🧾 AI Video Ads Freelancer: Pricing, Rights and Revisions

Search the phrase "ai video ads freelancer" and Google hands back Freelancer.com project listings, Upwork hire pages and portfolio profiles. No editorial page holds the keyword. The guides that rank nearby are almost all published by companies that sell the generation tools, which gives them a quiet conflict: they profit when you believe this service business is frictionless.

This post is about invoicing a client for ad creative. That is a different business from earning platform ad revenue on your own channel, which has different cash flow and different rules; we cover that side in the faceless channel monetization guide. Here the money comes out of somebody's marketing budget, and everything downstream follows from that one fact.

By the numbers#

• $0.40 per second for Veo 3.1 at 720p and 1080p, $0.60 at 4K, and $0.10 on Veo 3.1 Fast at 720p, per Google AI for Developers. Thirty seconds of kept footage at standard rates is about $12 of generation before a single re-roll
• $150-400, $400-1,000 and $1,000-3,500+ per 30-second video across beginner, mid-level and expert tiers, plus $50-150+ per finished minute and retainers of $1,500-20,000 a month, per Playcut's rate guide
• +30-50% per 30-day paid flight, +50-100% for exclusivity, +100-200% for a perpetual buyout, also from Playcut. That ladder is priced on top of an asset whose ownership the page never establishes
• October 21, 2024: the FTC's Rule on the Use of Consumer Reviews and Testimonials took effect, expressly covering "AI-generated reviews and testimonials" and content that misrepresents "the identity, experience, or existence of the reviewer", with civil penalties available, per Sidley Austin's analysis
• 2 August 2026: EU AI Act Article 50 transparency obligations became applicable. Providers must add machine-readable marks so synthetic content can be detected, and deployers must tell people when they are looking at a deepfake, per the European Commission
• June 1, 2026: Meta began using automated detection to find undisclosed AI content in ads and apply an "AI Info" label, on top of its existing rule that repeated failure to disclose "may result in penalties against the advertiser", per the Meta Transparency Center

What clients actually buy#

The person signing your invoice is usually a performance marketer at an ecommerce or DTC brand. They are not commissioning art. They are feeding an ad account that eats creative, and their problem is that they need twelve testable hooks this month and their agency quoted three weeks for two.

So the product is throughput. One approved body, twelve hook openers cut against it, three aspect ratios, captions on and off. That is what gets bought again next month, because it maps directly onto the line item the budget comes from: creative testing.

The work AI freelancers enjoy making is the sixty-second cinematic brand film. It sells once, to a founder, and then never again. Learn to like making variants.

So what: quote in clips and hook counts, never in hours, because throughput is the thing being purchased.

The AI testimonial trap#

The single most-requested format in this niche is a synthetic person looking into a phone camera saying a product changed their life. Fiverr sells it for $10. Every competing guide lists it as a core offer. It is also the format most likely to put your client in front of a regulator.

The FTC's rule, in force since October 21, 2024, bars creating, promoting or disseminating fake consumer reviews and testimonials, and it names AI-generated ones. The test is misrepresentation of the identity, experience or existence of the reviewer. A fabricated customer fails all three. Adding an "AI-generated" label does not cure it, because the deception is the endorsement itself and not the rendering technique. The rule also reaches "purported individual accounts not associated with a real individual", which is worth reading twice before you accept a brief for a fake creator persona.

What stays on the safe side of the line: AI b-roll and product demonstration, animated explainers, synthetic voiceover, motion graphics, and AI presenters who are presented as presenters or brand characters rather than as customers. Real testimonial footage from real customers, cleaned up and cut with AI tooling, is also fine, and the consent paperwork travels with it. The two tools clients actually ask for in this format are compared, testimonial realism against production-line pricing, in our Arcads vs Creatify breakdown.

So what: the highest-demand format in your inbox is the one to price with a written refusal, not a rush fee.

