Free PDF insideScript, test plan and disclosure checklist · Get the PDF
Book a call
← dependables. Playbooks

AI UGC video ads: 8 real examples and the 7-step playbook for running them

What brands are actually doing with AI avatars in paid video, what the sources report, where it goes wrong, and how to test it without faking a customer.

Paid socialAI UGCCreative testing

AI UGC means video ads that look like creator content but use an AI-generated presenter: a stock avatar, a cloned voice, or a person built from a script. Tools like Arcads, Creatify, HeyGen and TikTok's Symphony can turn a script into a talking-head ad in minutes, for a few dollars a video.

That makes it very good at one job: testing a lot of hooks and angles before you spend real money on production. It's bad at the job UGC was invented for, which is a real person vouching for a product they used. Most of the problems we see come from mixing those two jobs up.

Below are eight brands that have run AI-avatar ads, with what each ad does and the number its source reports. Then the playbook in seven steps: where it fits, how to script it, how to test it, how to label it and when to hand the winner to a real creator. Every example links to its source. Almost every number comes from the tool vendor or the platform, so read them as marketing, not as audited results.

Done for you

Creator Programs by dependables.

  • Creator sourcing and vetting
  • Briefs, contracts and usage rights
  • Content delivered ready for ads
  • One team running it end to end
Book a call or start now ›

Run by Stephanie Mendez ›

Built by the team behind Tabs Chocolate's 100M+ organic views a month.
Not ready? Get the free PDF playbook with every template. Get the PDF ↓

THE EXAMPLES

8 brands running AI-avatar video ads

Each card says what the ad does, where it ran when the source says so, and the number the source reports. We paraphrase and don't reproduce the ads. We haven't verified any figure ourselves, and none of these brands is a dependables client.

Vodafone

TikTok Symphony

TikTok · lead generation

Picked TikTok's stock avatars, which are licensed paid actors, and had a TikTok partner agency adapt them to Vodafone's look and tone of voice for ongoing lead campaigns.

45% lower cost per lead than branded assets from other channels (premium segment); 24% lower than branded assets on TikTok (mass segment)Platform-reported (TikTok)

The comparison is against Vodafone's other assets, not a controlled test. Source: TikTok for Business blog ↗

Teknasyon

Arcads

Consumer apps

Used AI actors to find a winning ad type, then scaled it by swapping in different avatars and scripts. Also made quick UGC-style ads that react to trends.

76% lower cost per purchase and 40% higher ROAS, for one of its appsVendor-reported (Arcads)

The source gives these for a single app, not the whole portfolio. Source: Arcads use case ↗

hipto

Arcads

Meta, Google, TikTok, Snapchat, Bing · lead gen

A lead-generation agency running avatar and template videos, including a campaign for seniors shopping for affordable health insurance.

49% more qualified leads and 14% lower cost per lead over one monthVendor-reported (Arcads)

Health insurance is a regulated category; the source doesn't describe the claims review. Source: Arcads use case ↗

Designrr

Creatify

Meta placements · 9:16, 1:1 and 16:9

An agency made 150+ variations in two weeks: avatar videos framed as a podcaster, a blogger and a course creator describing the product, AI product demos, and 20 to 30 different hooks on one core message.

CPA from $55 to $30 and ROAS from 0.77 to 1.33Vendor-reported (Creatify)

Persona 'testimonials' are the pattern step 3 warns about; script them as presenters. Source: Creatify case study ↗

1MORE

Creatify

Audio products

Replaced influencer and actor videos and some static ads with AI-avatar videos built from product pages, with AI-written scripts.

CTR up about 48%, purchases up 200%, under $3.90 a videoVendor-reported (Creatify)

No time period or method given, and the page states the CTR lift two slightly different ways. Source: Creatify case study ↗

Whole Life Pet

Creatify

Pet food · social

Swapped agency and influencer product videos for AI-avatar videos generated from product page links.

About $3.90 a video, against about $3,000 beforeVendor-reported (Creatify)

A production-cost number, not an ad-performance number. Source: Creatify case study ↗

Vitals

Creatify

Shopify app for merchants

Built AI-avatar and UGC-style video generation into its app as an add-on, so merchants make ads from their own product images.

1,000+ creatives a month; 40% uptake of the paid add-onVendor-reported (Creatify)

Shows adoption, not ad results. We left out a cost-per-customer claim the page doesn't support in its body. Source: Creatify case study ↗

trivago

HeyGen

TV · 30 markets

Used AI translation to localize its TV ads across markets while keeping one consistent brand face.

