Who Produces AI Video for Fintech, Healthcare and Other Regulated Industries?

A marketing lead at an insurer or a digital-health company almost never asks whether a video can be produced with AI. They ask a harder question: who is going to be accountable for what it says. That is the correct question, and it is the one this guide answers. In a regulated category a video is not only a marketing asset. It is advertising that someone can complain about, compare against a disclosure document, and demand an approval record for.

This is an operational guide to AI video production for regulated industries — fintech and payments, investment and wealth platforms, insurance, digital health and medical devices. It is not a budget guide and not a tool review. If you want the planning layer — what to order and in what sequence — that lives in our guide to scoping an AI explainer video, and there is no point duplicating it here. What follows is what changes inside a production when the client is supervised: claims review, disclaimer placement, synthetic people, product-screen data, disclosure, and the audit trail that proves all of it.

What makes a marketing video “regulated”

It is not the size of the company. A two-year-old payments startup is held to the same claims discipline as a bank, and a wellness app that promises an improvement in a clinical measure has moved into health advertising whether or not anyone on the marketing team intended it. The two triggers are simple: does the product require a licence, and does the video imply a financial or clinical outcome.

The second trigger is where most teams get caught, because they audit the script and forget the picture. The FTC’s health products compliance guidance states that marketers should not focus narrowly on individual phrases but should consider each advertisement as a whole, assessing the “net impression” conveyed by all elements, including the text, the product name, and any charts, graphs and other images. The guidance goes further with a worked example: a brochure showing doctors in white coats, microscopes, molecular structures and a stack of medical journals conveys an implied claim that the product has been clinically proven. Nobody said it. The frame said it.

For video that is a structural problem rather than a copy problem, because video is mostly picture. A voiceover that has been through legal three times can still ship an unapproved claim if it plays over a rising chart, a white coat, or a screen showing a return that nobody substantiated. The same guidance notes that when an advertisement lends itself to more than one reasonable interpretation, the advertiser is responsible for substantiating each interpretation.

On the financial side the boundary is written down. FINRA Rule 2210 defines a “retail communication” as any written, including electronic, communication distributed or made available to more than 25 retail investors within any 30 calendar-day period — which is essentially every piece of video marketing a US-facing broker-dealer publishes. In the EU the transparency layer arrives on a fixed date: the AI Act’s Article 50 comes into force on 2 August 2026 under Article 113.

How claims review works before a script exists

The single change that saves the most time is also the least glamorous: the claims table is written before the script. Not after the first cut, not alongside the storyboard — before anyone writes a line of voiceover.

A claims table is a flat document with five columns: the claim as the marketing team wants to say it; the source document that supports it, with a page or section reference; the person who owns that source; the approved wording; and the status. It is deliberately boring. Its whole purpose is to move the argument about what the video may say to a point in the schedule where changing the answer costs an hour instead of a re-shoot.

What the table has to withstand is specific. FINRA Rule 2210 states that communications may not predict or project performance, imply that past performance will recur, or make any exaggerated or unwarranted claim, opinion or forecast, and that all communications must be fair and balanced with a sound basis for evaluating the facts. On the health side the FTC states that claims about the health benefits or safety of health-related products generally require substantiation in the form of competent and reliable scientific evidence. Those two sentences kill most first-draft fintech and health scripts, and they kill them cheaply if you read them at the table stage.

The craft move is not to delete the claim. It is to change its shape. “Grow your savings faster” is a projection. “Set a risk level in four taps and see the fee before you confirm” is a demonstrable product behaviour that needs one screen recording’s worth of evidence and reads as more confident, not less. FINRA Rule 2210 also requires that any comparison in a retail communication disclose all material differences between the things compared, which is why a side-by-side competitor sequence is usually the most expensive ten seconds in a regulated edit.

We also flag the implied claims at this stage, because they are the ones that do not appear in any script document: a stethoscope in frame, a rising line graph with no axis labels, a “5-star” motif, a person in a suit at a trading desk. Each of those gets a row in the table like any spoken line.