Pricing from the render cost up#

Retainer pricing for this work gets pitched a lot. Watch one of those pitches, then hold it against the render-cost maths above — the revenue framing is real, the cost side usually is not shown:

Start with cost of goods, because the guides quoting "90%+ profit margins" get there by pretending re-rolls do not exist.

A 30-second spot is typically six clips of about five seconds. Thirty seconds of kept footage at $0.40 per second is roughly $12. Now add the multiplier that nobody publishes: on ad-grade work with a brand's actual product in frame, we budget four to six generations for every clip we keep. That turns $12 into $50-70 of generation for one spot, before voice, music, stock, or a minute of your time. Add three to five hours of scripting, direction and edit labour and the low end of the beginner band stops looking like a margin at all.

Two rules fall out of that. Price per deliverable or per package, never per hour, so re-roll risk sits inside a number you chose. And charge separately for volume: the second, third and twelfth hook variant cost you far less than the first because the body is already approved, which is exactly why a hook pack is the offer with a defensible margin.

For the retainer, tie it to testing cadence rather than a video count: a fixed number of new concepts plus a fixed number of variants per month, with rights and disclosure paperwork included. That is the structure that survives a client's quarterly budget review. Our per-model cost breakdown has the current per-second numbers if you want to rebuild this math for a different stack.

So what: if you cannot state your cost per finished 30 seconds within ten dollars, you are not pricing, you are guessing.

The revision trap is a rendering problem#

Every competing guide treats revisions as client management. It is not. It is a property of the tooling.

Generative video is non-deterministic. Run the same prompt twice and you get different framing, different pacing, different audio. A regeneration meant to fix one hand can come back with a different face, a rearranged room, a shifted colour palette and a new rhythm. There is no timeline to nudge; there is a dice roll to repeat.

Here is the mistake I paid for. I sold my first AI ad package with "3 rounds of revisions" copied straight off my old editing rate card. Round two was a four-word change to one line of voiceover. The regeneration came back with a visibly different presenter, which meant re-rolling the two clips either side of it for continuity, which meant re-grading the whole spot. About $180 of credits and two days, on my side of the ledger, to rebuild something the client had already approved.

The fix is to move every approval gate upstream of the expensive irreversible step, and to define a revision as a bounded change at a named gate rather than an open round on the final render.

Write it into the scope: revisions happen at the gate where the change originates, unlimited at the concept and script gates because text is cheap, one round at the keyframe and voice gates, and a post-render change that requires regeneration is a new deliverable at a stated percentage of the original rate. Clients accept this readily when you explain why, because "a re-render changes things you did not ask to change" is intuitive once someone says it out loud. Keyframe locking is the other half of the answer; the character consistency guide covers the techniques that make a gate hold.

What rights you can honestly grant#

The U.S. Copyright Office concluded that generative outputs are protected by copyright "only where a human author has determined sufficient expressive elements", and that this does not include "the mere provision of prompts". It also confirmed that using AI during creation, or including AI-generated material in a larger human-made work, does not bar copyrightability of the whole.

Read those two sentences together and the deliverables clause writes itself. Inside the finished ad there is real human authorship: the script, the shot selection, the sequencing, the timing, the captions, the grade. That is assignable. The raw model output on its own may carry nothing to assign.

So structure it as three moves. Assign what you authored. Sublicense only what your tool terms let you sublicense. Warrant delivery and non-infringement of the material you supplied, rather than warranting exclusive ownership of every frame. Selling a client a "perpetual buyout" of something you may not own is how a rate-card upsell becomes a misrepresentation claim two years later.

So what: say what you are transferring in plain words, because a vague grant is the clause that gets litigated.

Three platform-terms traps that actually bite#

One ranking guide states that most platforms "grant you full commercial rights to the content you generate on their paid plans". That is wrong in the three places that matter for client work.

