AI avatars and AI voice acting for creatives 2026: HeyGen, ElevenLabs, Arcads

Once every couple of weeks the founder writes to me: “we want UGC, but it’s expensive to shoot, can we use a neural network?” This request always means the same thing - you need not just one beautiful shoot, but a factory of options to quickly generate hypotheses. AI avatars and AI voice acting are just about this. I tested HeyGen, ElevenLabs and Arcads on combat links and below I will analyze not “what are neural networks for video”, but specifically: how much does the video cost, where the avatar enters the performance and where the viewer senses a fake in two seconds.

Immediately frame. This is not an article about “AI will replace the film crew.” Will not replace. This is about a tool that cheaply closes one section of the creative funnel - a mass test. Over 9 years in digital, I have shot different things, and AI video is not magic, but another way to reduce the cost of entering into a hypothesis.

An AI avatar is not a “cheap replacement for an actor.” This is a machine for stamping options. If you don’t have the “20 hypotheses per week” bundle, you’re not using the avatar for its intended purpose - you’re just saving on one video and wondering why it didn’t work.

1. What does this actually give?

Three tools cover three different tasks. HeyGen makes a “talking head”: you upload text, the avatar pronounces it with his lips in a lip sync. ElevenLabs - voice synthesis, separate from the picture: voice-over, voice-over, dubbing. Arcads - UGC format: a library of AI actors who play the scene “a person talking about a product from a phone.”

The main value is one - speed and price at scale. A live UGC video involves casting, briefing, filming, and editing. A week and tens of thousands of rubles apiece. AI option - an hour and a couple of hundred rubles. When you need to test 15 different hooks for one audience, the difference makes all the difference. I testing hypotheses packs, and the bottleneck has always been in creative production, not in campaign setup.

One more moment. An AI avatar removes dependence on a person. The actor got sick, changed his mind, asked for an additional payment for a new version - there are no such problems with synthetics. I changed the line in the script, regenerated it, rolled it out. For flow of advertising creatives This is the flexibility for which everything is started.

2. How much does the video cost?

The numbers below are ranges from my tests and subscription rates for 2026, not the price list of a specific client. A HeyGen or Arcads subscription is a fixed monthly fee and a pool of minutes or generations. If you spread it out by volume, one finished video comes out something like this.

The difference is in order. 10 options on AI = half a day and a couple of thousand rubles. 10 live options = a month and a budget, which many projects, in principle, will not accommodate. This equation explains why avatars have taken root in performance art. Not because they are “better”. Because they provide the volume at which A/B tests make sense at all.

3. Where does it come in?

An AI avatar works where the viewer does not expect intimate contact with a person. I'll list it specifically.

• Hypothesis test for a cold audience. You need to understand what hook is catchy, and not make a masterpiece. 20 avatars with different first phrases solve this problem in a day.
• Info-offers and online courses. The audience is accustomed to a “talking head in the background”; the AID avatar here is almost indistinguishable from a regular speaker at a webinar.
• Simple product demos. Show the interface, speak features voiceover on ElevenLabs - the viewer doesn’t care whether it’s synthesis or a live speaker.
• Localization. One video, ten languages ​​via synthesis voice acting. For an international performance, this is a huge saving on dubbing.

At NEMIFIST, I was targeting gaming coverage through bloggers, and they needed a live texture. But parallel performance links for cold traffic would work great on avatars - for testing hooks, the face of a particular person is not critical at all. This is the border: where recognition and trust are important, a living one is needed; Where the volume of checks is important, AI comes in.

4. Where the viewer feels fake

Now the downside. There are niches where the avatar drops metrics, and this is not a question of the quality of generation. It's about trust.

Medicine, finance, personal brand, premium segment. As soon as the viewer needs to trust the face - to believe that this person actually used the product, was really cured, really earned money - uncanny valley turns on instantly. The eyes are almost in the wrong direction, the microfacial expressions are dead, the intonation is even where a living person would falter. The viewer may not consciously understand what is wrong. But the finger had already scrolled further. Viewability drops, followed by CTR.

Separately, the face of the brand and customer reviews. If you make a “real review from a satisfied customer” on an AI actor, you risk your reputation: they will burn you, and your trust in the brand will sag for a long time. Here, saving on filming turns into an expensive disadvantage. I would not touch AI avatars in niches where trust is being sold. For everything else, try it.

5. HeyGen, ElevenLabs, Arcades and live shooting

Head-on comparison. Quality and prices are my observations from tests, not from press releases.

The bundle is often assembled as a hybrid: a picture or background is generated in Midjourney and DALL-E, the dynamics are adjusted to video generators, the voice is from ElevenLabs. The avatar is just one layer of this pipeline. If you want a generated scene, and not a talking head, look towards Veo - I analyzed it in guide for marketers.

6. What format should be done on AI

In order not to guess at each brief, I keep a simple matrix: the format of the creative versus whether to drag it on an avatar or live.

The rule is simple. If the viewer needs to believe a person, film the person. If the viewer needs information and you need a range of options, take an avatar. Check everything in between with an A/B test, not with intuition.

7. Risks of sites and the right to face

Two layers of problems that I put into the plan before the start.

The first is moderation. There is no direct ban on AI, but Meta and TikTok are taking a tougher look at synthetics in sensitive categories, and some sites already require the “AI-generated” label. Deepfake advertising for finance and medicine catches deviations more often than usual. This doesn’t mean “you can’t”, it means “set aside a rotation reserve and don’t build the entire funnel on one avatar.”

The second is the right to face and voice. You cannot clone someone else’s appearance or voice without written consent: this is personal data and the right to an image, and this comes not from the site, but from the court. The platforms' stock avatars are licensed, but read the terms - some of them prohibit certain niches and political advertising. For the face of the brand and employees, document the rights of use before the video goes on air.

Conclusion

AI avatars and synthesized voice acting are not “cheap video instead of filming.” This is a machine for testing hypotheses: cheap, fast, in the volume at which A/B generally works. Applies to cold hooks, info offers, product demos and localizations. It falls where trust is sold - medicine, finance, brand face, reviews. The border lies on exactly one question: whether the viewer should believe the person. Must - shoot a living one. Shouldn't - generate twenty options and merge the unsuccessful ones.

If you are thinking about where to insert AI avatars into your funnel and not experience a CTR drawdown in a trusted niche, write to me on Telegram @dipustovalov or leave a request via form. Starting consultation - 0 ₽: we’ll analyze your formats and where synthetics will really save, and where they will kill the metric.