy above it allows AI at all.
What the admin copilot can do
Read-only tools, and nothing else. When it wants to change something it writes aproposal that you approve by id. The model never holds a write, so a prompt-injected instruction in an email cannot become an action on your server.
Smart replies are suggestions
Nothing composed by a model is ever sent without somebody pressing send. There is no auto-reply, no auto-triage that deletes, and no background classification of mail nobody asked about.
Analytics on this website
Separately from the product: this site loads Google Tag Manager only after you accept, and the consent banner’s decline is the same size as its accept. Until you choose, no request is made to Google at all — which is asserted by a test on every build, not just intended.
The legal version is the data processing addendum, which covers the bring-your-own-inference model explicitly.