How to make a UGC-style video ad with AI, step by step

Plenty of tools will turn a product into a talking-person video in a few clicks. The harder part is the thinking before the clicks: what you may claim, what the video can actually show, and what you test first. This guide is the sequence I would follow, from claims to script to render to review.

It is a how-to, so it does not repeat other pages. For what a creator or an AI ad costs, read cheap UGC ads. For whether to use a real creator or an AI presenter at all, read real creators vs AI avatars. This page assumes you have decided to try an AI presenter and want the steps.

This page may contain affiliate links from which Blur Studio earns commissions.

What "UGC-style" may and may not claim

"UGC-style" describes a look: a person in a normal room speaking to the camera in plain words. It does not describe who is speaking. If the person in your ad is an AI presenter, the video is a demonstration or an explainer in a casual style. It is not a customer story.

The US Federal Trade Commission (FTC), the regulator for advertising in the US, finalised a rule in August 2024 that bans fake or false reviews and testimonials. Its announcement says this includes AI-generated ones attributed to people who do not exist or who had no real experience with the product, and that knowing violators can face civil penalties. Its Endorsement Guides add that an endorsement must reflect the honest experience of the person giving it. The FTC's questions and answers page for the rule adds a useful detail: the rule has no blanket ban on AI-generated avatars in marketing, and an avatar's message is a problem when the testimonial behind it is fake or false. It also warns that avatars can still be deceptive under the FTC Act and the Endorsement Guides. So my plain reading is this: an AI face that explains your product is allowed, but a line such as "I have used this for three months" spoken by a made-up face is a testimonial from nobody. This is information, not legal advice. Check the current FTC pages and local law.

So the safe pattern is this:

• Allowed in spirit: "Here is how it works." "People ask us about this." "This is the part most people miss."
• Not allowed: "I lost 5 kilos with this." "My skin has never been better." "I switched from my old supplier and saved a fortune."
• If you want real customer proof, use a real customer's own recording, with their written permission and a clear note of anything you gave them for free.

Before you start: the claims table

Magic Hour, a vendor of AI video tools, gives good advice on script preparation: remove any line your footage cannot show. I would go one step earlier and list every claim in a table before writing a word of script.

Claim	Evidence you hold	Where the video shows it	Keep or cut
"Ready in 10 minutes"	Your own timed test	On-screen timer, real footage	Keep
"Fits all phones"	None	Nowhere	Cut
"Loved by customers"	14 reviews with names and dates	Screenshot of real reviews	Keep, with the source
"I used this for a month"	None, the speaker is AI	Nowhere	Cut

The four rows above are made up to show the format. Your table should be about eight to ten rows. Any health claim, money-back, speed or comparison claim needs evidence you can show an advertising reviewer. If a claim has nothing in the second column, it does not go in the script.

Script shape: hook, problem, proof, benefit, action

Magic Hour's seven-step workflow uses this shape (hook, problem, visible proof, supported benefit, action). Aim for about 20 seconds, which is roughly 50 spoken words.

1. Hook (about 3 seconds). Name the problem or the surprise. "Most people store their invoices in three places."
2. Problem (about 4 seconds). One sentence on what it costs them. "Then they lose an hour at month end."
3. Proof (about 6 seconds). Something the viewer can see: the product in use, a screen, a before and after you actually made. Put your real footage here, not a claim.
4. Benefit (about 4 seconds). One benefit that your claims table supports.
5. Action (about 3 seconds). One thing to do. "Link in bio" or "Try the free version."

A made-up example for a small bookkeeping service:

• Hook: "Three apps for one set of invoices?"
• Problem: "That is why month end takes all afternoon."
• Proof: (screen recording of a real invoice being sorted)
• Benefit: "We keep them in one place so you find any invoice in seconds."
• Action: "Book a free 15 minute call. Link below."

Notice what is missing: no customer story, no invented result.

Step by step in Higgsfield Marketing Studio

Higgsfield's Marketing Studio is its workspace for ads. Its help center describes a template-first path (you pick a ready format, such as a UGC template, and connect your product), and its blog says clips run up to 15 seconds with audio and lip sync in one pass. The tool itself is covered in the Marketing Studio guide. Here is the order of work.

1. Prepare the product. Use clear product photos or a product page link. Check that the label, colour and shape imported correctly before you do anything else.
2. Choose the format. Pick a UGC-style video mode. Then pick a presenter from the avatar library, which the platform describes as a collection of ready-made AI presenters, or generate one. Higgsfield's blog also says you can upload a photo of a person whose face you own or have permission to use. Reuse the same presenter for a whole campaign.
3. Pick the setting and shape. A kitchen, desk or shop counter that fits the product. Choose vertical 9:16 for TikTok, Reels and Shorts before you generate, since a new shape is a new generation.
4. Paste the script. Only the words the presenter says. See the next section.
5. Look at the price on the Generate button. Higgsfield does not publish a fixed credit cost for Marketing Studio video, so the button is the only reliable figure. Higgsfield's own blog gives a range per video, but the pricing page does not publish one, so I do not quote it. Measure your own cost per approved ad instead.
6. Make one diagnostic render. One script, one presenter, one setting. Do not batch yet.
7. Review it with the checklist below. Fix the script or settings and render once more.
8. Then batch. Once one render passes, make the variants described in the test section.
9. Add captions after the audio is final. Captions made before a script edit will not match.

