I Made A Custom AI Video Avatar and Here’s What I Learned

Danny, our Chief of Staff, had plenty of fun when Sora 2 was released. Making funny, albeit unnecessary, videos no one on our team knew they needed (I’m being very liberal with my use of the word needed here). 

But Danny wasn’t the only one using AI video generation. AI video avatars have been all over my feed for months, and it got me thinking:

Could I create a believable video avatar of Nick, our CEO?

Lately, I’ve been wanting to do more video content, but Nick is hard to schedule with. He’s got a lot going on. AI Nick, on the other hand, is always available. 

So, as someone who helps businesses navigate marketing tools, I figured it was time to stop watching from the sidelines and actually build my own video avatar.

I went into this project equal parts excited and skeptical. 

What followed was a somewhat surprising, occasionally weird, and ultimately very educational few weeks. Here is everything I learned.

What Is an AI Video Avatar?

An AI video avatar is a realistic, AI-generated presenter that delivers scripted content on screen without anyone stepping in front of a camera. 

You may have also heard the terms an AI talking head video, a virtual spokesperson, or an AI spokesperson video. Different names, same idea.

For marketing teams, the appeal is obvious. No more booking studio time talent. No more reshooting because Nick fumbled a line or the mic didn’t work. Just write, generate, and publish. 

That’s the promise, anyway. I wanted to find out if that’s really the case.

What is the difference between a custom avatar and a stock avatar?

There are actually two types of AI video avatars available on most platforms, and they are not the same thing.

A stock avatar is a pre-built AI presenter the platform provides. You pick from a library of realistic-looking people, type your script, and the stock avatar delivers it. All AI video generators have  libraries of stock avatars, and I tested several. They are fast, easy, and a reasonable starting point if you just want to see how the technology works. They are faster and easier, but the presenter has no connection to your brand. For a generic explainer or an internal training video, that may be fine. For anything where trust and authenticity matter, like a paid ad or a brand-forward video, using a rando-AI-face to represent your company is a tough sell to perceptive audiences.

A custom avatar is built from footage of you or someone on your team. You record a training video of yourself, the platform analyzes your face, voice, gestures, and expressions. The result is an AI generated spokesperson video that actually looks and sounds like someone your audience can connect to your brand. This is what I spent most of my time building and testing, and it is the harder of the two to get right.

Custom avatars take more work, but the payoff is a spokesperson that is actually you. Kind of.

How are marketers actually using AI video avatars?

AI video avatars aren’t just a novelty anymore. Marketers are putting them to work across a wider range of channels than you might expect, and the use cases are growing fast.

Social media is the obvious starting point BUT if you’re going to use it there, it needs to be good. Short-form video for LinkedIn, Instagram Reels, TikTok, and Facebook seems like a natural fit because the format is quick and the volume demand is relentless. Posting consistently is a lot easier when you’re not trying to book Nick for 20 minutes every week. 

Paid advertising is where we think the ROI case is clearest right now. Video ads outperform static ads, but producing them at scale is expensive and slow. AI avatars make it possible to test multiple versions of an ad with different scripts, tones, or messaging without reshooting anything. If you’re running A/B tests on ad creative, that alone could save you days of production time per campaign.

Product explainers and demo videos are a strong fit too. If you’ve got a product line that changes frequently, or a SaaS platform with regular feature updates, an avatar can redeliver updated content without the full production cycle. No rescheduling. No reshooting. Just update the script and go.

Nobody’s winning awards for AI-generated onboarding videos. But if you’ve ever had to reshoot a 10-minute HR update because a policy changed, you understand the appeal immediately. Internal communications and training content are low-risk places to experiment and build confidence with the technology before putting it in front of customers.

The common thread isn’t complexity or budget. It’s volume. The more video you need on a regular basis, the stronger the case becomes.

Is AI video generation actually that good now?

It depends on who is doing it.

If you have experienced video and marketing professionals on your team who can invest real time into learning the platforms, dialing in the training footage, and iterating on the output? Yes, AI video generation can be impressive and it’s getting better fast.

If it is just you experimenting on the side? There is more to it than most of us were hoping for. A year ago, AI-generated video had the telltale signs everyone knows: stiff movement, glassy eyes, mouth shapes that never quite matched the words. Today, the gap between real and generated has closed. But closing that gap takes work, patience, and a fair amount of trial and error.

For short-form content, social clips, and explainer videos, AI video quality is now good enough that casual viewers may not notice. But “good enough for casual viewers” is not the same as “indistinguishable from real.” So let’s get real.

Can you actually tell AI video from real video?

Sometimes yes, sometimes no. Good AI video is genuinely hard to spot, but getting it to that point takes real effort. It does not happen automatically out of the box.

Here is what I noticed during testing: the moments that give AI avatars away are rarely the big obvious things. It is the small stuff. The micro-pause before a gesture that a real person would never have. The way enthusiasm in the voice does not quite match the energy in the face. A blink that happens a half-beat too late.

