AI assistant and writing style

Availability

The AI assistant is in beta and included in every paid plan, including your 14-day trial. There is no per-user AI fee. Chat will stay included. Deeper capabilities we add later may be priced separately.

Start a chat

1. Open AI assistant from the dashboard sidebar.
2. Describe the result you want, including the mailbox or time window when useful.
3. Review source cards and the exact effect before confirming a sensitive or bulk action.

Open a conversation in the inbox, then choose Ask the AI assistant in the toolbar or mobile controls, or press I, to toggle the AI assistant sheet with that conversation attached. You can also press ⌘K and choose Ask the AI assistant. Any text you typed in the command palette fills the composer for you to review before sending.

On wide layouts, a 380px AI assistant column sits beside the email. On narrower layouts it floats, and on phones it fills the screen. Its context follows the conversation you select. Removing the context chip keeps the same chat without thread context until you open another conversation. Choose New chat to start another chat, or Open in AI assistant to continue the chat on the full AI assistant page. Sheet chats also appear in recent chats.

Writing in the composer

Pro, Team, and Scale plans include AI writing help in every message composer: the Rewrite button improves or shortens a draft. Select text to change only that passage, or leave the selection empty to work on the editable message.

Shipmail always shows an editable preview before Apply. Undo is available on the web and Restore original is available on mobile. Forwarded and quoted messages are outside the AI scope, and links, email addresses, phone numbers, dates, numbers, and template placeholders are restored locally rather than rewritten by the model.

If you have an active personal Writing Style, Rewrite uses its bounded instructions and two short approved examples while treating the current draft as the primary voice. Without a saved style, Shipmail preserves the voice already present in the message. Rewrite never learns silently from Sent mail, and shared inboxes use the composing user's personal style. You can manage that style from Rewrite in a modal without leaving or losing the draft.

Model requests use no tools, request no provider storage, and share the organization's included AI allowance with AI assistant chat. The mobile app shows its provider disclosure on first use and keeps that acknowledgement with your account.

Rate and copy answers

Use the thumbs under an AI assistant reply to rate it. A thumbs down lets you add an optional comment about what went wrong. Click your selected thumb again to clear the rating. Your rating stays with the answer when you reopen the chat.

Use Copy beside the thumbs to copy the answer text. For anything else, use Chat with us in the sidebar.

Steps and questions

While the AI assistant works, it lists each step under its answer, such as Looked at your folders or Searched mail from a sender. When the answer arrives, the steps fold into one line, such as 3 steps. Click it to see them again, also after you reopen the chat.

When a request could mean several things, the AI assistant asks with options, for example the closest folders and one that creates a new folder. Click an option, or click Other and type your answer. Draft and forward recipients and invitation attendees must be addresses you type yourself: picking an option never adds one.

After you pick, the AI assistant finishes the request with what it found before it asked. If you pick Create Support, for example, one card creates the Support folder and moves the messages it found into it. Folders are shared by all your mailboxes, so the card names no mailbox.

Asking what needs a reply

When you ask what needs a reply, the AI assistant reads conversations you chose to track with Track reply, after the other person wrote back, plus follow-ups that are now due. It does not treat ordinary inbound mail as work just because the sender looks like a person.

The check always covers at least the previous two weeks, even when you ask about today. A tracked conversation that is still waiting on you stays in the list until you reply, mark it done, or set it aside.

Ask about delivery

Ask about delivery rates, sent-message volume, recipient outcomes, bounce types, or the latest bounced, complained, deferred, and failed deliveries. Rates use known recipient outcomes and do not measure inbox placement.

For example: “What was our delivery rate this week?” or “Why did delivery to customer@example.com fail yesterday?”

Delivery answers contain organization-wide mail data. They are available only to Owners and Admins with access to all mailboxes. If your role or mailbox access changes, Shipmail hides earlier delivery answers and their cards.

Domain health and sending reputation

Ask whether a domain is set up correctly or why sending may have a problem. The AI assistant can show the stored MX, SPF, DKIM, and DMARC verification state, name a failing record, and link to the domain page with the expected DNS values.

