My newest real AI experiment is Instagram carousel posts inside my daily social pipeline, and unlike almost everything else that pipeline does, it's deliberately kept off autopilot for now. The first real run taught me something about how trust actually gets built, and it wasn't what I expected going in.
What the experiment actually is
Once a day, at most, the social pipeline checks whether any of that day's five topics has genuinely enough real material behind it, three or more distinct real tips or story details already pulled from actual data, to justify a five-slide carousel instead of a single image. When one qualifies, it gets built as a real deck in my actual brand colors, with no invented mascot or made-up detail anywhere in it.
What has to be true before a carousel even ships
Every deck gets checked against a specific structure before it's allowed anywhere near scheduling: a cover slide, exactly three interior slides, and a closing call-to-action slide, no more and no fewer, with zero emoji, zero em dashes, and absolutely no invented mascot anywhere in it. A deck that breaks any one of those rules fails the same way a missing disclosure line would fail a Pinterest pin. It doesn't get a pass just because the format is new.
Why it's not on autopilot yet, even though everything else daily is
My daily social batch has run fully on its own since September 15th. Carousels specifically were left out of that standing approval on purpose, because that authorization covered the plain single-image, four-caption format I'd already watched work correctly for weeks. Any batch that includes a carousel still has to stop and wait for my own go-ahead every single time, until it's proven itself the same way the rest of the pipeline already has.
What did the first run actually teach me?
The most honest thing it taught me is that most days, nothing will even qualify. The bar is three or more real distinct beats already sitting in data I've already gathered that day, and on a normal day that bar simply isn't cleared. That's treated as completely normal, not a failure of the pipeline, which took some getting used to. I kept expecting a carousel to show up more often than it actually does.
The lesson underneath it
New capability earns its own trust separately, even inside a pipeline I already trust for everything else. Marie Gems' Pinterest pipeline needed roughly ten clean batches before it earned its own standing approval, and daily social needed its own separate stretch before September 15th. That's the normal path here, but it isn't the only one I've actually used: this site's brand-new Pinterest pipeline went to full autopilot after exactly one batch, a deliberate exception I made myself, the same day it launched, not a shortcut anyone talked me into. Carousels aren't getting that same fast track. I'm choosing to let them earn it the slow way on purpose, because a new format inside a pipeline I already trust is a different kind of risk than a whole new pipeline with its own separate quality gate.
What the experiment deliberately leaves out
The carousel format borrows its five-slide structure from a tool built for a completely different purpose, but it skips that tool's own character and mascot decision-making entirely. Marie Gems doesn't have a brand mascot, so nothing in a carousel deck is allowed to invent one just because the borrowed structure has a slot for it. Only the parts that were actually useful got kept, nothing else, which is the same restraint I'd want applied to any borrowed tool.
The same rule applies to bigger structural changes too
The new operations structure grouping my Marie Gems agents together, stood up this same week, was built the same cautious way. No new standing permissions were granted anywhere in it. The Pinterest and social pipelines keep exactly the same pre-authorized daily lane they already had, and everything else still routes through one real check-in with me before anything publishes. A reorganization doesn't quietly widen what's allowed to run on its own. Only actual, proven performance does that, one pipeline and one format at a time.
Where this leaves things right now
I'd rather have a genuinely rare, well-sourced carousel than a forced one every day just to hit a quota, and I'd rather wait on approving that specific format a little longer than hand it the same blanket trust I've already given the rest of the pipeline. Some things earn autopilot. Some things just aren't there yet, and that's fine.
The bigger picture on how I think about what to hand off and what to hold onto is in code, craft, and create.