Where disclaimers actually belong

Most disclaimer problems are placement problems, not wording problems. The wording has usually been settled by legal for years. What has not been settled is whether a viewer on a phone, with sound off, in a feed, can actually read it.

FINRA Rule 2210 puts this plainly: information may be placed in a legend or footnote only in the event that such placement would not inhibit an investor’s understanding of the communication. Read as a production instruction, that rules out three habits we see constantly — a disclaimer that exists only in the video description, a disclaimer burned in at a size that is illegible at mobile scale, and a disclaimer that appears for less time than it takes to read.

So we treat the disclaimer as a design element with a shot number, not as an afterthought applied in the export:

• It is in the cut, not the caption. The video has to be compliant when it is embedded, downloaded, re-uploaded to a partner’s channel or played in a meeting room without any surrounding page.
• It is timed to reading speed. We hold it long enough to be read at the pace of the narration rather than matched to an edit beat.
• It is legible on a phone. Minimum type size, contrast and safe-area are locked in the storyboard, and we check the frame at mobile scale before animation, not after.
• It sits next to the claim it qualifies. A risk statement at the end of a ninety-second film does not qualify a performance graphic at second twelve.
• Every platform version keeps it. The 9:16 cut, the 6-second bumper and the silent autoplay version each need their own disclaimer treatment, and the bumper is where it usually gets quietly dropped.

Why we won’t generate people who look like your customers

This is the line we hold hardest, and it is worth explaining rather than asserting. Generated humans in our work are non-specific, are never framed as customers, patients, clinicians or advisers, and never deliver a line that functions as a testimonial. We do not clone the likeness or the voice of any real person, and we do not build a synthetic spokesperson for a regulated brand.

There are three separate reasons, and they stack.

The first is that a synthetic customer is an unsubstantiated claim wearing a face. The FTC’s worked example about lab coats and microscopes is the mild version; a generated person who reads as a satisfied investor or a recovered patient is a testimonial that cannot be substantiated because the person does not exist. There is no consent, no typicality evidence, and nothing behind it.

The second is regulatory definition. The EU AI Act defines a deep fake, in Article 3(60), as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. Article 50(4) then requires deployers of a system that generates such content to disclose that it was artificially generated or manipulated. A generated “real customer” walks straight into that definition; a stylised, obviously-crafted human figure in a brand film generally does not.

The third is reputational, and it is the one marketing directors actually feel. Google’s Ads misrepresentation policy lists manipulated media and unreliable claims among the practices it prohibits, and specifically bars making it appear that you are supported by another brand, organisation or government entity when you are not, or lying about services that could put people’s health or safety at risk. In a supervised category, one complaint about a fake patient is not a creative note. It is an incident with a file number.

Product screens, data and what never leaves your environment

The most common data incident in video production is not a breach. It is a screen recording taken from a staging environment that still contained real records, cut into a film, approved by four people who were watching the transition rather than the table, and published.

Our rule is simple: we rebuild the interface, we do not record it. Practically that means we ask for your design system or component files rather than a video capture, and we animate the screen from those. The interface in the finished film is a construction, so every element in it is under our control and under your sign-off.

That leaves the values, which is where the real work is. Before animation starts we agree a synthetic data sheet with your team covering:

• Identities — invented names, avatars and locations that are checked not to collide with a real customer, a public figure or a competitor’s brand.
• Money — balances, transaction amounts, fee figures and portfolio values chosen to be plausible and, critically, chosen so that no combination of them reads as a performance claim. A balance that grows across three shots is a projection.
• Clinical values — vitals, ranges and results set inside normal reference bands unless a specific abnormal value is required by the story and signed off with the clinical claim behind it.
• Identifiers — account fragments, policy numbers, member IDs and dates generated from reserved or clearly invalid ranges.
• Charts — series data written out as numbers and approved as numbers, because a chart shape is a claim even when no axis is labelled.

Two things follow from doing it this way. There is no production capture to scrub, so there is nothing to redact under time pressure three days before launch. And the sheet itself becomes an audit artefact: when someone asks in six months where the figures on screen came from, the answer is a signed document rather than an editor’s memory.

