Knowledge base retrieval
Knowledge is split into passages and embedded; each reply uses the passages closest to the customer’s question.
Skip to contentAI replies
AI email replies, drafted from your own knowledge
AI email replies are drafts — or sent answers — written by a language model from your own knowledge and the conversation so far. In RelayRow each mailbox chooses how far the AI may go: Off, Draft (a person approves every reply) or Auto (it sends only above your confidence threshold and drafts otherwise).
Updated
Three reply modes, set per mailbox
| Mode | What happens | Good for |
|---|---|---|
| Off | No AI replies; triage and rules still run | Sensitive mailboxes, e.g. legal@ |
| Draft | The AI writes a draft; a person edits and sends | Most support and sales mailboxes |
| Auto | Sends when confidence is above your threshold, drafts otherwise | Proven, repetitive questions |
Replies grounded in your knowledge
The whole thread
The draft sees the conversation history, not just the last message.
Your voice per workspace
Each workspace sets its own reply persona, and each mailbox its own signature.
Writing assistant
Improve, fix, shorten, lengthen, change tone or translate any reply in the composer.
Thread summaries
Summary, key points, next step and sentiment for long conversations, on demand.
Follow-ups
Time-based rules can send a follow-up when a customer hasn’t answered for a set time.
Why draft first
A wrong autonomous reply costs more than the minutes it saves. RelayRow starts every mailbox from a safe default and lets you graduate the mailboxes — and, through automation rules, the kinds of message — where the AI has proven itself to Auto.
Automation rules can also call the AI reply action with their own instructions and a minimum confidence, so a rule like “refund requests in support@” can always stay a draft while “where is my invoice?” is answered on its own.
AI reply questions
Can AI reply to customer emails automatically?
Yes, if you allow it. Set a mailbox to Auto and RelayRow sends replies whose confidence is above your threshold; anything less certain stays a draft for a person to approve.
Where does the AI get its answers?
From the mailbox’s knowledge base (and any workspace-wide knowledge), plus the thread history. Retrieval picks the passages most relevant to the question.
Which model writes the replies?
Whichever you pin for the reply feature — for example a Claude, GPT, Gemini or open-weight model via OpenRouter — with automatic fallback to other providers.
Can a person always review before sending?
Yes. Draft mode never sends on its own; every reply waits for a person to approve, edit or discard it.
Related
- AI email triageAI email triage that scores spam and phishing, then labels every message with category, urgency, sentiment, language, intent and a one-line summary.
- Any AI modelRun inbox AI on OpenRouter, Anthropic, OpenAI, Gemini, Groq, DeepSeek, Mistral, xAI or any OpenAI-compatible API — per feature, metered at real token cost.
- Email automationAutomate a shared inbox without code: 20+ conditions, AI yes/no questions, follow-up timers and actions like assign, label, forward, Slack and webhooks.
🚀 Ready when you are
Run every inbox from one place.
Connect your first domain in minutes — no mail server, no IMAP, no glue code. AI and automations included.