Post-production time cut in half; 3 to 4 months savedVendor-reported (HeyGen)

TV, not paid social. Included because localization is one of the clearest uses. Source: HeyGen customer story ↗

01 / 07 · FIT

1. Know when AI UGC fits and when to hire a real creator

AI UGC is a good fit
  • Testing many hooks and angles before paying a creator
  • Explainers where a presenter talks about the product
  • Screen-recorded apps, software and lead-gen offers
  • Localizing one script into many languages
  • Fast versions of a trend or a seasonal offer
Use a real creator
  • Anything that says "I used it and it worked"
  • Products people need to see in real hands: texture, fit, taste
  • Health, money and other claims a regulator reads closely
  • Audiences that trust a face they already know
  • Your best-selling angle, once you've found it

Our split, based on the examples and research on this page.

Look at what the examples above have in common. Vodafone and hipto sell leads. Teknasyon and Designrr sell apps and software. trivago used it to translate. In almost every case the avatar explains something, and the product shows up on screen as a recording, a demo or a page.

Now look at the research on the other side. NielsenIQ found viewers often spot AI-generated ads and rate them as more annoying, boring and confusing than traditional ads, with weaker memory activation even for high-quality ones. Digiday, citing Billion Dollar Boy research, reported that the share of consumers who prefer gen-AI creator content over traditional creator content fell to 26%, from 60% in 2023.

So AI UGC is a testing and explaining tool. It's at its best when the job is to say something clearly and many times over. It's at its worst when the job is to be believed: a skincare result, a taste, a fit, a before and after.

There's also a reputational floor. McDonald's Netherlands pulled an AI-generated Christmas ad in December 2025 after a backlash calling it lazy and creepy. That was a brand film, not UGC, but the lesson carries: if people feel tricked, the ad becomes the story.

How to set it up

  1. List the angles you want to test this month.
  2. Mark each one: does the ad need someone to have used the product, or just to explain it?
  3. Send the 'explain it' angles to AI UGC and the 'used it' angles to real creators.
  4. Check your category. Health, finance, insurance and anything with a before-and-after needs a claims review whichever way you make it.
  5. Write down what you'll do with a winner before you start (step 7).
What it looks like (illustrative)A budgeting app wants to know whether 'stop overdrafting' or 'see where your money goes' sells better. Six AI-avatar videos answer that in a week for the cost of one creator video. Once 'stop overdrafting' wins, a real creator who has used the app films the version that carries the budget.

The mistake that breaks it: Using AI avatars to fake the one thing UGC is for: a real customer's experience. It's the version most likely to get you a policy problem and the version viewers are best at spotting.

In the free PDF: Fit checklist: 10 questions that sort an angle into AI UGC or real creator.Get the PDF ↓

Use AI to find out what to say. Use people to say it when trust is the thing you're selling.Our rule of thumb

02 / 07 · SCRIPT

2. Write the hook before you pick the avatar

0–2sHook

Name the viewer or their problem in their own words.

2–6sProblem

One specific moment the viewer recognizes.

6–15sProduct moment

Show it working: screen recording, demo clip or b-roll.

15–22sProof

A fact you can back up: feature, price, guarantee, review count.

22–30sAsk

One action and the offer.

The 5-line script. About 70 to 80 words read at a normal pace.

Avatar tools make the face the fun part, so teams start there. The script decides the result. Designrr's agency, for example, built its batch around 20 to 30 hooks on the same core message.

Keep each video to one message and five beats: hook, problem, product moment, proof, ask. At 25 to 30 seconds that's about 70 to 80 words. Longer scripts make avatar delivery flatter and give the viewer more time to notice it isn't a person.

Write the hook in words your viewer would use. A good hook names who it's for or the exact moment the problem happens. 'Busy professionals' is a category. 'Rewriting the same email three times' is a moment.

Put the product on screen as footage: a screen recording, a demo clip, b-roll of the real thing. Avatars are good at talking to camera and weak at holding, pouring, applying or wearing a product.

How to set it up

  1. Write 10 hooks for one message. Use six types: callout, number, question, 'stop doing X', demo-first and objection.
  2. Write one body (problem, product moment, proof, ask) that works under all ten.
  3. Read every line out loud. Cut anything you'd never say to a friend.
  4. Check every proof line against something you can show: a feature, a price, a policy, a public review count.
  5. Only now pick an avatar that fits the viewer the hook is talking to.
What it looks like (illustrative)Hook: 'If you've ever rewritten the same email three times…' Problem: a 40-minute inbox morning. Product moment: screen recording of a reply written by voice in 20 seconds. Proof: 'Works in Gmail, Slack and Notion.' Ask: 'Try it free for 14 days.'