The same discipline applies to everything else that reaches us. Brand assets, design files and any documents you send for claims substantiation stay inside the project, are not used to train anything, and are not passed to a third party outside the named production pipeline.

Do we disclose that it was made with AI? Yes

Plainly, and without being asked. We use generative tools — Midjourney for design frames and look development, Kling, Veo and Seedance for motion — alongside conventional storyboarding, cinematography direction, colour grading and sound design. We say so, on our own site and in the projects we deliver, and we help clients draft their own disclosure line during scoping so that compliance approves the wording with the script instead of arguing about it after the edit.

The obligations are converging fast enough that a position taken now is cheaper than one taken later:

Our house position is that disclosure is a positioning asset, not a liability. A regulated brand that states its AI use clearly is a brand whose other statements become easier to believe — which is exactly the argument one commenter made in the Hacker News discussion cited further down.

The audit trail: what you get and what it’s for

The audit trail is the deliverable nobody asks for at the briefing and everybody asks for at the review. We assemble it during production rather than reconstructing it afterwards, because reconstruction is where the gaps appear. It has five parts.

1. The approved claims tableEvery spoken and on-screen claim, the source document behind it with a reference, the approved wording, the approver and the date. This is the document that answers “why does the video say that”.
2. The shot logShot by shot: what was generated, with which tool, what was live-action or client-supplied, and what was composited. This is what lets you answer a disclosure question precisely instead of defensively.
3. The synthetic data sheetEvery value visible on a product screen, agreed before animation. It proves that nothing on screen came from a real account, a real patient or a real transaction.
4. The disclosure recordThe decision on whether and how AI use is disclosed, the approved wording, the platform-specific labels applied at upload, and who signed that decision off.
5. The version historyWhich cut each approver saw, what changed between versions, and which version was released. When a complaint arrives eight months later, this is the only artefact that reliably settles it.

One note on rights, because it is the question that always follows: rights are defined per project in the quote.

Fintech vs. health vs. insurance: what actually changes

The process is the same. What differs is which step expands, and knowing that in advance is most of scheduling accurately.

Internal audiences behave differently again. Compliance training, onboarding and policy-change films are regulated in a second sense — they have to be accurate, versioned and re-issued when the policy changes — which is a different production pattern and one we cover under AI training video. Multi-brand groups running a shared library across several supervised entities usually start at AI video for companies.

The seven-step process with an approval gate

1. Compliance intake, not a creative briefOne session with marketing and whoever releases advertising in the room together. We leave with the claim you want to make, the documents behind it, the named approver, and any disclosure position you already hold.
2. Claims table, drafted and circulatedWe write the first pass, including the implied claims carried by imagery. Your side edits and signs it. This is the deliverable that unblocks everything else.
3. Script and storyboard against the signed tableEvery line maps to an approved row. Disclaimers get shot numbers and on-screen durations here, not in the edit.
4. The approval gateThe named approver signs the storyboard and the claims table together, before any animation. This is the whole point of the process: a change here costs an hour, the same change after the first cut costs a week.
5. Screen build and synthetic dataThe interface is rebuilt from your design files and populated from the agreed data sheet. Nothing is captured from a live environment.
6. Generative production and finishingDesign frames in Midjourney, motion through Kling, Veo and Seedance, then the conventional craft layer — edit, grade, sound design, mix — under the Creative Director. The shot log is written as this happens, not after.
7. Delivery with the audit packMaster plus platform versions, each with its own disclaimer treatment and disclosure labels, delivered with the five-part audit trail.

Where the time really goes on a regulated project

What people are actually saying

Three public discussions capture the state of the argument better than any vendor page, including ours. All three are on Hacker News and all three are linked below.

