Every marketing team in Malaysia has now been shown an AI generated video and asked why their content costs what it costs. It is a fair question, and it deserves a more useful answer than either of the two on offer. The dismissive answer, that AI video is a novelty that cannot match real production, is already out of date. The credulous answer, that it replaces production entirely, does not survive contact with a client brand guideline. This article looks at where AI video genuinely fits in a Malaysian brand’s content mix, what it does well, what it still cannot do, and how to brief it so the output is usable.
What AI Video Actually Means in Practice
The term covers several distinct capabilities that get bundled together, which is the source of most confusion. It is worth separating them, because they have very different reliability.
- Generative footage. Producing moving images from a text prompt or a still image. Impressive, improving quickly, and still unpredictable over durations longer than a few seconds.
- Image to motion. Taking an existing still, such as a product render or a photograph, and animating it. Considerably more controllable than pure generation, because the starting frame is fixed and approved.
- Synthetic presenters and voice. Generated people delivering scripted lines, and cloned or synthesised voiceover. Useful for volume and localisation, weak on genuine warmth.
- Assisted post production. Rotoscoping, upscaling, noise reduction, transcription and rough assembly. The least discussed and by some distance the most immediately valuable.
The fourth category is already saving real money on conventional productions without any client ever knowing it was involved. The first three are where the marketing conversation happens.
Where It Genuinely Works Today
AI video performs best where the requirement is volume, variation or abstraction rather than brand critical accuracy.
Social content at scale is the clearest case. A campaign needing thirty variants of a fifteen second piece for different segments, platforms and languages is a poor use of a full production crew and a good use of generative tools. The individual pieces do not need to be perfect; they need to be plentiful, timely and on message.
Abstract and atmospheric sequences work well because there is no ground truth to violate. Flowing particles, light, texture and environmental mood have no correct appearance a viewer can check against reality. This is why AI generated material frequently appears in title sequences and transitional moments even in high end productions.
Animating approved stills is the most reliable application of all. When a client has already signed off a 3D product render, bringing that exact frame to life for five seconds carries almost none of the risk of generating a product from scratch, because the product is already correct in the source image.
Concept visualisation and pitch material is another honest use. Showing a client an approximation of a proposed treatment, clearly labelled as indicative, communicates intent far faster than a mood board.
Where It Still Fails
The failures are consistent and worth knowing before a client is promised something.
Brand critical product accuracy is the hard limit. A generative model does not know a client’s product. It will produce something that resembles it, with the logo subtly wrong, the proportions altered and details invented. For any brand where the product must be exactly right, and that is most of them, generation from scratch is not viable. The workaround is to start from an approved render or photograph.
Human performance remains detectable. Synthetic presenters are adequate for functional information and unconvincing for anything requiring warmth, humour or trust. A testimonial from a synthetic person is worse than no testimonial, because if an audience notices, the credibility damage exceeds any benefit.
Continuity across shots is unreliable. Maintaining the same character, the same room and the same lighting across a sequence of generated shots is still difficult, which is why most convincing AI work is brief and self contained.
Text and typography inside generated footage is still routinely wrong, which matters in a market where signage, packaging and multilingual copy frequently appear on screen.
The Local Considerations That Get Overlooked
Malaysian brands face a few specific issues that global demonstrations of these tools do not address.
Representation is the most immediate. Generative models trained predominantly on Western data produce a default population that does not look like Malaysia. Getting a genuinely representative multiethnic cast, appropriate dress including hijab, and recognisable local environments requires deliberate and repeated prompting, and the results are inconsistent. For any brand whose audience is the whole country, this is a serious limitation rather than a detail.
Language is the second. Bahasa Malaysia, Mandarin and Tamil voice synthesis has improved but remains weaker than English, and prosody errors are more noticeable to native listeners than to the people approving the work. Mixed language delivery, which is how many Malaysians actually speak, is handled poorly.
Environment is the third. Generated street scenes tend towards a generic Asian city that is recognisably not Kuala Lumpur. If the setting needs to feel local, filmed plates or real photography usually remain the better answer.
A Sensible Content Mix
The productive framing is not AI against traditional production. It is matching the method to the job.
- Brand films and hero content. Conventional production. These carry the brand’s credibility and cannot afford uncanny detail.
- Product visualisation. 3D from CAD data, which is accurate by construction, with AI assisting on environments and atmosphere around the accurate hero.
