An AI UGC creator is a casting brief, not a menu pick.

How to describe the person on camera so the ad reads as a person, which fields actually change the result, and where a generated performer still gives itself away.

casting.json

{
  "creator": {
    "age": "mid thirties",
    "look": "no makeup, hair tied back",
    "wardrobe": "grey sweatshirt, one size too big",
    "setting": "kitchen, morning light, clutter on the worktop",
    "delivery": "unrehearsed, half smile, one small stumble",
    "camera": "handheld at arm length, slight drift"
  },
  "aspect": "9:16"
}

The three fields that decide realism are the last three, not the first two.

What an AI UGC creator is

A synthetic performer built from a written brief, standing in for the person you would otherwise book.

An AI UGC creator is a face, a voice and a delivery generated for a video ad. No casting call, no shipping the product to someone, no waiting a week for a file that turns out to be filmed in the wrong orientation. You write who the person is and what they say, and a vertical take comes back.

You will see the same idea sold under three names. AI UGC actors usually means a performer generated per brief. A UGC AI avatar usually means a fixed identity from a library, reused across many videos so a campaign has one recognisable face. Creator gets used for both. The distinction that matters when you are choosing is whether you can describe the person you want or only pick one that already exists.

Libraries are faster to start with and better for continuity, because the same avatar in ad four looks like the one in ad one. Described creators are better for testing, because the person is a variable you can move. Most accounts end up using both: a library face for the campaign that is working, and described ones while hunting for the next angle. What both produce is the same kind of file, covered on AI UGC videos.

The casting fields that change the result

Most briefs spend all their detail on the face. The face is the field that matters least, because every model already renders a plausible one.

Age and life stage

The single field that moves performance most, because it decides who the viewer thinks the ad is about. A claim about waking at three in the morning lands differently at 26 and at 45.

Setting

Rooms carry more signal than faces. A kitchen with things on the worktop reads as a home, and an empty room with a plain wall reads as a set.

Wardrobe

Slightly wrong clothes make a person look real. Anything crisp, matched or freshly pressed drifts back toward a brand shoot.

Delivery

Ask for unrehearsed rather than confident. Confident is the default and the default is what people have learned to scroll past.

Camera behaviour

Held at arm length, with drift and a reframe, is the tell that says phone. A locked static frame says tripod, and tripods belong to brands.

Voice

Pace, accent and breath decide whether the lip sync survives a second viewing. A slightly slower read gives the mouth easier work to do.

Directing the take

A generated performer will do exactly what the brief says, including the parts you left on default. Defaults are what make an ad look generated.

Write for a mouth, not a page
Read the script out loud and cut anything you stumble on. Contractions, short clauses and one filler word do more for believability than any render setting.
Give one physical beat
A person doing nothing with their hands looks like a talking head. Pick the product up once, on a specific line, and the whole take stops feeling like a presentation.
Leave the imperfection in
A half second of dead air before the close, a glance away, an unbalanced frame. Ask for these directly, because no default setting will give them to you.

direction notes

notes for take 2

- drop the word "amazing", she would not say it
- start talking before the shot settles
- one physical beat: pick the sachet up on
  the line about the change
- eyes to the lens on the close, not before
- leave the pause after "three in the morning"
- no music, room tone only

Notes for a second take are just text, so they can be replayed on the next product.

Where it still reads as fake

Worth knowing before you spend, because most of it can be designed around rather than fixed.

Hands doing fiddly things. Peeling a seal, twisting a cap, passing an object from one hand to the other. Shoot those beats as separate product inserts and cut to them.
Long unbroken takes. The longer the camera stays on one generated face, the more time a viewer has to notice the eyes. Cut every four to six seconds.
Text on packaging. Any label the model has to render tends to come back close but wrong. Keep the pack out of focus, or composite the real one over it.
Perfect eye contact for twenty seconds. Real people look away. Ask for one glance off the lens and the take relaxes immediately.
Studio audio. Room tone, a little reflection and no music read as a phone. Clean voiceover over silence reads as an ad.
Anything that needs a specific real person. A generated creator cannot be your founder, your customer or a named expert, and pretending otherwise is a disclosure problem rather than a quality one.