Reputation is the risk, not the technology

The thread on McDonald’s pulling an AI-generated Christmas ad after backlash is the clearest read on audience risk. One commenter argues that AI is deeply unpopular with a large and vocal share of the public, that this makes AI-generated assets reputationally risky for brands in any public-facing project, and that marketing managers need to recognise it. Another makes a craft argument rather than an anti-AI one: the tools can produce work of good quality, but taste is what is lost when the professionals are removed from the process. Both points land harder in a supervised category, where a reputational incident tends to arrive with a regulator attached. Our reading is not “avoid AI” — it is that the pipeline needs a named creative director standing between the tool and the release, which is also the argument behind evaluating a studio’s portfolio properly before you hire.

Labelling the AI work is what makes the rest credible

The thread on the EU moving to mandatory labelling of AI-generated content runs the disclosure debate in both directions. Sceptics predict “may contain AI content” stickers on everything, and one commenter points out that the EU legislating labelling is not the same as enforcing it. But the most useful comment in the thread makes the reputational case: if an organisation is known to tag its AI-generated material, its untagged material gains weight as original. Another asks for disclosure to be a short standardised label naming what AI was used for — code, assets, text — rather than a blanket warning, which is close to how we structure the shot log.

Nobody can buy certainty yet — so build the record

The Show HN thread for an offline-first EU AI Act compliance tool is the most practical of the three. One commenter’s caution is the important one: the interpretation of what compliance with the AI Act actually requires is still in high flux and changing frequently. Another asks the author to state plainly that coding assistants were used in building the project, “if only for intellectual honesty” — a norm argument about disclosure that arrived from the community rather than from a regulator. Together they describe the position we work from: you cannot buy certainty about a moving target, but you can keep a record good enough to show what you did and why.

A compliance-readiness scorecard

Five mistakes that send a video back to concept

Cost and delivery

Delivery is 7–10 working days from an approved brief to a delivered master — the same window we run on any other production. The word doing the work in that sentence is approved. The brief is not approved until the claims table is signed and the synthetic data set is agreed, and on a regulated project that is where the calendar actually lives. Teams that arrive with those two documents ready ship at ordinary speed; teams that start them after the kickoff call add whatever their internal review cycle costs.

Pricing is per project against a written scope rather than a rate card, and Arcanewiz productions start From $1,500. What moves a regulated quote is specific: the number of deliverables and platform versions, how much of the on-screen interface has to be rebuilt from design files, how many approval rounds your compliance path requires, and whether a claims table already exists or we are writing the first one. B2B software teams comparing studios on process rather than price usually start with our breakdown of AI video studios for B2B SaaS.

The pre-brief checklist

Bring these to the first call and the compliance intake collapses from three meetings into one.

• The single claim you most want the video to make, written as one sentence.
• The document that supports it, with the relevant page or section marked.
• The name of the person who releases advertising, and their availability over the next three weeks.
• Your product design files or design system — not a screen recording.
• Any existing AI-disclosure position, or a note that you do not have one yet.
• The platforms the film will run on, and whether the ad accounts are certified where certification applies.
• The disclaimer text your legal team has already approved for other channels.
• Any wording, imagery or comparison that has previously been rejected internally, and why.

Where Arcanewiz fits

Arcanewiz is a cinematic production studio that uses a generative pipeline, in that order. Creative Director Daniel Atzil brings more than twenty years of traditional production — storyboarding, cinematography, colour grading and sound design — and the AI layer sits underneath that craft rather than replacing it: Midjourney for design frames and look development, Kling, Veo and Seedance for motion. The process above is not a compliance add-on we bolt onto a standard job; it is how the job is scheduled from the intake call.

Two things we will not do, stated plainly because regulated buyers are right to ask. We do not clone the likeness or voice of a real person, and we do not build synthetic people who are presented as your customers, patients or advisers. Our cleared client roster — Samsung Israel, Anipet, Fun Forest, Homey Panda and SolarEdge — is enterprise and consumer brand work, and we are not going to dress any of it up as a regulated-industry case study it was not. What we bring to a supervised category is the process, the audit trail and the craft, and we would rather show you the claims table than a testimonial.

If you are earlier in the decision and still comparing approaches rather than studios, the AI video for fintech and AI video for healthcare pages go deeper into the deliverables themselves.

Frequently asked questions