Email channel guide
AI assistant
The ✨ icon at the top right of each conversation gives you four helpers:
- Summarise conversation: summary, key points and what the customer expects. Handy for long threads or when you pick up a colleague's conversation.
- Draft a reply: a draft for you to review and send yourself. For email it opens in the formatted editor. The AI only knows what is in the conversation: it does not invent prices, dates or availability; if something is missing, it marks it for you to fill in.
- Security check: the AI reviews the last message received and tells you whether it looks legitimate, suspicious or a scam, why, and what to do. The result is saved; you can ask it to analyse again.
- Translate message: translates the last message received (with its subject, for email) into the language you choose. The search accepts any language: type “French”, “Japanese” or its name in that language. Your panel language is selected by default and it remembers the last one you chose. It tells you which language it translated from, and if you ask for the same language again you see the saved translation without using credits.
The assistant only runs when you click it, using models on European Union servers. It can make mistakes: always review what it suggests.
AI replies
Assign an agent on the channel page, under Configure AI → AI agent. If you do not have one yet, the “Create a new AI agent” link creates it and links it to the channel automatically.
In email the AI is more cautious than in a chat, because an email reply is in writing: if the request is ambiguous, sensitive or it cannot resolve it safely, it does not improvise. It replies that someone from the team will follow up and the conversation stays in the inbox for you.
Does it reply by itself or prepare a draft?
On the channel page, in Configure mailbox → When the AI replies, you choose what the AI does with its reply:
- Prepare a draft to review (default, recommended): the AI writes the reply as it would do automatically, but does not send it. It stays as a draft inside the conversation for a person to review.
- Reply automatically: the AI sends the reply by itself. Turn it on once you trust how your agent answers.
Start with the draft and, once you find you almost always send it as it is, switch to automatic.
It is a setting of each mailbox, not of the agent: the same AI agent can handle several mailboxes or channels, each with its own mode. It only exists on email channels and only takes effect if the mailbox has an AI agent assigned (on “Humans only” there is nothing to review).
How the draft works:
- An email arrives and the AI writes the reply (same prompt, same spam and advertising filters and the same rules as always).
- The conversation is tagged “AI draft” in the list and the bell alerts you (and Telegram, if connected).
- When you open it, above the writing box you see the draft with three buttons: “Send as it is”, “Edit and send” (it opens in the formatted editor so you can change it) and “Discard”.
- When sent, it goes out as a normal reply in the thread, with your signature and a copy in your mailbox “Sent”. The alert disappears from the bell and from Telegram.
- From Telegram you get what the customer wrote and the proposed reply, and you tap “✅ Send” or “🗑 Discard”; “✏️ Edit in the panel” opens the conversation. Once decided, the alert is updated for everyone who received it. If your Telegram is in “alerts only” mode, no texts are shown: you can only open the panel.
- If the customer writes again before you decide, the draft is rewritten with the new message: only one is ever pending. If you reply by hand, the draft disappears.
- Tools: in this mode the agent only uses the tools you allow. By default they are: checking free slots, services and appointments; creating tickets (a person works them anyway and the customer gets their number); and booking appointments with your approval: if the appointment type “requires approval”, the AI leaves the request pending and alerts you (bell and Telegram) so you approve or reject it; if it does not, it offers the slot in the draft and you book it. The others —cancelling or changing appointments, GDPR rights…— are not used: it proposes them in the draft and you do them, so nothing changes in your system without a person seeing it. You can adjust it tool by tool in Configure mailbox → When the AI replies → Agent tools in draft mode (for example, letting the AI book directly). With “Reply automatically” the agent uses all its tools as usual.
- Credits and privacy: the draft uses AI credits just like an automatic reply. It is stored encrypted; resolved drafts are deleted after 30 days, ones nobody attends after 14, and when a contact is erased under GDPR their drafts are deleted too.
- Who can decide: owners, administrators and supervisors of the workspace, and whoever has the conversation assigned. The first one to decide wins.
