AI UGC: how to make a creator-style video ad

A creator-style ad used to need a person willing to be on camera. AI UGC removes the casting problem without removing the person: you still write the script, you still perform it, and only the face that delivers it changes. What follows is the production workflow in the order you actually do it, and the checks worth making before you pay for any tool.

What AI UGC changes, and what it does not

AI UGC is a video in the creator style, shot like a phone video rather than a commercial, where the person on screen was generated instead of hired. If the category itself is new to you, the guide to UGC covers the definition, the three things marketers mean by the word, and why the style works in a feed. This page assumes you already want to make one.

What changes is casting and logistics. There is no creator to brief, no product to ship, no shoot day to schedule, no usage window to renegotiate when the ad keeps working. What does not change is the writing, the performance and the law. Someone still wrote those words, someone still decided where the emphasis falls, and an ad that misleads is still an ad that misleads whoever appears to be saying it.

Write the script before you open any tool

Three constraints do most of the work, and none of them involve software.

One idea per video. A hook names one problem, one surprise or one number. Two ideas in twenty seconds means neither one arrives.

Say the claim, then show it. Spoken claims are cheap and viewers know it. The hand opening the box, the app loading, the stain going, the before and the after: those are the seconds people believe.

Close by saying what to do. A spoken call to action outperforms a graphic card in this format, because the viewer is listening rather than reading.

There is a fourth constraint specific to generated presenters: moderation. The audio is checked before anything renders. Sosie's own guidance when a render is rejected on the audio is to re-record without background music and to avoid naming other companies, medical or money claims, or reading a web address out loud. Writing that way from the start saves a round trip.

Record the take that drives the video

With performance capture, the recording is not a formality, it is the input. Runway's Act-Two documentation describes the driving performance as what captures the movement, expressions, audio and gestures transferred onto the character. Whatever you do badly here arrives on the other face intact.

In practice that means: front-facing, phone at arm's length, a room with enough light to see your eyes, and somewhere quiet, because your audio is kept rather than replaced. Hold the energy slightly above conversation level. Keep your hands where the frame can see them if the gestures are part of the point. Shoot one take per hook rather than one long take you plan to cut, since the take is the unit the model works on.

Choose who delivers it

Three routes exist, and they differ in how much control you get and what rights you need.

A catalogue presenter is the fastest: someone already cleared for this use, picked from a grid. A described presenter is generated from a sentence you write, which is the route to take when the person needs to look like your customer rather than like a model. An uploaded presenter is a photo of a real person, and it is the only route with a legal prerequisite: written permission from that person, and never a public figure. Serious tools block the latter outright and refund what they block.

Decide what is behind them

This is the choice most people skip and most viewers notice. A generated presenter standing in a generated studio reads as generated, because that studio has appeared in a thousand other ads. A presenter standing in your actual shop, kitchen, gym or workshop reads as a video someone shot there. If the tool can place the presenter in your own room from your recording, that setting is usually the more convincing option for an ad that is meant to look unplanned.

What to check before you pay

The AI actors guide compares the underlying techniques. Once you have picked one, these are the commercial details that decide whether the tool is usable:

• Output spec: resolution, orientation and whether a watermark is burned in.
• Whether the licence covers paid placements, not only organic posts.
• What happens when a render fails, and whether you pay for it.
• The consent policy for uploaded faces, and whether public figures are blocked.
• How language is handled. Script-driven tools need a synthetic voice in your language; performance capture speaks whatever you spoke.
• The pricing unit, per video or per second, and what a longer video costs.

Disclosure is part of the build now

Three obligations apply to a generated presenter, and none of them are optional.

Endorsements must be honest. The FTC's Endorsement Guides require a material connection between speaker and advertiser to be disclosed, and the Commission's rule on consumer reviews and testimonials, effective 21 October 2024, prohibits testimonials from people who do not exist or never used the product. A generated presenter is both.

Synthetic video has to be declared in the EU. Article 50 of the AI Act applies from 2 August 2026 and requires deployers of deepfake content to disclose it clearly, at first exposure at the latest.

Platforms ask for a label too. TikTok requires creators to label realistic AI-generated content and applies its own label when it detects it.

How Sosie fits

Sosie is the performance-capture route. You record up to forty seconds on your phone, saying your script, in any room. You pick a presenter from the catalogue, describe one and Sosie generates them, or upload an image of someone who gave you written permission. Your voice, your timing, your expressions and your gestures are kept and delivered by that presenter. By default the presenter is placed in your own room, with its walls, its floor and its light; you can switch to the presenter's own backdrop instead.

What comes back is a vertical 1080p MP4 with no watermark, cleared for commercial use including paid placements, in about five minutes for a twenty-second video. One credit renders up to twenty seconds and two cover up to forty. Credits come as monthly subscriptions or as one-time packs that never expire, the smallest being €39 for five credits, and there is no plan without payment. A render that fails is refunded automatically. The terms prohibit uploading anyone who has not consented, and the product's own guidance is presenter-and-demo copy rather than a fabricated testimonial.

A first week that produces something

Pick one offer. Write three hooks for it. Record all three on your phone in one sitting, generate the same presenter for each, and run them against each other. The tool is the smallest variable in that week; the hooks are the experiment, and they stay yours whichever face ends up delivering them.