Tier gates. ElevenLabs' terms say that free users "may only use the Services for non-commercial purposes", while paid subscribers may use them commercially and retain rights in their output. A free-tier voiceover on a paid client job is a terms breach on day one.

Sublicensing limits. Synthesia licenses its stock avatars under a "non-sublicensable, non-transferable, non-exclusive, limited license". You can use that avatar in work you deliver, but you cannot pass the avatar right through to the client, which is precisely what a buyout clause claims to do. If a client wants avatar rights they can own, that is an argument for a custom avatar or a different vendor; the Synthesia alternatives comparison sets out the trade-offs.

Subscription expiry. That same Synthesia license runs "during the term of an applicable subscription". Cancel in December and the client's January flight is running on a lapsed right. Nobody notices until someone does.

So what: before you quote, read the licence tier you will actually generate on, not the marketing page.

Whose disclosure is it, anyway?#

This is the part every competitor skips, because the duty lands on the client's ad account rather than yours.

Meta requires advertisers to disclose when a social issue, elections or political ad contains photorealistic AI imagery or realistic synthetic audio, and says it will reject undisclosed ads with penalties for repeat failure. From June 1, 2026 it also detects and labels automatically. TikTok requires AI-generated content labelling on paid and branded content. YouTube requires disclosure of realistic altered or synthetic content, with a specific carve-out that beauty filters, colour and lighting adjustment, and "cloning one's own voice to create voice overs or dubs" do not need one. And anything running in the EU now sits under Article 50, applicable since 2 August 2026, with machine-readable marking on synthetic output and deployer disclosure for deepfakes; adherence to the Code of Practice on Transparency of AI-generated Content is the recognised route, and anyone declining it must use "alternative equivalently adequate means". We unpack the marking side in the EU AI Act watermarking guide.

Your job is not to file any of this. Your job is to hand over the facts: which model produced which shot, whether any real person's likeness or voice appears, whether the audio is synthetic, and whether provenance metadata survived your export. Then the contract names who files. Put that sentence in every agreement and you have turned a liability into a selling point. If the brief is a recurring branded character rather than a one-off ad, the New York and EU disclosure regimes covered in our AI virtual influencer guide apply with even less room to skip them.

Likeness, voice and consent paperwork#

If a real face or a real voice is anywhere in the deliverable, the consent record is part of the product. Get it in writing, name the specific use, name the term and the territories, and keep it where you can find it in two years. Cloned voices need consent from the voice's owner, and the YouTube carve-out is a useful illustration of where the line falls: cloning your own voice for narration needs no disclosure, while making a real person appear to say something they did not say does. If you are building repeatable synthetic presenters, the AI avatar guide covers the consent questions each approach raises.

The deliverable spec clients expect#

Ship 9:16, 1:1 and 16:9 from one master, hook variants cut from a shared body, captions burned in and supplied separately as an SRT, and filenames that survive an ad-platform bulk upload. Agree upfront what "source files" means when the source is a prompt: our default handover is the edit project, exported masters, the prompt and seed log with model versions, and the voice tracks. Caption accuracy is worth a pass of its own, since burned-in errors cannot be patched after the render; the caption accuracy guide has the checks.

Competing with a $10 floor#

You cannot win on rendering. Somebody with the same subscription will always undercut you, and marketplaces price that work at ten dollars. What they cannot deliver is a package that survives a legal review: a compliant format choice, a defensible rights grant, a consent file, a disclosure sheet handed to the media buyer, and a revision structure that does not blow up the schedule.

That is the pitch. Not "AI videos, fast and cheap", but "ad creative your compliance team will sign off on, at testing volume". Brands with real budgets are already nervous about this category, and being the freelancer who removes that anxiety is worth more per deliverable than being the fastest renderer on the platform.

If you also want the platform-revenue side of AI video, that is a separate business with separate math, and the seven monetization models breakdown ranks it honestly. Client work pays this month. Channel work pays later, if it pays.