The mistake that breaks it: Ten avatars reading one script. You'll learn which face Meta liked this week, not which message sells.

In the free PDF: Script template, hook starters in six types and a proof-line check.Get the PDF ↓

03 / 07 · RULES

3. Cast the avatar as a presenter, never a customer

OK: presenter"This app turns a podcast into a blog post. Here's what that looks like, start to finish."
Not OK: fake customer"I tried it for 30 days and doubled my traffic."

Same avatar, two scripts. The second one describes an experience that never happened.

In the US, the FTC's 2024 rule on fake reviews and testimonials bans testimonials that misrepresent who is speaking, and it names AI-generated fake reviews specifically. It also covers testimonials from people with no real experience of the product. The FTC can seek civil penalties against people who break it knowingly.

An AI avatar saying 'I tried this and it worked' is close to the definition. It describes an experience no one had. Some vendor case studies show avatars framed as personas, like a podcaster or a course creator talking about the product. Write those as presenters explaining, not as users reporting results.

The safe pattern is simple. The avatar talks about the product the way a host or a spokesperson would: what it is, how it works, what it costs, what the guarantee is. Real experiences come from real customers, with their permission, and real creators who used the product.

This isn't legal advice, and other countries have their own rules. If your category is regulated, run every script past whoever signs off your claims.

How to set it up

  1. Ban first-person experience lines for avatars: 'I tried', 'I used', 'my results', 'changed my life'.
  2. Write avatar scripts in second person ('you') or about the product ('it').
  3. Keep every number in the script traceable to a source you could show a regulator.
  4. Use real reviews only as on-screen text with the reviewer's permission, never read out as the avatar's own story.
  5. Keep a log of which ads use AI presenters, for disclosure and for the review in step 6.
What it looks like (illustrative)An avatar can say: 'Here's how the subscription works, and here's what's in the box.' It shouldn't say: 'I've been using this for a month and my skin has never been better.' The first is a presenter. The second is a testimonial from someone who doesn't exist.

The mistake that breaks it: An avatar presented as a customer with results. It's the pattern the FTC rule targets, and it's the one the scam accounts reported by Media Matters relied on.

In the free PDF: Line-check sheet: phrases avatars can and can't say, with rewrites.Get the PDF ↓

dependables.AI UGC video adsFREE PLAYBOOK · PDF
Free PDF

Get the full playbook, with every template

Everything on this page, plus the working pieces you can copy straight into your program:

  • Fit checklist
  • Script template (25–30 seconds)
  • Avatar line-check sheet
  • Variant matrix
  • Test plan
  • Disclosure checklist
  • Creator hand-off brief

We'll email you about creator programs now and then. Unsubscribe any time.

04 / 07 · VARIANTS

4. Change one thing per variant

Avatar AAvatar BAvatar C
Hook 1Round 1
Hook 2Round 1
Hook 3Round 1
Hook 4Round 1
WinnerRound 2Round 2Round 2

Round 1 changes only the hook. Round 2 keeps the winning hook and changes only the avatar.

Cheap production tempts you to change everything at once: new face, new script, new music, new offer. Then one ad wins and nobody knows why.

Teknasyon's approach, as Arcads describes it, was to find a winning ad type first and only then multiply it with different avatars and scripts. Designrr's agency held the core message steady and changed the hook. Both kept most of the ad fixed.

Order matters. The hook usually moves results the most, so test it first. Then the presenter, because a face changes who stops scrolling. Then the body and the offer.

Keep variants different enough to count. Platforms tend to group near-identical ads, so five hooks that all start with the same three words won't give you five real tests.

How to set it up

  1. Pick one message and write four or five truly different hooks for it.
  2. Make Round 1 with one avatar, one body and one edit style.
  3. Name every file with the variable it tests: msg-hook2-avatarA-body1.
  4. Carry the winner into Round 2 and change only the avatar.
  5. Keep a running sheet of what won each round and why you think it did.
What it looks like (illustrative)Round 1: four hooks, one avatar, one body. Hook 3 wins on cost per lead. Round 2: hook 3 with three different avatars. Round 3: the winning hook and avatar with two offers, a free trial and a discount.

The mistake that breaks it: Changing the avatar and the hook together. You'll scale a face that only won because of its script, or kill a good hook because of its face.

