One of our AI influencers was asked to make a sponsored reel for a real product: a small, screenless wearable you clip to your jacket and talk to. It answers out loud, so your phone can stay in your pocket. A simple brief, and a product that is easy to explain in a few seconds.
It took six attempts over two days to get an ad we would rate seven out of ten. Every attempt failed in a different way, and almost none of the failures were about the image or video models being "bad". They were about how the ad was planned, checked and handed from one step to the next.
This is the honest list, in order, with what each failure taught us. If you make AI influencer content for brands, you will probably hit most of these.
Six attempts, six different problems.
Attempt 1: the product was invented
The first version looked polished, and showed a gadget that does not exist. The written idea named the product, but the image model had never seen it, so it drew something plausible: a glowing white sphere. The real product is a flat black disc.
What fixed it: looking the product up and saving real photos of it before anything is drawn, then giving those photos to the image model for every frame that shows it. We wrote this up in detail in How to Make an AI Influencer Ad With a Real Product.
Attempt 2: the product was right, but it was just a prop
A good AI UGC ad is decided before anything is generated: in about 60 words of script. Here is the five-beat structure our AI influencers' producer writes to, how many words fit in 20 to 30 seconds, and the mistakes that make a script sound like an ad.
With the real photos, the product looked correct. But the idea was a fashion joke about an all-black wardrobe with "one exception", and the gadget was a glowing ornament in it. It never did anything. The team had treated a sponsored ad like any other viral post, starting from a trending format and a joke.
What fixed it: an ad gets its own brief before any idea is written. What the product does for the person using it, one key message, and a "proof moment": the single shot where the product visibly does its job. For a voice gadget, that means she talks to it and it answers. Every idea has to work inside that brief.
Attempt 3: one long shot that never moved
The next version followed the brief, but it was drawn as a few frames of the same street corner. She was "walking", yet every frame had the same shop window behind her, from the same distance. When the video model animated between them, nothing really went anywhere.
The cause was an instruction meant for a different kind of scene: every frame was told to keep the first frame's background exactly, which is right for a café table and wrong for a walk. Nobody had decided where the camera stood for each shot either.
What fixed it: planning each shot like a film crew would, with where she is, where the camera stands and what is behind her, and only using the previous frame to match her face, outfit and light. We covered this in Why AI Storyboard Shots All Look the Same.
Each shot needs its own camera position, not a copy of the first.
Attempt 4: she talked, but nothing came out
This version had the right product, the right idea and real movement. In the last shot she dipped her chin to the gadget, lips moving, as she spoke to it. But the prompt for the video model said "no dialogue", so she mouthed silently, and the gadget never answered. Then the video stopped mid-gesture, with her mouth open.
Two problems hid in one shot. The last cut asked for four actions in three seconds (dip, talk, touch, look up), so the model ran out of time. And the video's length was simply the planned shot times added up, with no room for an ending.
What fixed it:
Any shot where she talks must have its line written out, word for word, and the time it takes to say it. A shot of someone talking with no words is rejected before anything is drawn.
One action per shot, with each shot timed like an editor would: about a second for a glance, two to three for a movement.
Every video ends on a held moment of a second and a half, so it finishes instead of stopping.
Nothing under ten seconds. Shorter reels felt too slight to be worth watching.
Attempt 5: the wrong "product photos"
A new version came back with a generic black button where the gadget should be, and the reviewer flagged it. The problem was ours: the photos saved as "the product" were the wrong pictures. One was a glossy marketing render with a glow the real device does not have, and one was the companion app's icon. The image model drew what it was given.
What fixed it: choosing product photos that show the real object plainly (held in a hand, worn on a jacket), and having the reviewer compare the product in every frame with those photos, detail by detail: shape, logo, the small camera notch at the top.
Attempt 6: a frame quietly disappeared
The last failure was the sneakiest. One planned shot was a close-up of the gadget on her collar, with her chin just at the top of the frame. Our automatic face check compares each picture with the influencer's face, and with only a chin in shot it scored too low three times, so the frame was dropped. The video plan still had five shots; the storyboard now had four.
The reviewer caught it, because by then it was checking the whole set against the script: five cuts in the plan, four frames, and the missing one was the shot that sells the product.
What fixed it: a storyboard frame that misses the automatic face check is no longer thrown away. It goes to the reviewer, who looks at it next to the character's reference sheet and decides.
What finally worked
The version we kept is 21 seconds long. She stands in a tailoring atelier holding a pinned skirt in both hands while her phone buzzes on the table. She looks at the phone, then at her full hands, says to the camera that the gadget has no screen and you just talk to it, then tells it what to remember. It answers out loud. She nods. The product is real in every shot, including a close-up where its logo and camera notch are sharp.
It isn't perfect. The first few seconds move too little, the buzzing phone could be clearer, and in the last shot the skirt she is holding reads as one she is wearing. But a stranger understands what the product does within a few seconds, and that is the point of an ad.
The lessons, in short
Give an ad its own brief. What it does, one message, and a proof moment. Don't run it through the same process as a joke post.
Draw from real product photos, and check that they really show the product.
Plan every shot: camera, place, what's behind her. Use earlier frames only to match the person and the light.
Write every spoken line, time every shot, and leave room for an ending.
Review the whole set against the script before paying for video. Most of our failures were visible in still frames costing a few cents, long before a video cost a few dollars.
Never let a step fail silently. A dropped frame, a cut-off sound line, or a rejection that gets overridden should be visible to whoever approves the post.
Every fix above is now part of how Influgen's influencer teams work: an ad manager writes the brief from the real product, the creative director plans and times each shot, and the critic checks every frame and the script together before you approve anything, with every check it ran shown on the post. You still decide what gets made into video and what gets posted.
Most AI ads give themselves away through the sound before the picture: dead air, a voice with no room around it, music fighting the words. Here are the seven fixes we now build into every talking ad, with the exact settings we use.
Our AI-made product ads were short, quiet and polished, and they felt fake. So we pulled apart the creator ads that actually sell gadgets on TikTok. Almost all of them do the same five things, and silence isn't one of them.
