AI spokesperson video: what it can and cannot say

A clinic wants a short video explaining what a first appointment involves. An agency wants one introducing a service nobody has heard of yet. An estate agent wants one that walks a buyer through a listing before they drive across town to see it. All three are the same job: one person, talking to camera, explaining something. In all three cases the business would rather that person were not the owner.

What the video is being asked to do

A spokesperson video is explaining, not selling. The script is usually correct on the first draft, because the business knows its own service better than any writer does. The difficulty sits entirely in the delivery. Somebody has to stand in reasonable light and sound like they mean it, and then do it again in three months when the service changes, and again when a second location opens.

That is the demand behind the search. It is not really a demand for a face. It is a demand for a delivery that does not require the owner to block out an afternoon, find a clean wall and get their hair cut first.

Two ways a generated spokesperson gets built

In the script route you type the words. A voice model reads them and a face is animated to fit the audio it just produced. Nobody records anything, so the whole video exists as text until the moment it renders.

In the performance route you record a take. A person says the words out loud on a phone, and the system moves that delivery onto a different face. The timing, the pauses and the emphasis come from the recording rather than from a model's reading of the punctuation. The longer comparison of the two routes covers how each one fails.

For a spokesperson video the choice is narrower than it looks, because of what these scripts contain.

Where a script read stops working

Service explanations are mostly reassurance, and reassurance lives in the concessions. The sentence that earns trust in a dental video is not the one about technology, it is the one that goes: it takes about ten minutes, and yes, the first two are uncomfortable. A voice model puts the stress where the grammar suggests, so that sentence arrives as a flat recitation of two facts, and the concession stops being a concession.

The same problem turns up in anything with a number in it. A price, a wait time or a guarantee period is a line the speaker needs to slow down on, because the listener is doing arithmetic. Read at an even pace, it washes past.

Then there is pointing. A spokesperson in a service video usually has to indicate something physical: a floor plan, a screen, a machine, a doorway. A presenter animated from an audio track has no idea any of those objects exist, so the gesture either goes missing or lands on nothing.

The face belongs to somebody

A generated presenter still looks like a human being, whether it was built from a photograph or invented by a model. That resemblance is the part that attracts legal attention, not the generation.

Using a real person's likeness in advertising needs that person's written permission, and the permission has to cover advertising specifically. A release signed for a staff headshot does not extend to a paid ad running in three countries. A public figure's likeness is simply not available for this, at any budget, and a resemblance close enough to be recognised is close enough to be a problem.

What you have to disclose

Three obligations attach, and none of them care which technique produced the face.

The FTC's Endorsement Guides define an endorser to include parties that merely appear to be an individual, group or institution, which brings fabricated endorsers and computer-generated characters inside the guides. Where a material connection between that endorser and the advertiser would not be expected by the audience, it has to be disclosed.

The Commission's rule on consumer reviews and testimonials went into effect on 21 October 2024 and prohibits writing, creating or selling a testimonial that materially misrepresents that the testimonialist exists. A generated presenter delivering a first-person customer story is precisely that, which is why the spokesperson framing matters: a business explaining its own service is on solid ground, and the same face claiming to be a satisfied patient is not.

In the EU, Article 50 of the AI Act applies from 2 August 2026. Deployers who generate or manipulate video content constituting a deepfake must disclose that it was artificially generated or manipulated, in a clear and distinguishable manner, and no later than the viewer's first exposure to it. Design the label into the frame rather than adding it after the creative is signed off. The guide to AI actors covers the casting side of the same question.

How Sosie fits

Sosie takes the performance route. You record up to forty seconds on a phone, saying your own script, in whatever room you happen to be in. You then choose a presenter from the catalogue, describe one for Sosie to generate, or upload an image of someone who gave you written permission; public figures are blocked.

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 and its light, which is the part a service business tends to care about: a clinic video shot in the clinic rather than in a stock studio. A separate backdrop is available 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 covers up to twenty seconds and two cover up to forty. Credits are sold as monthly subscriptions or as one-time packs that never expire, from €39 for five credits, and there is no plan without payment. A render that fails returns its credit automatically.

One test before you commission anything

Find the hardest sentence in your script, the one with a caveat or a number in it. Record yourself saying it once on a phone, then read it back flat, with no inflection at all.

If the two versions say different things, your script needs a performance and a typed one will not carry it. If they say the same thing, a script-driven spokesperson will do the job and you can stop reading here.