In the free PDF: Variant matrix and a file-naming convention.Get the PDF ↓

05 / 07 · TESTING

5. Set up the test so it can tell you something

MetricWhat it tells youUse it to
Hook rate3-second views ÷ impressionsPick openings, not winners
Hold rateHow many watch to the product momentFix the middle of the script
CTRClicks ÷ impressionsCheck the ask and the offer
CPASpend ÷ purchases or leadsDecide what gets more budget

Judge on cost per result. The other three tell you which part of the script to fix.

A test needs a question, a fixed setup and a decision rule. Write all three before anything goes live.

Judge on the result you pay for: cost per purchase or cost per lead. Hook rate is a diagnostic. Creatify's write-up on DeNA's mobile game tests is a good reminder: the creative that converted best had a lower hook rate than the two other leaders. DeNA ran rounds of 10 to 12 variants changing hooks, visuals and angles. The source doesn't say whether avatars were used.

Run a control. Put your current best ad, ideally a real-creator ad, in the same ad set. Without it you can't tell whether AI UGC beat your account or just beat a bad week.

Give each round enough spend to reach a decision, and stop rounds on a schedule, not on a mood. Our rule of thumb is a week per round and at least a few dozen results per ad before calling a winner.

How to set it up

  1. Write the round's question in one sentence: 'Which of these four hooks gets the lowest cost per lead?'
  2. Use one ad set per round with the same audience, budget and placements.
  3. Add your current best ad as the control.
  4. Run for a fixed window and don't edit ads mid-round.
  5. Pick the winner on cost per result. Use hook, hold and click rates to plan the next round.
What it looks like (illustrative)One test campaign, one ad set per round, four or five ads per ad set, the same audience and budget for each round, seven days per round. Decide winners on cost per lead. Use hook rate only to explain why a loser lost.

The mistake that breaks it: Declaring a winner on hook rate or CTR after two days. The ad that stops the most thumbs isn't always the one that sells.

In the free PDF: Test plan template with a decision rule for each round.Get the PDF ↓

06 / 07 · DISCLOSURE

6. Label it where the platform asks

WhereWhat it asks of you
MetaAds made with Meta's own AI tools get an "AI info" label, placed next to "Sponsored" when they show a photorealistic AI person. Meta has started detecting third-party AI ads too. Political and social-issue ads must disclose realistic AI people or events.
TikTokNon-Spark ads that are fully or largely AI-generated must switch on the AI-generated content disclosure. Symphony avatars are labeled automatically.
Google / YouTubeElection ads must disclose realistic synthetic people or events. YouTube creators must label realistic AI content.
FTC (US)Testimonials from people who don't exist, including AI-generated ones, are banned under the 2024 rule.

Summary as of Oct 23, 2026. Rules change; check each link in the sources before launch.

Meta labels ads made or heavily edited with its own gen-AI tools as "AI info". When an ad shows a photorealistic AI person, that label sits next to "Sponsored" instead of being tucked into a menu. As of June 2026 Meta says it has started detecting ads made with third-party AI tools and labels those in "About this ad". Political, election and social-issue ads must disclose realistic AI-made people, events or audio.

On TikTok, non-Spark ads whose video, image or audio is fully AI-generated or significantly changed by AI must turn on the toggle "This ad contains AI-generated content". Ads made with Symphony's avatars are labeled automatically.

Google requires election advertisers to disclose realistic synthetic content, and YouTube asks creators to label realistic AI-made content. Neither changes the FTC rule from step 3, which applies whatever the platform says.

One more change worth knowing: Adweek reported in October 2026 that Meta's updated commercial terms make advertisers responsible for what Meta's own AI features generate or change in their ads. Check your ads after any automatic 'enhancement'.

How to set it up

  1. Keep a column in your ad sheet: AI presenter yes/no, tool used.
  2. Switch on TikTok's AI disclosure for every non-Spark ad with an AI presenter.
  3. Treat Meta's label as expected, and never edit an ad to dodge detection.
  4. For political or social-issue ads, disclose in the ad setup every time.
  5. Review Meta's automatic creative enhancements before and after launch.
What it looks like (illustrative)A TikTok ad with a stock avatar made outside TikTok: switch on the AI-generated content disclosure in Ads Manager. The same video as a Meta ad: expect an 'AI info' label if Meta's tools or detection flag it, and don't fight it. If it runs as a social-issue ad: disclose it yourself.

The mistake that breaks it: Treating the label as a performance problem to engineer around. Hiding it risks the ad account, and the ads that need hiding are usually the ones faking a customer.

In the free PDF: Disclosure checklist for Meta, TikTok and Google, with links.Get the PDF ↓

07 / 07 · HAND-OFF

7. Hand the winners to real creators

AI UGC20–50 cheap variants
→
Winning anglehook + message + offer
→
Real creatorssame angle, real hands, real voice
→
Both runAI keeps testing, creators carry spend

AI finds the angle. People earn the trust.