When the AI does not reply (to avoid loops and annoyance)
The email always reaches the inbox, but the AI does not reply in these cases:
| Case | Why |
|---|---|
| Automatic replies (“out of office”, “received”) | Two auto-responders would answer each other forever |
| Delivery failure notices (MAILER-DAEMON) | There is nobody on the other side |
| Newsletters, mailing lists and bulk emails with an unsubscribe link or footer | They are not conversations |
| no-reply@ senders | Nobody would read the reply |
| Emails you send yourself from the mailbox | So it does not reply to itself |
| Your mailbox is only in copy (CC) | If “Reply only when in To” is on |
| Many emails in a row from the same sender | Safety limit: at most a few automatic replies per sender per hour |
| Advertising, spam or platform notices (if the filter is on) | They are not customer enquiries |
| Emails flagged as possible phishing | Replying to a scammer only confirms your mailbox exists |
| Unknown sender (if you choose “known contacts only”) or in your “Never reply” list | You decided so |
You decide who it replies to
In the mailbox settings, section “When the AI replies”:
- Reply with AI to all senders, or only to known contacts. Known contacts are those already in your CRM through another route (WhatsApp, form, manual entry…) or those you authorise. The safest option if you do not want the AI to reply to strangers.
- Advertising and automatic notice filter (on by default). Before waking the agent, a fast model checks the email: if it is advertising, spam or a platform notice (invoices, orders, alerts), it is not answered; if it is unclear, the bell notifies you so you decide.
- Your own rules for the filter, in your own words. For example: “Emails from suppliers must also be answered. Ignore job applications and CVs”. They take priority over the general criteria.
- Always reply to and Never reply to: lists of emails or domains (one per line, for example
[email protected]or@company.com).
When the AI does not reply to an email, the conversation shows the reason and a “Reply with AI” button so it answers anyway. If the sender was not a known contact, you can also click “Authorise and reply with AI”: from then on the AI will always reply to them.
The email filter: what it does with each type and what counts as pending
The filter is a small, fast AI that looks at each email before your AI agent sees it. It decides whether the agent should answer it. It only works on mailboxes with an AI assigned, and it does not use keyword or language lists: it judges by the meaning of the email. It puts each email in one of four categories:
| Category | What it is | What happens |
|---|---|---|
| Customer enquiry | A real person who asks, requests or writes about something your business handles (customer, prospect, supplier writing personally). | It goes to your AI agent. |
| Advertising or spam | Newsletters, offers, SEO agencies or cold sales that write to everyone. | The AI does not answer. If the sender only sends you this kind of email, it moves to the “Automatic” folder. |
| System notification | Automatic notices: confirmations, bounces, “no-reply”. | The AI does not answer either. Like advertising, it goes to “Automatic” if the sender only sends notices. |
| Unclear | It cannot tell whether a reply is needed, or it is a sensitive or legal matter where an automatic reply would be risky. | The AI does not answer. It stays unread and the bell rings so a person decides. |
What counts as “unread” (the number in the side menu and in the list):
- AI in “Reply automatically”: an email the AI handles does not count, because there is nothing for you to do.
- AI in “Prepare a draft for review”: it does count, because you have a draft to review and send.
- “Unclear”: it does count, and the bell rings. If the AI does not handle it, nobody will answer it but you.
- Advertising and notices that go to “Automatic”: they do not count in the number and do not ring. They stay unread inside that folder (Contacts → Automatic), in case you want to check them. If a sender ends up there by mistake, move it to Contacts and its emails will count again.
- Mailbox without AI (“Humans only”): every email counts.
If the filter fails or the account has no credit balance, the email goes to the AI as a customer enquiry, so you do not lose customers because of a technical failure.
Security: who is really writing
In an email, the sender can be forged. That is why Agent8Ai checks the verification your mail server performs when it receives the email (DKIM, SPF or DMARC). If the sender is not verified, the AI can give information (opening hours, services, availability), but to look up, cancel or change appointments it asks for a code sent to their real email. That way nobody can act on someone else's behalf.
Alerts (bell)
The panel bell alerts you when an email needs a person: when the conversation is assigned to someone, when the AI is paused, or when there is no AI and nobody is handling it. If the AI is replying, it does not alert you, so you are not flooded with notifications. If a customer sends several emails in a row, you see a single alert with the number of messages. Emails flagged as phishing do not trigger alerts.
When the AI leaves a draft to review, the “AI draft” alert stays on the bell (and on Telegram) until someone sends it, discards it or replies by hand.
Alerts arrive instantly, without reloading the page. The bell shakes when a new one comes in and the tab title shows how many you have unread, for example “(2)”. If the panel loses its connection for a while, the bell catches up by itself when you come back to the tab, the network returns or the sidebar counters change. If your session expires, the panel takes you to sign in and, once you are in, the bell shakes if you have alerts you had not seen yet. If you read an alert in one tab, the others update.