AI Video for Banks and Fintech: Where the Film Dies in Compliance
A retail bank approves a ninety-second film for a deposit product. The visuals are clean, the voice is calm, the branch staff like it. Then compliance reads the script line by line and finds four problems. The rate is stated as a single number instead of a range with a validity date. A person in the background is wearing a lanyard from another institution. The phrase "guaranteed return" sits in a subtitle that nobody checked. And the disclaimer at the end is on screen for 1.4 seconds.
The film is pulled two days before launch. Nobody argues about the images, because the images were never the problem. The script, the numbers and the approval trail were not built to survive a regulator.
We produce AI video in Tashkent for clients across several regulated categories, and banking is the one where pre-production discipline matters more than image quality. A beautiful film that legal rejects has a production cost and no value. A plain film with an auditable script ships.
What the film actually has to do
A bank video is rarely one job. In practice it carries three, and they pull in different directions.
The first is explanation: move a viewer from "I do not understand this product" to "I know what it costs and what it requires" in under a minute. The second is deflection: reduce the volume of the same five questions at the branch counter and on the call line. The third is record-keeping: leave behind a version of the claims that someone can point at eighteen months later and say "this is what we told customers, and here is who approved it".
Most bank films fail the third job silently. Marketing treats the video as a creative asset; compliance treats it as a financial promotion. When those two views meet late, the project stalls.
Lock the script before anyone renders a frame
The single most useful artifact in a banking video is not a storyboard. It is a claims table: one row per assertion the film makes, with the approved source document, the exact wording, and the person who signed it off.
Once that table exists, the script stops being a creative negotiation and becomes a build. Sentences either map to a row or they get cut. The table also gives you a cheap way to localise: translators translate rows, not vibes, so the Russian and Uzbek versions cannot drift into claims the English version never made.
From one approved script we normally cut three lengths: a thirty-second vertical for social, a sixty-second loop for in-branch screens with no sound, and a ninety-second version for the website. Cutting from a single locked script costs a fraction of producing three separate films, and it keeps the numbers consistent across all three. Nothing damages trust faster than a rate of 22 percent on Instagram and 21.5 percent on the branch screen.
The advisor who is not an employee
Banks want a human face. They also rarely want to commit a real employee to forty videos over a year, and they should not.
Generated presenters solve the scheduling problem and create a different one: a synthetic advisor must never be placed in a slot that implies employment. "Meet our team" pages, branch signage with names attached, and onboarding films that present a person as staff are the wrong places for a generated face. For product explainers, generic service walkthroughs and safety content, a generated presenter is fine, provided the look is consistent and the voice is native.
Consistency is the hard part. A presenter who changes jawline between videos reads as untrustworthy at a level viewers cannot articulate. We fix a look bible at the start: age band, wardrobe palette, hair, framing distance, lens feel, and a locked voice. Every subsequent video is built against that document, and we re-render rather than accept drift.
Voice matters more than face in this category. A Russian-accented voice reading Uzbek copy sounds like a call centre reading a form. If the audience is over forty-five, subtitles alone will not carry the message, and a native voice performance is worth the extra cost.
Screens, rates and the footnote nobody reads
Every number a bank shows is a liability. Do not let an image model draw numbers. Generative models produce digits that look plausible and are wrong, and a wrong rate in a video is a regulatory event, not a design error.
Instead, generate the footage clean and composite the interface separately: a screen plate rendered at twice the delivery resolution, text set in the brand typeface, numbers inserted as data. This keeps text crisp after compression and makes a rate change a one-hour job instead of a re-shoot.
Then respect reading time. A six-word line on a phone screen needs roughly three seconds. A legal footnote needs at least four, set no smaller than 24 pixels at 1080 width, in a colour that passes contrast against the actual frame behind it, not against a design mockup. We test every text frame by watching it once at normal speed on a phone. If our own team cannot read it, no viewer will.
The last habit that saves money: keep variable numbers out of the voiceover. If the rate lives in the voice track, every change means a new voice session and a new compliance review. If it lives on a screen plate, a rate change is a text edit and a re-approval of one frame.
Where AI still fails in a bank context
Honesty about limits is cheaper than re-rendering.
Hands and cash remain unreliable: finger counts, note edges, the way a thumb holds a bundle. Anything where a document is read, signed or handed over tends to smear. ID cards, uniforms with insignia, and lanyards with legible text break quickly. So do reflections in branch glass and any wide shot of a real interior that the client will recognise, because a generated interior that is almost the client's branch reads as a mistake rather than a stylisation.
Testimonials are the category we refuse to generate. A synthetic customer saying they were approved in two days is a legal exposure, not a creative choice.
So we photograph a small set of plates: branch exterior, an advisor at a desk, a card in a hand, a phone in a hand at the right angle, a call-centre headset, a meeting room. Ten plates is often enough. AI then carries the transitions, the environment extensions, the product metaphors and the versions in other languages. The mixed approach costs one shoot day and buys back the shots that would otherwise be argued about for weeks.
Language versions and terminology traps
Uzbek financial terminology is not fully settled. Omonat and depozit both appear in official documents, and interest is written several ways. Pick one glossary at the start, put it in a shared file, and hold every translator and voice artist to it. Without that, three videos in the same campaign will use three words for the same product.
Russian copy must match the contract language, not a marketing paraphrase, because customers compare. Keep legal phrasing intact even when it is clumsy; a smoother sentence that changes meaning is a defect.
For subtitles, assume forty-two characters per line at phone width, two lines maximum, and never split a number across lines. Burn in Uzbek subtitles rather than relying on platform captions, which mangle Latin Uzbek diacritics.
Approvals, gates and what actually drives turnaround
Three gates work better than open review. Gate one is the locked script and claims table. Gate two is the first cut with final voice. Gate three is the colour and text pass. Each gate has one named approver, and feedback after a gate costs a change order rather than an email.
With a claims table in hand, a ninety-second film in two languages takes five to eight working days. Without it, the same film takes three weeks, because every review round reopens the script.
KAMOD is an AI video studio based in Tashkent, working with clients worldwide. If you have a rate sheet and an approver, we can start on a first cut this week. Send us the product one-pager and the list of claims you are allowed to make, and we will come back with a script, three lengths and a version plan before anything is generated.