Domain checks use Shipmail’s most recently stored verification result. They do not run a live DNS lookup, and the AI assistant cannot change DNS. Check the last-checked time and use the domain page when you need the current values or want to verify the domain again.

Reputation answers show the recorded sent, hard-bounce, and complaint counts behind each rate. Inbox placement also depends on content, recipient engagement, provider filtering, and other factors Shipmail cannot see. The AI assistant cannot promise that a DNS or reputation change will move mail into the inbox.

Access follows your current role and mailbox scope. Domain details and organization-wide reputation each appear only when you can view that part. If permission blocks one part, the AI assistant treats it as unavailable rather than as an empty or healthy result.

Webhook health and replays

Ask which webhook is failing or why an endpoint stopped receiving events. The AI assistant lists each endpoint by host and path, says whether it is enabled, and gives its failed and total deliveries for the last 7 days, deliveries waiting for a retry, and the HTTP status and recorded error of the latest failure. It never shows payloads, signing secrets, or URL query strings.

Ask the AI assistant to replay failed deliveries, for all endpoints or one you name. It lists the exact deliveries first, grouped by endpoint with their event types, and replays nothing until you confirm. One replay covers at most 20 failed deliveries from the last 7 days that were not replayed already. After it runs, the card shows each delivery's result, including any that could not be replayed. If an endpoint's URL or events change after you review the list, the AI assistant replays nothing to it and asks you to review it again. Deliveries you can no longer access when you confirm are counted, not described.

Webhook answers and replays are available to Owners and Admins, as on the Webhooks page. A delivery that carries mail from a mailbox you cannot read is not counted, listed, or replayed. An endpoint whose deliveries are all beyond your access shows that its history is hidden.

Conversation history

The AI assistant saves each chat in your conversation history. Pin chats you return to often, rename them, or archive finished work. Archived conversations stay readable and can be restored before you continue chatting. Chats are private to your user in the current workspace.

Drafts and confirmations

Review AI assistant drafts in the Inbox, edit them inline, then send them with the same undo-send window as any other reply. On a confirmation card, choose Cancel to decline the proposed change without applying it.

When the AI assistant drafts a reply, its card appears at the end of that Inbox conversation. The Draft ready pill also marks the conversation in AI assistant results and Inbox thread rows. Open the AI assistant drafts view from the Inbox to review them. Only the person who started the chat, or the owner of the automation that created the draft, can see it and edit it inline before sending. If that person's account is removed, the draft becomes an ordinary shared draft in the mailbox's Drafts folder.

Replies the AI assistant creates stay private to the person who started the chat, or the owner of the automation, as described above. Sending one uses the same undo-send window as any other inbox reply. Reversible single-message actions can run directly. Scope-changing, bulk, or higher-risk actions show the exact effect and require confirmation. Eligible bulk actions remain undoable for seven days.

Ask the AI assistant to delete or remove messages and it moves them to Trash, never deleting them permanently. Only chat offers this, and you always confirm the exact messages first. Undo moves them back, and you can also restore them from Trash yourself.

Undo restores only the part of each message that the AI assistant changed. For example, a later read or star change does not stop a folder move from being undone. If the action created its destination folder, Undo removes that folder after it restores the messages, but only while the folder is unchanged, empty, and not in use. An existing folder is never removed. The result card reports any messages that could not be restored, and Inbox counts refresh.

Writing Style

Open AI assistant > Writing Style and paste one or more replies that sound like you. Drafts are written to match your examples, plus any optional instructions you add. Each save creates an immutable version. Only text you paste is used. Shipmail never learns from Sent mail silently.

Drafts use your style right away. An internal safety evaluation protects automatic sending. Existing drafts and active automations stay pinned to the exact style version they were created with.

Security boundaries

• Email bodies, HTML, attachments, calendar content, and tool output are untrusted data.
• The AI assistant can act only on resources available to your current workspace membership.
• The model cannot invent tools, URLs, recipients, permissions, billing changes, credentials, or arbitrary code.
• Every side effect is typed, authorized again at execution time, idempotent, and audited.
• Shipmail signs AI assistant model requests with Web Bot Auth. Receiving services can verify Shipmail through its public Ed25519 key directory.