AI actors for video ads: how they work and the rules

The phrase "AI actor" carries two conversations at once. One is about cinema and the actors' unions. The other, much smaller and much more practical, is about a shop owner who needs twelve vertical ads by Friday and does not want to be in any of them. This guide is about the second conversation: what an AI actor is in advertising, the three ways one is made, what each way gets wrong, and the rules you have to respect before the first render.

What an AI actor is, in an ad

In advertising, an AI actor is a presenter who appears on screen and delivers your message, and who was generated by a model rather than filmed. They may be entirely synthetic, or a synthetic face on a real performance. What they are not is a customer. That distinction runs through everything below.

The reasons people want one are mundane. They do not like being on camera. They cannot afford a creator for every variation. They sell in three languages and cannot film in three. They want to test ten hooks and only have one face. An AI actor answers all of those, at the cost of some honesty work that a human presenter would have forced on you anyway.

The three ways AI actors are made

Tools that sell "AI actors" use one of three techniques, and they are not interchangeable.

Script-to-avatar

You type the script. The tool synthesises a voice, picks or generates a presenter, and animates the mouth and some of the head to match the audio. No camera is involved at any point.

It is the fastest path from a blank page to a video, and the cheapest per unit. Its weakness is the performance. Every human decision in a take, where to pause, which word to lean on, when to smile, what the hands do, is invented by a model that has never read your offer. The face is usually fine. The delivery is where viewers notice.

Performance capture

You film the take yourself, on a phone, and a model transfers that performance to a different presenter. Runway's Act-Two is the reference model here: it takes a driving video of a person and a character image, and produces the character performing what the person performed, including facial expression, head movement and, depending on the input, body gestures.

The result keeps your voice, your timing and your gestures. What changes is the face and, if you want, the room. It costs you a real recording, which is the point: the performance is yours, so it has the emphasis a script generator cannot fake. This is the technique Sosie is built on.

Full generation

You describe the scene and a text-to-video model invents everything: person, room, motion, sometimes speech. It is the most flexible and, for a talking presenter holding a specific product, the least reliable. Product shapes drift, hands do strange things, and a second take is a different person. It suits b-roll better than a spokesperson.

What each technique gets wrong

Knowing the failure modes tells you which one to buy.

• Script-to-avatar fails on delivery and on hands. If your ad needs the presenter to hold, open or point at something, expect trouble.
• Performance capture fails when the input is poor. A dark room, a moving camera, a face turned away or a mouth covered by a hand gives the model nothing to transfer. It also inherits your acting: if the take is flat, the render is flat.
• Full generation fails on consistency. The same presenter rarely survives two prompts, which makes a campaign of variations hard to hold together.

None of the three fixes a weak script. All of them make a strong one cheaper to test.

What AI actors cost

Pricing in this category is per video or per minute of output, usually sold as credits. Two things to check before comparing prices.

First, the length a credit buys. Sosie's unit is one credit for a video of up to twenty seconds, and twenty-one to forty seconds costs two credits, because the underlying cost is per second of render. A tool that quotes a lower price per video for thirty seconds may be quoting a different product.

Second, what happens when a render fails. With Sosie a failed render is refunded to your balance automatically; not every tool does that.

For reference, Sosie's one-time packs run from €39 for five videos to €219 for forty-two, and monthly subscriptions from €49 for eight videos to €279 for seventy, with unused subscription credits rolling over one month. Pack credits never expire. There is no plan without payment. Current prices are always on the home page.

The rules before your first render

You need the rights to the face. A real person's likeness may only be used with their written consent, and public figures are prohibited by any tool that takes the question seriously. Sosie's terms require that you upload only yourself or people who have consented in writing, that the people shown are adults, and that you keep proof of consent; the rendering provider also runs automated public-figure checks.

The presenter is not a customer. The FTC's Endorsement Guides require endorsements to reflect an honest opinion and material connections to be disclosed. A generated presenter has no opinion. Write the ad as a presentation and a demonstration, not as testimony. "This is the product, this is what it does, here is how to get it" is the safe register; "I used this and it changed my life" is not.

Synthetic media has transparency rules in the EU. Article 50 of the EU AI Act sets disclosure obligations for AI-generated or manipulated content, phased in on the Act's timetable. If you run ads in Europe, plan on labelling.

Platform policies apply on top. Ad platforms have their own rules on synthetic media and on misleading claims, and they change. Read the current policy of the platform you buy on, not a summary of it.

How to perform for an AI actor

If you choose performance capture, the recording is the product. A few practical rules, learned the hard way:

1. Face the camera, at arm's length. A phone held out in front of you is exactly right. The model needs to see your whole face.
2. Light from the front. A window in front of you, not behind you.
3. Record somewhere quiet. Your audio is kept and becomes the presenter's voice.
4. Keep it under twenty seconds if you can. It costs one credit instead of two and it is the length the feed rewards anyway.
5. Say the hook first. The first two seconds are the ad.
6. Do the gestures. They transfer. A flat take gives a flat render.

How Sosie fits

Sosie is a performance-capture tool for vertical video ads. You record up to forty seconds on your phone. You pick a presenter from the catalogue, describe one in words and Sosie generates them, or upload an image of someone whose likeness you have the rights to. Sosie keeps your voice, timing, expressions and gestures and puts them on that presenter. By default the presenter is placed in your own room, with your walls, floor and light; you can switch to the presenter's own backdrop instead.

Output is a vertical 1080p MP4, landscape if your clip was landscape, with no watermark and cleared for commercial use including paid ads on Meta, TikTok, YouTube and Google. A twenty-second clip renders in about five minutes. Because your own audio drives the video, the presenter speaks whatever language you spoke, in your voice and your accent. The product's guidance is the presenter-and-demo style described above, never a fabricated testimonial.