Let us start with the thing that should stop a bad brief in its tracks.

Runway's Terms of Use, last updated 23 October 2026, state in section 4.4 that your Inputs and Outputs may be used by the company to train and improve its AI models, under a licence described as non-exclusive, irrevocable, perpetual, worldwide and royalty-free. The terms describe no opt-out. They also say Runway does not claim ownership of your inputs or outputs and does not restrict your commercial use of them, and they direct enterprise customers to separate Enterprise Services terms. Read section 4.4 yourself — it takes two minutes.

Now consider what a brand hands an agency for a product film: unreleased packaging, pre-launch product plates, talent footage, brand reference material. If that goes through a self-serve account on a platform with terms like these, you have granted a perpetual training licence over your unreleased assets, and nobody on the call mentioned it. That is not an argument against AI video. It is an argument for reading the production stack before you approve a budget, which is where our project scoping now starts.

With that established, here is the honest position on what this technology can actually do for you.

Should you use AI video instead of shooting?

No, and at your budget the question is a distraction.

At ₹3 lakh and upward in India you can commission a genuinely competent shoot — experienced crew, decent glass, a studio or a real location, proper post. The pitch that generative video democratises high-end production is aimed at someone with ₹30,000, not someone with ₹3 lakh. You are not the buyer that argument was written for.

And you will end up shooting anyway. The exact stitching on the shoe, the refraction through the bottle, the specific finish on the packaging — a diffusion model cannot hallucinate your actual SKU to brand-guideline accuracy. If the product is the message, the product goes in front of a camera.

So the useful framing is different: AI video is a multiplier on what you shoot, not a replacement for shooting it.

What is the Message Shot?

Go through your storyboard and mark every shot where the message actually lives. The product in use. The face delivering the line. The hands doing the thing. The before and after.

Those are Message Shots. They get a camera, a lens and a human being.

Everything else — the environment, the scale, the aerial, the city at dusk, the impossible interior, the abstract transition, the texture bed under the voiceover — is negotiable, and that is where generative tooling earns its place. It is cheaper than a location, faster than a build, and nobody is examining it closely because it is not what the viewer is looking at.

The failure mode in almost every disappointing AI brand film is the same: somebody tried to generate a Message Shot. The hands go wrong, the product drifts between frames, the eyes are dead, and the viewer's attention lands precisely on the thing that cannot survive attention.

Where does generated video still reliably fail?

Vendors document this themselves, and academic benchmarking of generative video consistently finds the same cluster of weaknesses. The ones that matter commercially:

• Sustained human performance. Short bursts can look remarkable. A held close-up with a line of dialogue and a change of expression is still where it breaks.
• Hands and faces at close range. Improving, still the first place a viewer's eye catches an error.
• Continuity across shots. Keeping one identity, one product and one space consistent across a cut remains hard. Reference conditioning helps and does not solve it.
• Text and typography in frame. On-pack copy, price points, legal lines — expect to composite these rather than generate them.
• Physical plausibility. Liquids, collisions, cloth, anything involving mass behaving correctly. Benchmark work on physical commonsense in generated video finds current models score poorly on exactly these tasks, because they have no underlying model of how the world works.
• Your actual product. Close enough is not a brand guideline.

Two practical consequences. First, generations are cheap but usable generations are not, because you discard many to get one — and no vendor publishes a yield rate, so your real cost per shot is unknown until you measure it on your own project. Second, clip lengths are short and coherence degrades as you extend, so plan in beats of a few seconds rather than in sequences.

Where is the real win for an Indian brand?

Not the generated shot. The audio layer — and almost nobody commissioning this work is using it properly.

One shoot, performed once, becomes a genuinely multi-market package: voice generation, voice cloning and lip synchronisation across Indian languages, at a cost conventional dubbing never allowed. For a brand selling across three or four language markets, that is a larger commercial effect than any visual generation.

The infrastructure is real and improving. Sarvam AI documents voice cloning and synthesis across 12 Indian languages, and AI4Bharat's Indic-Parler-TTS covers 22 Indian languages under an open licence. That is a materially better position than existed two years ago.

But the gating factor is per-language quality, not the tool's language list. A platform listing 100-plus languages is telling you what it will accept, not what it does well. Deep phonetic alignment exists for some Indian languages and is thin for others, and the two layers fail independently — the voice can be convincing while the mouth is wrong, or the reverse. A Hindi plate force-mapped onto a Dravidian or eastern-language phonetic set produces mouth shapes a native speaker clocks instantly.

Here is the part that makes this expensive to get wrong. A brand manager in Mumbai or Indore reviewing a Tamil or Odia cut cannot hear what is wrong with it. Neither can their agency, usually. The audience in Chennai or Bhubaneswar can, and what they perceive is not "impressive technology" but "cheap advertisement." So the rule is simple and non-negotiable: every language version is signed off by a native speaker before it ships, and you budget for that review rather than assuming it.

Treat it as a staged rollout. Start with the languages where the tooling is demonstrably strong, prove the quality with native review, and expand. The same discipline that applies to measuring any campaign properly applies here: test one market before you commit to nine.

