If your feed has any international reach at all, most of your comments arrive in a language you don't speak. A community manager either answers a fraction of them in English and ignores the rest, or hires four more people. An engine takes the third path: it detects the language, replies in it, in your voice, while the post is still warm.
Why "reply in every language" is a real feature, not marketing
Instagram's Graph API doesn't do this. Neither does any scheduler. What they can do is show you comments; what they can't do is respond in ways that feel like you, in a language you don't speak, within the tight window where a reply actually earns follows. That window is small · the first 30 minutes after a post drops. Reply then and the algorithm rewards you with more distribution. Reply the next morning and the comment might as well be in a museum.
For a persona feed like @ninna.artist that runs 24/7 with a global audience, the reply layer is not a nice-to-have. It's the reason the feed feels alive.
The four things a reply engine has to get right
1. Language detection
This is the easiest problem in the stack. A modern LLM detects language with near-perfect accuracy from as little as three words. Even mixed-language comments ("no way, incredible tbh") route correctly. OnBrand supports EN, ES, RU, FR, DE, PT and IT at launch · not because more is hard, but because those seven cover the overwhelming majority of comment volume on the accounts we've profiled.
2. In-voice reply
This is where most tools fail. Translating "thanks so much!" into six languages produces six versions of the same generic reply. That is worse than not replying, because it flags to the reader that they got the auto-response. A real reply engine takes the brand's voice rules (lowercase, offhand, one emoji max, present tense, no exclamation marks · whatever your rules are) and applies them in the target language. In Spanish that means dropping capitalization even where formal Spanish would keep it. In German it means resisting the language's structural formality. The output has to read as if a native speaker with your voice wrote it, not as a translation.
3. Spam, troll and sensitive-topic filtering
The single most important rule in the reply layer: the engine never auto-replies to something it shouldn't touch. Every incoming comment is classified before a reply is written. If it looks like spam (link bait, follower-farming, repeated boilerplate), it gets reply_status: 'flagged' and the reply is skipped. Same for obvious trolling, and same for anything the classifier flags as sensitive · mental health, politics, breaking news. Those bubble up to your queue with the flag reason. You handle them personally or hide them. The engine stays quiet.
4. Speed
Warm replies compound; cold ones don't. The scheduler runs every five minutes, so worst-case reply latency is around six minutes end-to-end (five for the tick, one for detection, classification, generation and the API call). That's inside the algorithm's warm window on every post we've measured.
What "in your voice" actually looks like
Here's a real example, verbatim from the seed data:
maya.k_: obsessed. how is the light even doing that??
→ ninna.artist: i asked the sun nicely 🤍 more of this era tomorrow
lucia.mdrd: ¡Diosa! 🤍
→ ninna.artist: gracias mi amor · esta era es para vosotros 💚
dasha.spb: богиня
→ ninna.artist: спасибо, родная · дальше будет ещё красивее ✨
Three languages, three replies, one voice. Lowercase, warm, present tense, one emoji, no exclamation marks · the rule set holds across languages because the rules are semantic, not tied to English grammar.
When the engine stays quiet
Silence is a feature. The default disposition on any comment that isn't obviously friendly should be "don't respond." Specifically:
- Anything containing a URL or promotional keyword.
- Anything from an account with fewer than 5 posts and less than 10 followers (spam signal).
- Anything flagged by the classifier as sensitive (health, politics, personal crisis).
- Anything that could be sarcasm the engine misreads.
All four cases route to the flagged bucket. You see them in the app; the engine leaves them alone.
Where this shows up in the product
The reply layer runs on the same 5-minute scheduler that publishes and drafts. Every posted post accumulates comments (from the Graph API when live, from a simulator that mirrors real distributions when in simulation mode). The reply worker walks the pending queue for the brand, batches by language, generates the responses, hits the reply endpoint, marks them as sent, and writes a comments.replied event with the languages served. That event lands in your Brain Activity feed on the Overview page ("replied to 14 comments · ES, RU, FR"), so you can see the layer working without having to open Instagram.
More on the loop it fits inside: what is an AI content engine? and can AI post to Instagram automatically? Join the waitlist for founding access.