The seven checks every auto-sent reply has to pass

The scariest sentence in social media automation is “it posts for you.” We built auto-pilot anyway, because triaging forty mentions a day is a real tax — but we built it with the assumption that a single bad auto-reply costs more than a hundred good ones earn.

So before any reply ships without you, it clears seven checks. In order:

The rails

• —Confidence threshold — the draft's voice-match score has to clear the bar you set (default 85%). Below it, the reply waits in your inbox.
• —Sentiment check — if the incoming mention reads negative or controversial, it always waits for you. Disagreement deserves a human.
• —Profanity and spam filter — the outgoing draft is screened; anything that trips the filter is held.
• —Account-age check — brand-new accounts never get an automatic reply. That's where impersonators and baiters live.
• —Rate limits — hourly and daily caps keep the cadence human. Replying to forty things in four minutes looks exactly like what it is.
• —Quiet hours — you set a window; nothing sends inside it. Your 3am timezone overlap shouldn't imply you're awake.
• —Blacklist — anyone you've blocked is never answered automatically, no matter how high the confidence score.

And when it's wrong anyway

Every send — automatic or manual — has a five-minute undo that deletes the reply from the platform. Not “hides it from your dashboard.” Deletes it at the source.

The design principle underneath all of this: auto-pilot never makes a judgment call on your behalf. It ships the obvious ones and holds everything else. When it skips a reply, it logs why — “confidence 78 below threshold,” “incoming mention is negative” — so you can tune the rails instead of guessing.

Everything that doesn't qualify simply waits. Auto-pilot never denies a reply on your behalf — declining is a human decision too.