Once AI UGC has found an angle, you know something real: which message and hook your buyers respond to. That's the most expensive thing to learn with creators, and the cheapest to learn with avatars.

Then give the angle to people. A real creator can show the product in hand, tell a true story about using it and answer the objection in their own words. That's the trust layer AI UGC can't supply and the FTC rule says it mustn't fake.

Digiday's reporting points the same way: agencies say demand for creators who feel authentic and a bit messy went up after a flood of AI content, and some clients ban AI creators outright or require disclosure.

Run both. Keep a small AI UGC budget testing new angles every week, and put most of the spend behind creator versions of the winners. If a creator version loses to the AI version, test the creator's hook before you give up on creators.

How to set it up

  1. Turn each winning angle into a one-page creator brief: the hook, the message and what to show on camera.
  2. Brief two or three creators who actually use products like yours.
  3. Ask for the hook word for word plus two of their own.
  4. Run creator versions against the AI winner in the same ad set.
  5. Keep a standing AI UGC budget for next month's angles.
What it looks like (illustrative)AI UGC finds that 'packing for a week in one carry-on' beats every other angle for a luggage brand. Three creators who travel film it with the real bag: the zip, the fold, the overhead bin. The AI version keeps running in a test ad set while the creator versions take most of the budget.

The mistake that breaks it: Treating AI UGC as a replacement for creators. You'll save on production and lose the ads that carry trust, which are usually the ones that scale.

In the free PDF: Creator hand-off brief template (one page).Get the PDF ↓

THE HONEST PART

What it takes to run this yourself

The software is the cheap part. A steady AI UGC program needs the same roles as any creative testing program, plus a claims check. Here's who covers what.

ScriptwriterWrites the hooks and bodies, and keeps every line inside what you can prove.
ProducerRuns the avatar tool, picks presenters and cuts in real product footage.
Media buyerSets up rounds, keeps the control running and calls winners on cost per result.
Claims reviewerChecks scripts against the FTC rule, your category's rules and platform labels.
Creator leadTurns winning angles into briefs and gets real creators filming them.

If you have those roles covered, run the playbook. The last step is where we come in: Creator Programs puts real creators on your winning angles, with sourcing, briefs, contracts and usage rights handled by one team.

Book a call

dependables.AI UGC video adsFREE PLAYBOOK · PDF
Free PDF

Get the full playbook, with every template

Everything on this page, plus the working pieces you can copy straight into your program:

  • Fit checklist
  • Script template (25–30 seconds)
  • Avatar line-check sheet
  • Variant matrix
  • Test plan
  • Disclosure checklist
  • Creator hand-off brief

We'll email you about creator programs now and then. Unsubscribe any time.

Questions

Is AI UGC allowed on Meta and TikTok?

Yes, within each platform's ad policies. Meta labels some AI-made ads as "AI info" and is starting to detect third-party AI ads. TikTok requires an AI-generated content disclosure on qualifying non-Spark ads. Political and social-issue ads have stricter disclosure rules on both.

Can an AI avatar give a testimonial?

In the US, a testimonial from someone who doesn't exist breaks the FTC's 2024 rule, which names AI-generated reviews. Script avatars as presenters who explain the product, and leave experiences to real customers and creators.

Which tool should I use?

The examples on this page used Arcads, Creatify, HeyGen and TikTok Symphony. We're not affiliated with any of them and don't recommend one over another. Pick the one whose avatars suit your audience and whose terms cover paid ads in your markets.

Are the results in the examples typical?

We can't say. Nearly all of them come from the vendor's or the platform's own case studies, without full methods. Treat them as proof that brands use it, not as a forecast.

How many variants do I need?

Our rule of thumb: four or five per round, one variable per round, a week per round. Designrr's agency made 150+ across formats, but that was a two-week sprint with a large budget.

Will AI UGC replace creators?

We don't think so. It's good at finding the angle. Real creators are still how you show a product in real hands and earn trust, and that's where most of the budget tends to go once you know what works.

What's in the PDF?

The seven steps plus the working pieces: the fit checklist, script template with hook starters, avatar line-check sheet, variant matrix, test plan, disclosure checklist and the creator hand-off brief.

Can you run this for us?

Creator Programs is the real-creator side: sourcing, vetting, briefs, contracts, usage rights and content delivered ready to run, for $5,000 a month. Book a call and we'll talk through where AI UGC fits next to it.

Creator Programs by dependables.Free PDFBook a call