Casting from the window you already work in

Picking a creator from a grid of thumbnails is pleasant once. It is the slowest part of the job by the fortieth ad.

A casting brief is structured text. An agent can hold five of them, run the same script through each, and hand back five files named after the person.
Direction notes turn into another run without a context switch. You reply in the chat, the agent adjusts the inputs and reruns, and the previous take is still on screen for comparison.
The creator description can live in the same repository as the rest of the campaign, so the person in ad forty is provably the person in ad one.
Nothing has to be re-picked. A dashboard forgets your choices between sessions. A file does not.

Wireflow is our own product, and it exposes its generation workflows over a hosted MCP endpoint for exactly this reason. Listing the workflows and the models is read only, so an agent can look before anything is spent. The same approach applied across a client roster is on AI UGC for agencies.

cursor

You

Same sleep tea script, three different creators: late twenties, mid thirties, mid fifties. Kitchen, handheld, unrehearsed.

list_models{ "capability": "video" }

↳ avatar and lipsync models available, 9:16 supported

Agent

I will keep the script, the setting and the camera note identical and change only the age band, so the comparison is clean.

run_workflow{ "workflowId": "ugc-ad-v4", "inputs": { "creator": "f-55-kitchen", ... } }

↳ executionId exe_9c04a, two more queued

get_execution{ "executionId": "exe_9c04a" }

↳ status completed, output tea-creator-55.mp4

Agent

Three takes. The fifty five read is the calmest and holds the pause before the close. The late twenties one rushes the second line, so I can slow that read and rerun just that one.

list_models, run_workflow and get_execution are real tools on the server. Only run_workflow spends credits.

Where the rest of this lives

Once the person on camera is settled, the next decisions are about the script and the buy.

The script structure the creator has to deliver: AI UGC ads.
Placements, lengths and specs for the ads they end up in: UGC video ads.
The connect once and generate from chat setup: the AI UGC generator overview.

Questions people ask about AI UGC creators

What is an AI UGC creator?

It is a synthetic performer generated for creator style video ads: a face, a voice and a delivery, produced from a written brief instead of booked for a shoot. Some tools call the same thing an AI UGC actor or a UGC AI avatar. The practical difference between them is whether you pick from a fixed library or describe the person you want.

Are AI UGC actors and UGC AI avatars the same thing?

Close enough in everyday use, though the words hint at different products. Avatar usually means a fixed identity from a library that you reuse across many ads. Actor or creator usually means a performer generated per brief, which gives you more range and less continuity between videos.

Can I keep the same AI creator across a whole campaign?

Yes, and you should when the ads are meant to feel like one person. Reuse the same avatar identity or the same seed and casting fields across the batch. Keep the description in a file so a run six weeks later starts from exactly the same person rather than a close cousin.

Where do AI UGC creators still look fake?

Hands and product handling are the weak spot, along with long unbroken takes, over even eye contact and any packaging text the model has to render. Short takes, one physical action per shot and product inserts filmed separately hide most of it. If a shot needs a label read on camera, film that shot or composite it.

Do I need a real person to license their likeness?

Not for a generated performer, because there is no real person to license. That is also the constraint: if your ad depends on a specific known face, this is the wrong category of tool. Use generated creators for volume testing and book real people for the ads that trade on who is saying it.

How do I direct an AI UGC creator from an agent?

Connect the MCP endpoint to Claude, Cursor or n8n, then describe the person and the read in plain language. The agent calls the generation workflow with those fields filled in, polls the run, and returns the video URL. Notes for a second take are just another message in the same conversation.

Describe the creator and run the take

Connect the endpoint, write the casting brief in a sentence, and read the result in the same window. The read only tools cost nothing, so you can look before you generate.