What must you disclose in India, and from when?

This changes the asset, so it belongs in the brief rather than in a post-production conversation.

ASCI's labelling guidelines take effect in late December 2026. Published on 29 September 2026 with a three-month runway, they apply a risk-based test: disclosure is required where synthetically generated content materially shapes consumer decisions and leaving it undisclosed could mislead. Media coverage of the release sets out the structure and the guidelines are published in full.

Prohibited regardless of any label: fabricated endorsements, synthetic depictions that exaggerate product results, and deepfakes without consent. A disclosure does not cure a misleading claim. If someone proposes generating a product outcome you cannot deliver and labelling it, the label is irrelevant and you are the one in breach.

Label required: synthetic presenters and virtual influencers; replication of a real person's likeness or voice in personalised messaging even where that person consented; invented events or settings that affect how the ad is understood; demonstrations of products that do not yet exist; sound effects tied to core product features; and paid AI-generated product recommendations, which must identify the sponsor.

No label required: routine editing such as colour correction, lighting and minor blemish removal; decorative or obviously fantastical elements; text-based uses including AI-written ad copy; and accessibility features such as subtitles and translations.

Read those three lists against the audio strategy above and you will see the consequence immediately. A cloned voice in a personalised message needs a label even with a signed consent. A dubbed version using a synthesised voice in place of the original performer is squarely in the labelling conversation. Budget for the label as a design element from the storyboard stage, not as something bolted on in the final week.

Separately, India's IT Amendment Rules have been in force since 20 February 2026 and reach audio, visual and audio-visual content — text is excluded. They place obligations on platforms and significant social media intermediaries rather than on you directly, with sharply shortened timelines including 36 hours for grievances about unlawful synthetic content. Khaitan & Co's summary is the clearest account. The practical effect for an advertiser is that platforms now ask you to declare synthetic content at upload, and answering carelessly puts your campaign at risk.

Can you generate a character who looks like a celebrity?

No, and Indian courts have been unusually quick here.

In September 2023 the Delhi High Court restrained the unauthorised commercial use of Anil Kapoor's name, image, voice and persona traits, in a case that explicitly engaged AI misuse. In May 2024 the same court made orders protecting Jackie Shroff's name, image, voice and likeness, including against an AI platform whose chatbot used his personality attributes.

Worth noting the limit, because the protection is not unbounded: in the Shroff matter the court declined an interim injunction against one YouTube video it considered arguably a tribute, citing free expression. So this is a developing area rather than a blanket rule — but for advertising it is settled enough. Prompting a model toward a recognisable public figure's likeness, voice or signature manner is not available to you, and buying consent does not remove the ASCI labelling obligation that follows.

What do you actually own, and what can you not use?

Three things to settle before production, because all three are contractual rather than creative.

Copyright in generated footage is doubtful. Indian copyright requires a human author. Purely generated output may not be protectable, which means a competitor copying it may face no enforceable claim from you. Substantial human direction, editing and arrangement improves the position, which is one practical reason the hybrid route beats the fully generated one — and a reason to document the human contribution rather than just make it.

Music licensing moved, and one tool left the market. Following its settlement with Universal Music Group, Udio locked downloads, including for tracks users had already created, which removes it as a route for brand work. Suno ties commercial rights to subscription status at the time of creation, so check the current terms before you rely on a track. And note the gap that applies to all of it: a platform permitting commercial use is not the same as you owning the music.

Training rights on your own assets. Back to where we started. Ask your agency, in writing, which platforms and which account tiers your material passes through, and what those terms say about model training. If unreleased product plates are involved, that answer should determine the stack.

How a hybrid project should actually be specified

Six things that belong in the brief. Most of them are decisions made before the camera turns on.

1. Mark the Message Shots. Those are shot. Write it into the storyboard so it is not renegotiated in post.

2. Shoot for extension. Clean plates, framing headroom so the same take reframes to 16:9, 1:1, 4:5 and 9:16 without recomposing, consistent lighting so generated environments can be matched, and deliberate reference capture of the product and talent from multiple angles.

3. Specify the versioning matrix upfront. Languages, aspect ratios, durations and offer variants. This is where one production day becomes many finished assets, and it is the number your quote should be judged on — not the day rate. Our social and content work is usually where that matrix pays for itself.

4. Name the review gates. A technical pass on generated elements, a brand pass on product fidelity, a native-speaker pass per language, and a compliance pass against the ASCI categories above. Each with a named owner.

5. Settle the stack and the rights. Platforms, tiers, training terms, music licensing, talent consent, and what records are kept per asset.

6. Agree the label treatment. If any version needs a disclosure, design it into the frame now.

When not to do this at all

Four situations where we would tell you to commission a straight shoot and skip the generative layer entirely.

The product is the whole film. Jewellery, apparel fit, food texture, device detail — where the viewer is studying the object. The failure is expensive and it will not show up in a performance report for two quarters. This covers a lot of retail and ecommerce work.