Keep spoken words and actions apart

Most tools take two inputs: what the person says and what the person does. Magic Hour's guide says to keep them in separate fields. If you mix them ("she smiles and says it only takes ten minutes while holding the box up"), the tool may speak the stage direction aloud or ignore half of it.

• Speech field: only the exact words. No brackets, no directions.
• Action field: "Holds the product at chest height with the label facing the camera. Nods once. Smiles at the end."

Keep actions simple. Hands holding products are the usual failure point, and the product label is the next. If the action is hard, use a product shot with a voice-over instead. The free alternative is also worth remembering: an owner or staff member filming a 20 second phone clip from the same script costs nothing and carries no AI risk.

Review checklist before anything goes live

• Label and logo. Pause at three points. Is every letter of the product label correct and readable?
• Hands and fingers. Count them. Look at where the product meets the hand.
• Mouth and voice. Do lips match the words? Does the voice sound the same all the way through?
• Claims. Read the final spoken words against the claims table. Anything not in the table is cut.
• No fake customer. Search the script for "I used", "I bought", "my results", "I switched". All of them go.
• Captions. Do they match the final audio? Are spelling and prices right?
• Safe area. Is text clear of the top and bottom strips where the app puts its buttons? Meta publishes specific safe zones for Reels ads, so check them.
• Sound off. Does the ad still make sense with the sound off?

For a wider quality routine, see avoiding the AI look.

Disclosure by platform

This is information, not legal advice, and rules change. Open the pages below on the day you publish.

• TikTok. Its support page says a label is required on realistic AI-generated content. An AI presenter who looks like a real person falls under that. TikTok also applies labels automatically in some cases.
• Meta (Facebook and Instagram). Meta applies its own "AI info" label for content made with its tools or detected through industry signals. Meta's June 2026 ads-transparency post mentions "Paid for by" disclosures for social-issue, election and political ads, and separate rules cover altered media in those ads (check Meta's policy page). Do not read that as permission to hide AI use. Your duty not to mislead under advertising law is separate from the label.
• New York. A synthetic performer ad disclosure law took effect on 9 June 2026. Per the text of bill S8420-A on the New York Senate site, an ad must disclose when the advertiser knows a synthetic performer is in it, with a $1,000 penalty for a first violation and $5,000 for each later one. If you sell to New York audiences, check how it applies to you.
• European Union. Article 50 of the EU AI Act applies from 2 August 2026, and requires deployers of deepfake image, audio and video to disclose that the content is artificially generated or manipulated. Check what applies to a realistic presenter in your market.
• Disclosure of paid relationships. If a real person promotes your product and you paid or gave them free product, the FTC says material connections must be disclosed clearly and conspicuously.

The label question and the truthfulness question are separate. A clear AI label on an ad where the AI face claims to be a happy customer still leaves a fake testimonial.

Test: one variable at a time

Higgsfield's own guide suggests changing only the opening line when you test hooks, and I agree. Here is a plan you can run:

1. Write three hooks for one script. Keep everything after the hook the same.
2. Render all three with the same presenter and setting. Three renders is a small, affordable number.
3. Run them at equal small budgets for the same days to the same audience.
4. Judge by the result that matters to you, such as purchases or booked calls, not only views or watch time. Magic Hour makes the same point.
5. Keep the winning hook. Next time change the proof beat, then the presenter.

Set the pass and fail rule before you spend: for example, "the winner must beat the others on cost per booked call by a clear margin, or we stop." For deeper hook work see ad hook tests, and for scaling to many versions without chaos see how to test 20 ad variations. No independent benchmark showing that AI UGC beats or matches real UGC is known. Only vendors assert it, so let your own test decide.

Who should skip this

• Anyone who needs real proof. Health, finance, safety, or any product where trust is the sale: use real customers and real creators.
• Physical products that must be handled on camera. Hands and labels are where AI presenters fail most often, so film it for real.
• Owners who already have a willing customer or staff member. A phone video from a real person beats an AI presenter for trust and costs nothing.
• Anyone starting from a product URL who wants a first draft fast. Arcads and Creatify are built around that, and I compare them with Higgsfield in this comparison. A single-purpose tool may be simpler for you, and Arcads alternatives with prices you can check lists some.

Higgsfield is a weaker fit if you only want one avatar video a month, and you can see the Marketing Studio plans on its site before deciding. Its value shows when you test several versions and want other formats in the same account. Subscription credits expire at the end of each billing cycle, so check the credits guide before you commit.

How I'd approach this

I would write the claims table first and show it to whoever signs off the ad. Then I would write one 20 second script, make one test render, and run the checklist. Only if that passes would I make three hook versions. I would not let any presenter, human or AI, say they used the product unless a real person did.

If you make these ads for several products or clients, the repeatable part is the claims table, the script template and the review step. Blur Studio builds that kind of approval workflow and shared asset library around these tools, so a new ad starts from a checked process instead of a blank page.