Human communication is packed with unconscious signals we absorb without thinking. AI avatars are getting better at the words and the big movements. Where they still fall short is in the thousand tiny nonverbal cues that make a person feel genuinely present.

Which platforms were tested?

I tested two: InVideo AI and HeyGen. I just don’t have time to test a bunch and these popped up in my research and looked like a great place to start.

First, I had to make a training recording. >2 minutes of me talking. (Ugh.) I had to record and upload several different attempts to learn what worked. I tried scripted, unscripted, animated gestures, no gestures. My first tries produced really weird results (read the next section to get a good laugh).

InVideo AI was the rougher experience. The gestures felt disconnected from the content, like the avatar was doing a presentation for a different script. Some facial expressions read as exaggerated in a way that felt more cartoon than human. For a brand that values authenticity, it would be hard to put that in front of an audience.

HeyGen was noticeably better. The lip sync was tighter, the overall feel was more polished, and the output felt closer to something usable. That said, it still had moments where you could feel the AI behind the curtain. While you can enhance the voice inflection with amazing results some of the pauses, “chuckles”, and “sighs” that it inserted were flat. And there were subtle visual issues with expressions that broke the illusion just enough to notice.

Neither platform is perfect. HeyGen is the stronger starting point, but both require realistic expectations about the output.

What is the biggest factor in avatar quality?

Your training recording. No question.

The AI can only work with what you give it. 

Feed it a stiff, monotone, poorly lit recording, and you will get a stiff, monotone, darkly lit avatar. On the other hand, overly excited gestures can give you bizarre, and frankly, hilarious results. 

Expressions that are slightly off, gestures that feel disconnected from the words, an overall vibe that is somewhere between “something is wrong here” and “why does this person look like they are being controlled by a puppet master.” 

Things can get genuinely weird and creepy in ways that are hard to describe until you see it.

So let me show you!

Disclaimer: This video avatar of me is embarrassing, a little unsettling, and a solid reminder that no one should let AI make their first impression for them without some serious prep work. 

My first avatar test with InVideo:

Garbage in, garbage out. This rule does not get more true anywhere than it does here.

After a bunch of time experimenting and moving to HeyGen, I created this avatar for Nick and testing it out. It’s still not right, but we’re getting closer.

What I learned about making a good AI avatar training recording.

This took me a few attempts to figure out, and more time than I expected. Plan to spend a solid two to four hours just on your training footage when you factor in setup, multiple takes, reviewing the output, and going back to reshoot. 

Here are the five lessons I learned:

1. Be expressive, but keep it controlled. You want your face, voice, and hands to be animated and natural, but you do not want wild gestures or exaggerated expressions. Think: engaged presenter, not hype person. The AI uses your range to build its own, and if your range is huge and unpredictable, the output will be inconsistent.
2. Speak with intention and vary your tone naturally. Use a script to avoid filler words like the dreaded “ummmm”. Hit your emphasis words the way you would in a real conversation. 
3. Look at the camera as if it were a person. Not through it, not past it. Be direct, warm, present. This is the single biggest thing that separates a convincing avatar from a creepy one.
4. Keep your background clean. A plain white wall with even lighting works great. Opt for front-facing natural light or a softbox. Shadows on your face give the AI less to work with.
5. Record more than you think you need. Most platforms want 2 to 5 minutes of footage minimum. Give them closer to 10 if you can. More variety in your natural expressions and speech patterns means a more flexible, convincing output.

Is an AI Video avatar worth it for marketing?

It depends on what you need it to do, and how much time you are willing to put in upfront.

I found that the demos leave out the reality of what you’re in for: this is not a few clicks and suddenly you have a dozen polished videos ready to go. 

Learning the platform, building a quality training video, reviewing outputs, iterating, and troubleshooting the weird stuff takes real time and effort. Just like any other marketing tool worth using. The time savings kick in eventually, but you have to earn them first.

 Once you have put in that foundational work, AI avatars are genuinely useful for high-volume, low-stakes content like internal updates, FAQ videos, or social content The ongoing time savings are real, and the quality is workable.

But for me, real video still wins for brand-forward content where trust, warmth, and human connection are doing the heavy lifting.  The small things that make a person feel real are also the things that make a brand feel trustworthy. An AI avatar that is slightly off in its expressions or timing can do more harm than good if your audience picks up on it, even subconsciously.

Where does that leave me? Using AI avatars as a supplement, not a replacement. They are a solid tool for the content that I don’t have time to film properly, but they’re still not yet ready to be the face of our brand in high-stakes moments.

Are we ready to offer AI video avatars as a service?

Not yet. I’d rather be honest than rush something out the door that is not ready. That’s just how we roll.

What we know for sure is that this technology has real potential and that means more testing, more iteration, and a lot more embarrassing results before we are ready to call it a service. In another 6 months the tech will almost assuredly get better, but by how much, who knows.

So stay tuned. I’ll keep sharing what we learn, including the parts that didn’t go well, because that is usually where the most useful information lives.