The film carries documentary authenticity. Founder stories, customer testimony, factory and process films. The entire persuasive value is that it happened. Synthesising any part of it removes the only thing it had.

You need one evergreen asset, not many. Generated footage carries a visible period signature, in the way that early CGI is instantly datable now. A single brand film built heavily on generation will look its age within a year or so; a well-lit film shot on good glass will not. If you are making one film to last three years, shoot it.

You cannot staff the review. Native-language sign-off, product fidelity checking and compliance review are real work. Without those gates, volume becomes risk — the same mechanism behind most marketing programmes that quietly fail.

And the objection we would make if we were sitting on your side of the table: an agency shifting production in-house to generative tooling captures a larger share of your budget as margin, because less of it passes through to a production house and crew. A reasonable response is to commission the shoot where fidelity matters and buy the generative extension separately. We think the integrated version produces a better result, but you are entitled to be sceptical about who benefits from that view, and our pricing is set out openly partly so that question can be asked.

What nobody can tell you

There is no reliable public benchmark for Indian production costs at this budget level — no association rate card, no published aggregate for a ₹3 lakh advertising shoot. Build your own baseline by taking blind quotes from two or three local production houses with an identical brief. The same gap applies to generation yield rates: no vendor publishes how many attempts produce one usable shot, so measure it on your first project and price the second from your own data.

Specifications in this area also move monthly. Clip lengths, resolutions, pricing and licensing terms on every tool mentioned here may have changed since publication, so verify against vendor documentation rather than any article, including this one. How we think about this whole area is set out in our approach to AI-powered marketing.

Want a hybrid scope built for your project? Send us the brief, your language markets and your delivery surfaces, and we will come back with which shots we would put a camera on, which we would generate, what the versioning matrix produces from one shoot day, where your ASCI labelling obligations land, and what we would require in writing about training rights before anything of yours goes into a platform. If the honest answer is a straight shoot, we will say so. Talk to Midgrow, or message +91 9632398896 with your category and language markets. Our digital marketing services page sets out how projects are structured.

Frequently asked questions

Can AI video replace a film shoot for an Indian brand?

Not at a ₹3 lakh and upward budget, where a competent conventional shoot is affordable. Generated video still fails reliably on sustained human performance, hands and faces at close range, continuity across cuts, text rendered in frame, physical plausibility and the exact fidelity of a real product — which are the shots carrying the message. Use it for environments, scale, stylised sequences and versioning instead.

What is AI video actually good at for advertising?

Shots that were previously out of budget: scale, aerial, crowds, impossible or period locations. Concept and pitch work before anything exists to film. Stylised or abstract sequences where photoreal fidelity is not the goal. Environment extension around live plates. And high-volume variant and version production from footage you already have.

Does AI dubbing work for Indian languages?

Partially, and quality varies sharply by language. Sarvam AI documents voice cloning and synthesis across 12 Indian languages and AI4Bharat's Indic-Parler-TTS covers 22, which is a real improvement. But a platform's language list describes what it accepts, not what it does well, and voice quality and lip synchronisation fail independently. Have every language version signed off by a native speaker before release.

Do I have to label AI-generated video ads in India?

From late December 2026, ASCI's guidelines require labelling where synthetically generated content materially shapes consumer decisions and non-disclosure could mislead. Synthetic presenters, replicated likeness or voice even with consent, invented settings, and demonstrations of non-existent products all require a label. Routine colour correction, decorative elements, AI-written copy and accessibility features do not. Fabricated endorsements and exaggerated synthetic product results remain prohibited whether labelled or not.

Can I use an AI-generated version of a celebrity's face or voice?

No. The Delhi High Court restrained unauthorised commercial use of Anil Kapoor's name, image, voice and persona in September 2023, and made similar orders protecting Jackie Shroff in May 2024, including against an AI platform using his personality attributes. Even with documented consent, ASCI's guidelines require a label where a real person's likeness or voice is synthetically replicated.

Who owns AI-generated video footage in India?

Ownership is unsettled. Indian copyright requires a human author, so purely generated footage may not be protectable, meaning you may be unable to stop a competitor copying it. Substantial human direction and editing improves the position, which is a practical argument for the hybrid route and for documenting the human contribution.

Will my brand assets be used to train AI models?

Possibly, depending on the platform and tier your agency uses. Runway's Terms of Use, last updated 23 October 2026, grant it a perpetual, irrevocable, royalty-free licence to use inputs and outputs to train and improve its models, with no opt-out described, while separately stating it does not claim ownership and does not restrict commercial use. Enterprise use is governed by separate terms. Ask your agency in writing which platforms and tiers handle your unreleased material.

Can I use AI-generated music in an advertisement?

With care. Udio locked downloads after its settlement with Universal Music Group, including for already-created tracks, which removes it for brand work. Suno ties commercial rights to subscription status at the time a track is created, so verify current terms. And a platform permitting commercial use is not the same as you owning the music, since purely generated output may lack the human authorship copyright requires.