AI Reply Agents
A reply agent is a named profile — a tone, your company context, and a knowledge base — that a campaign can point at so AI-drafted replies (see Inbox) sound like you and know what you actually offer, instead of a single generic global voice.
Configuration
- Tone — Professional, Friendly, Casual, or Formal.
- Company name / domain
- Additional info — free-text context always included in drafting.
- Autonomous replies and follow-ups toggles.
- Draft on arrival — write a ready-to-approve draft for every eligible new reply without sending, even with autonomous replies off (see below).
- Reply policy — what the agent does for each kind of reply (see below).
- Reply send window — the hours an autonomous reply may go out (see below).
- Confidence gate — a minimum AI confidence below which an auto-send is held as a draft (see below).
Reply policy
By default an agent with autonomous replies on sends a reply to the “safe” intents (Interested, Question, Informational, Conversation, Unclassified) and leaves the rest for a human. A reply policy lets you decide per intent instead. For each reply type you pick one action:
- Auto-send — compose and send with no human in the loop.
- Draft — compose and park it for you to review and send. You can set a timeout: if you don’t action the draft within N hours it either auto-sends, archives (drops out of the waiting view, nothing sent), or keeps waiting — your choice, per intent.
- Schedule (out-of-office only) — read the “back on 27 June” date from the auto-responder and schedule the reply for that day. If no date can be read, it falls back to a Draft for you.
- Referral (wrong-person only) — send a short, polite “who’s the right person?” note. If the reply names a referred email, that contact is also captured as a new lead (deduplicated by email) so the right person isn’t lost.
- Ignore — do nothing; leave it in the inbox for you to read.
Some actions are deliberately not offered, because getting them wrong damages your sending reputation:
- Unsubscribe replies are only ever ignored — an automated reply to an opt-out is both a legal risk and a complaint magnet.
- Not interested cannot auto-send — a machine arguing with a “no” earns spam complaints. You can still draft it for a person, or ignore it.
- Out of office cannot auto-send now (that is how a mail loop starts) — it can only be scheduled for the return date or drafted.
Leaving the policy unset keeps the original behavior exactly, so nothing changes until you configure it.
Draft on arrival
Turn this on and the agent prepares a draft reply for every eligible new conversation and parks it for approval — you open the inbox and the answer is already written, you just press Send (or edit first). It works even with autonomous replies off, so you get the speed of AI drafting without anything ever leaving without you. Drafting spends one AI credit per reply, so it is off by default.
Reply send window
Optionally restrict the hours an auto-sent reply may go out — a start hour, an end hour (0–23), and a timezone — so the agent never mails a prospect at 3am. Outside the window an auto-send is held until the window reopens. Drafts and scheduled out-of-office replies are unaffected; leave the window unset to send any time.
Confidence gate
Set a minimum confidence (0–100) and any reply the agent isn’t sure enough about is held as a draft for you instead of being auto-sent — so a shaky answer never goes out on its own. It applies to auto-send and referral replies; drafts, scheduled, and ignored replies are unaffected. Leave it off (the default) to send whatever the policy allows.
Drafts waiting
When drafts are waiting for your approval, Sendbox can notify your integrations with a recurring
agent.digest webhook (see Webhooks). There is no email digest —
watch the inbox, or subscribe a webhook to be told when drafts need you.
Knowledge base
Train an agent on:
- Documents — PDF, DOCX, or TXT, up to 10MB each. Extracted, chunked, and embedded for retrieval automatically.
- A website — point it at a domain, review the discovered pages, and confirm which ones to actually index. Re-indexing replaces the previous page set rather than appending to it.
Use Search knowledge to preview exactly what the AI drafter would retrieve for a given question — useful for confirming the knowledge base actually covers a topic before relying on it.
Assigning to a campaign
A campaign with no reply agent still drafts (politely, without inventing specifics); assigning one gives the AI real, grounded context to draw on.