AI Auto-Reply#
The AI Community Manager module handles incoming comments and direct messages the instant they arrive, across every connected platform, sentiment analysis included.
Processing starts the moment the platform's webhook fires, not on a polling schedule: a background task runs right after the message is saved, without making the platform wait for it. A safety-net check (every minute) only catches whatever that instant task missed.
[Comment or message received via webhook]
↓
[Background task fires instantly, doesn't block the platform's response]
↓
[Keyword rule check] → match found → send instantly (skips the AI, bot hours still apply)
↓ no match
[Conversation context and account rules are loaded]
↓
[Sentiment and intent are classified, escalation keywords checked]
↓
[AI reply generated: GPT-OSS 20B primary, DeepSeek fallback]
↓
AUTO Mode → sent after bot hours + configured delay (safety-net cron sends it if the instant path missed the window)
REVIEW Mode → draft saved + notification sent (email, and Telegram if Turbo is linked), pending approval in InboxEvery setting below is scoped to the individual project (Settings → AI). A workspace with several brands configures each one independently: different mode, different hours, different tone.
auto_reply_commentsbooleanFor comments: true sends automatically (AUTO), false holds a draft for manual validation (REVIEW). There's no separate on/off switch : this one field doubles as both.
auto_reply_dmsbooleanSame true/false AUTO vs REVIEW behavior as above, scoped to direct messages instead of comments.
reply_delay_minutes1 | 2 | 3 | 5 | 10Deliberate pacing delay before automatic sending (AUTO mode only), so replies feel less robotic. Keyword rule matches skip this delay.
bot_hours_start / bot_hours_end0–23 (heure locale du projet)Active timeframe for the bot. Outside this range: messages (and keyword-rule replies) wait until hours resume.
Beyond generic AI replies, a project can define exact-response rules: "if a comment contains this phrase, send this canned reply". Useful for a call to action like "write 'I'm interested' in the comments" answered with a link every time, word for word, with zero AI cost.
Matching is substring-based (“contains”), not exact: a rule on je suis intéresséstill fires on “Bonjour, je suis intéressé !” or “je pense que je vais l'utiliser, donc je suis intéressé”.
A rule match sends instantly regardless of the project's auto/review mode (the merchant wrote the exact response, higher confidence than a generic AI draft), but bot hours and an active per-conversation pause still apply.
Configured in Settings → AI → Keyword auto-reply rules.
From any open conversation (mobile or desktop), the bell icon in the header pauses auto-reply for that one thread only: for 1 hour, 24 hours, or indefinitely until resumed. Useful when a specific customer needs a human touch without turning off auto-reply for everyone else.
In REVIEW mode, the AI draft is ready the instant the message arrives, not only when the Inbox page happens to be opened. A notification goes out right away: by email always (works for every account), and additionally by Telegram if the account has an active Turbo session. A bot can only message someone who has started a conversation with it first, so email is the one channel guaranteed to reach every user.
Setting up instant auto-reply, keyword rules and notifications
Was this documentation helpful?