- Volume social content. AI led, working from approved brand assets, with human editing and quality control before anything publishes.
- Localisation and variants. AI assisted, with native speaker review of every language version without exception.
- Internal, training and explainer content. Strong AI fit, since the audience values clarity over polish.
- Concept and pitch work. AI, clearly labelled as indicative so nobody mistakes it for a deliverable.
What It Costs and Why It Is Not Free
The assumption that AI video is nearly costless comes from watching someone generate one clip. Production cost does not sit in the generation; it sits in everything around it.
A usable AI led video still requires planning, scripting, art direction, asset preparation, extensive iteration to get acceptable takes, editing, sound design, music licensing, captioning and quality control. Generation replaces the camera, not the craft. A packaged AI video service in Malaysia is therefore quoted much like any other production, in the thousands of ringgit for short form pieces and considerably more for longer packages, with the saving showing up as more output for the budget rather than a token invoice.
Where the saving is real is variation. Producing a second, third and tenth variant of an existing piece costs a fraction of producing the first, which is exactly inverted from conventional production economics and is the strongest commercial argument for including it in the mix.
Briefing It Properly
Teams that get good results brief differently. Start from approved assets wherever possible, because an approved still is the single most effective quality control available. Keep generated shots short and cut between them rather than asking for long continuous sequences. Specify the local context explicitly rather than assuming it. Budget iteration time realistically, since a usable take often follows several unusable ones. Finally, put a human editor and a native speaker between the output and the publish button, every time.
Disclosure, Rights and the Questions Clients Now Ask
Two years ago no client asked whether AI had been used. Now it appears in briefing documents, and procurement teams increasingly want it in writing. Getting ahead of that conversation is easier than being caught by it.
Disclosure norms are still forming, but a reasonable position is that audiences deserve to know when a person on screen is synthetic, and that abstract or assistive uses do not require a label. A generated presenter presented as a real employee is the case that causes damage. Atmospheric particle work in a title sequence is not.
Rights are the more concrete issue. Commercial usage terms differ between tools, and some outputs carry restrictions that matter for advertising. Before generated material goes into a paid campaign, confirm what the tool’s terms permit for commercial use, whether any indemnity is offered, and whether the client’s own contract with their agency or media buyer imposes further conditions.
Likeness deserves particular care. Generating a person who resembles a real individual, or cloning a voice without documented permission, creates exposure that no production saving justifies. If a synthetic presenter is based on a real person, get written consent covering the specific uses.
Keep a simple record of which tools produced which assets. When a client asks six months later, and they increasingly do, the ability to answer precisely is worth the small effort of noting it as you go.
Building a Workflow the Team Can Actually Run
Most organisations that try AI video do it as an experiment run by one enthusiastic person, and it collapses when that person is busy. Making it durable requires treating it as a process rather than a tool.
Start with an asset library. Approved product renders, brand colours, typefaces, logo files and reference stills are the raw material that makes generation reliable. A team generating from a blank prompt gets inconsistent results; a team generating from approved assets gets usable ones.
Write down the prompts that worked. Prompt craft is genuinely cumulative, and a shared document of what produced good output for your specific brand saves enormous time compared with each person rediscovering it.
Build a review gate before anything publishes. At minimum, a human editor checks for artefacts, wrong hands, distorted logos and mangled text, and a native speaker checks any non English audio. This gate is what separates an efficient workflow from an embarrassing one.
Finally, measure the output rather than assuming. If AI led social content performs comparably to conventionally produced content, expand it. If it does not, the mix is wrong and should be adjusted rather than defended.
How to Apply It
Audit the content calendar and sort each item by how much brand critical accuracy it requires. Anything where the product or a real person must be exactly right stays with conventional production or accurate 3D. Anything that is abstract, high volume, internal or a variant of something already approved is a candidate for AI led production.
Start with the safest application, which is animating stills you have already approved, and expand from there once the team has seen what the output actually looks like at broadcast quality rather than in a demonstration.
At Avanguardia, we produce AI led video alongside conventional production and 3D for brands across Malaysia, including packages that combine AI generated sequences with accurate product renders. If you are trying to work out what belongs where in your content mix, talk to our team.
References
Wyzowl. (2026). The state of video marketing 2026. Wyzowl. https://www.wyzowl.com/video-marketing-statistics/
Think with Google. (2026). Consumer insights. Google. https://www.thinkwithgoogle.com/consumer-insights/