AI Video Ads for Pharma and Regulated Brands: Faces, MLR, and What Is Actually Allowed

What is actually allowed in an AI video ad for a pharma or regulated brand? The same things that were allowed before AI, checked by the same people: your MLR team still reviews every claim, every reference, and every super, and nobody gets to put a real person's face in your ad without that person's consent. AI does not exempt a single frame from that review, and that review does not forbid a single frame from being generated. What changes is the production discipline underneath: whose faces appear, where the references live, how the supers get built into the master, and whether every version can be audited a year from now.

We are Filmito, a film studio in Ahmedabad, in business since 2017 and AI-powered since 2023. We have shipped pharma and medtech films, and we keep a short list of things we will not do in this category. We are not your counsel or your regulatory team; they get the final word on every frame.

The rules did not change. The production did.

On a live shoot, compliance risk lives in two places: the script and the edit. What the camera saw is what you approved. With AI generation a third place appears: the generation itself. A model can invent a face, a label, a logo, a line of text, a device shape that does not exist, and none of it was approved by anyone. That is the real difference, and it is a production problem before it is a legal one.

Our method is called HumanXAI: humans direct, AI produces, and every frame is signed off by a person before it leaves the studio. In most categories that sign-off is about taste. In this one it is the control point.

Faces: why we refuse recognizable people, and put it in writing

Two rules, no exceptions. First, no recognizable real faces: a real face in a pharma ad implies a real patient or clinician vouching for the product, and that needs a release, an approved script, and often more. Second, no generated person derived from a real individual. Both are warranted in our contract, so your legal team is not relying on a promise in an email.

The second rule is the one people miss. Some engines can take a photograph of a real person and generate someone who is almost them: same bone structure, different hair, slightly different age. That is a worse position than using the real photo, because there is no release and no clear line about who the character is. So we never point an engine at a real person. Our characters are built from written descriptions, then locked as a reusable digital cast member so one face holds across every film in the campaign. One of our engines only accepts real photographs as references; in regulated work we aim that at packaging, devices, and locations, never at a person.

The honest tradeoff: if your concept needs a real key opinion leader on camera or a real patient telling a real story, we are not your vendor. That is a live shoot, and it should be.

MLR-aware scripting: writing for the review, not around it

Most compliance pain in video comes from scripts written as if the review were a formality. We write the other way round. The script goes to you as a package a reviewer can read without watching anything, and we do not generate a frame until it is stable.

• Every claim carries its reference number in the script document, not in a separate email.
• Supers are written into the script with their planned on-screen duration, so timing is decided before the edit.
• Required statements and safety information get their own lines and their own seconds.
• Each scene carries a note on what the engine will not be asked to render: no labels, no text, no faces of anyone real.

One practical detail: engine moderation can refuse a prompt without saying why. We run a free preflight check on every prompt before any generation is charged, so a refused scene gets rewritten at the script stage rather than discovered as a gap in the edit.

References, supers, and versions: what the master actually contains

Here is a limitation that turns into an advantage in this genre. On-screen text inside generated footage garbles, and no amount of prompting fixes it reliably. So we generate clean plates and add all text in post. For a pharma master that means every reference, super, and required statement is typeset in the edit, spelled correctly, sized to your standard, and kept as an editable layer in the master file. Nothing your reviewer cares about is burned into a generation nobody can change.

Compliance edits from MLR are included in every revision round. A changed reference number, a super that needs two more seconds, a statement that has to move earlier: those are not change requests we push back on, they are the job.

Versioning is where the audit trail comes from. Every cut is numbered, cut 1.0, cut 2.0, cut 3.0, with change notes attached, so a reviewer a year from now can see which super changed in which version and why. Our Delivery Report, computed from our own archive this August, covers 742 projects with versioned cuts: the median approval came at cut 3.0, 66.2 percent were approved by cut 3 (the first cut plus the two included revision rounds), and 3.6 percent took nine or more cuts. The breakdown is on filmito.io/en/ai-ads-report. One caution: we have not split those figures by industry, so treat the median as a studio-wide number, not a pharma promise. Plan for more rounds, not fewer.

“Anyone can generate with AI. We direct it.”

What AI is strong and weak at in this genre

Strong: environment and mood. A patient's morning, a pharmacy at dusk, a hospital corridor at scale, a visual metaphor for relief or for a condition, all of these generate well and need no location, crew, or permit. Strong too: editions. Vertical cuts and additional languages are straightforward once the master exists; across our archive, 454 delivered files are vertical or short-form editions and 33 projects shipped in more than one language. Two of our films in this category are public: ACIK SafePass, a pharma and safety explainer, and RescueIQ, a medtech film, both on filmito.io/en/ai-video-ad-examples.

Weak, and we say so before you sign:

• Text in frame. It goes in post, always.
• Exact device geometry and anatomy. An engine will give you a plausible inhaler, not your inhaler. If the shape of the product is the message, that shot is a product photograph or 3D, not generation.
• Long continuous action. Single takes hold reliably for roughly five to eight seconds, so films are cut from short takes. We cannot generate a ninety-second uninterrupted demonstration.
• Precise hands. Injection technique, dosing, device assembly: the engine gets fingers wrong often enough that we do not put those shots in front of a reviewer.

When we are not the right choice, and what it costs when we are

Not us: a real clinician or patient on camera. Not us: instructions-for-use or mechanism-of-action films where every detail has to be anatomically and mechanically correct; a 3D animation team will serve you better. Not us: a process that needs the vendor working inside your review system with in-house regulatory affairs staff. We are a small studio. We deliver versioned masters with change notes and editable text layers, and your team runs the system.

If the work is a brand film, a disease-awareness piece, an explainer, or a campaign of short editions with the references and supers handled properly, that is what we are built for. Our prices are the same for every client and public on filmito.io/en/pricing: a Launch one-off ad is $595 for 15 seconds, $895 for 30, $1,495 for 60, first cut in three days, two revision rounds included, full usage rights transferred on delivery. MLR compliance edits are included inside those rounds. If your review will need more than two rounds, say so at the quote stage so the price reflects it.

To see whether your brief fits, send it through filmito.io/en/see-your-ad. We reply within one business day with the idea, the price, and the date. If you need a live shoot or a 3D team instead, we will say so.