How AI studios actually ship brand work

The counts tally every mention across a contributor's answers, from main stack to lean setup. No one runs the whole table. What the working studios share is narrower and more boring than the tool discourse suggests: a handful of tools they trust under deadline, and a practised indifference to everything else.

Script and brand research

Claude is the front of almost every pipeline: brand research, scripts, shot lists, prompt systems. Christophe Martin, founder of Eugène Studio with 25 years leading campaigns for global brands, goes furthest: he runs Codex and Claude Code so the creative process itself becomes a system rather than a sequence of prompts. Charly, AI video ads creator for DTC brands, drives Higgsfield's picture and video generation from inside a Claude conversation over MCP.

Images: reference first, always

The craft rule that separates this panel from hobbyists is that nobody goes straight to video. Agha Abdullah, UGC and product-commercial creator partnered with several AI platforms, locks character and environment reference sheets before any video generation; in his experience, final-cut consistency problems almost always trace back to skipping that step. Olena Sinner, who has shipped 190+ product videos for US Amazon brands, never runs text-to-video for client work at all: every shot starts from an approved still frame, so the client signs off on composition and label accuracy before any credits are spent on motion. And Anirudh Kandpal, founder of creative studio KRI4TIV, gives the rule its shortest form: “a weak source image will usually create a weak video, regardless of how good the video model is.” Three contributors, three independent phrasings of the same law. Image-first is the closest thing this cohort has to a universal rule of production, and it is the conclusion the filmmaking panel reached from the craft side.

Video

Kling is the volume workhorse, named by seven, and the heaviest shippers pick it for the least romantic reason: it does what the prompt says, shot after shot. Seedance 2.0 is the specialist, reached for when a sequence needs multiple shots with one consistent face. Higgsfield is the authored-motion pick for cinematic ad work, and the panel splits on it hard enough that it gets its own treatment below. Veo appears in three stacks, usually via Google Flow.

Sound and the cut

ElevenLabs for voice, Suno for music, and the cut has not changed: CapCut for social-first delivery, Premiere Pro and After Effects where a client expects broadcast finishing. Ritesh Parihar, video editor with five years in traditional post, keeps the full Adobe chain and slots AI in as an accelerator rather than a replacement. For image generation he runs ComfyUI with Flux models locally, trading a steeper learning curve for the control and reproducibility commercial work demands, the same pro-control frontier our filmmaking panel described.

The quiet shift: the agent becomes the interface

Look at how the most agile contributors actually operate and a pattern emerges that no survey question asked about. Or Ben David, award-winning one-man production house and 2026 AI Mover and Shaker, works agent-first from a visual-canvas agent that carries every model he uses daily. Charly's Claude-to-Higgsfield setup is the same move. Two contributors independently named OpenArt Director, where you talk to an AI and the work gets done, as their smallest viable stack. And Anu Anuja, founder of ad agency Skyrise AI Studio, builds her own agentic pipelines directly against model APIs. Nobody called this a trend. Five stacks quietly being rebuilt around it is one.

The most citable single finding in this issue comes from the contributor with the most shipped volume. Product work lives or dies on whether the label survives a client zooming in, and Olena's field verdict after 190+ videos is exact: “There are three tools that survive client zoom-in: Nano Banana Pro, GPT Image 2, and Flux. Everything else distorts label text sooner or later.” Her one caveat: Flux tends to darken the frame, so she watches exposure.

This is where the field data meets our lab data. The Vibedex image leaderboard covers 18 models, blind-judged and scored on the VibeDex Score, our 0–5 blind-benchmark score. On it, GPT Image 2 (High) leads outright at 4.155, and the Nano Banana and Flux families sit in the upper and middle tiers. A shortlist built purely from client work tracks our blind leaderboard. When the people spending their own credits and the benchmark agree, that is about as strong as evidence gets in this field.

The strongest strategic through-line in the data is a business rule, not a tool choice. Aqib Shahzad, creative director of Scale My Beauty Brand, says stop selling “AI-powered” altogether: founders buy the outcome, meaning speed, scale and price point, not the tool. “The moment a client thinks 'AI-generated,' they start looking for flaws instead of results.”

Christophe frames the same instinct as craft discipline: decide what must remain invariant (product, character, visual language, the strategic idea) before generating anything, then change one variable per batch. And Evangelia Katsogiannou, fashion designer and 2026 AI Mover and Shaker in Fashion, names the failure mode of the opposite approach: the belief that prompting alone replaces professional expertise. AI can render a beautiful garment; it does not understand garment construction, textile behaviour, or decades of design knowledge.

Put together, the pitch that wins is the one where the client never has to think about the tools: the AI does the work, the human owns the taste, and the deliverable is judged as an ad, not as an AI ad.

We asked every contributor for their best ad by ROAS, CTR, CPA or hook-rate, and here is the honest answer: most do not have those numbers, for a structural reason. Creators and studios rarely own the media-buy data, and to their credit several said so plainly rather than inventing a figure.

What the data does contain is two real performance signals. Anu reports a UGC-style Meta ad that hit roughly 2× ROAS, the strongest creative in its campaign and the one hard paid-performance number in the dataset. And Olena, whose videos run as A/B variants on Amazon listings for a US supplement brand, distilled 190 videos of variant testing into a rule worth the whole section: “Openings where the product is already in use within the first second beat intro shots every time. The hook isn't the first shot, it's the first half-second.”

The rest of the evidence is organic. Evangelia's Caimera fashion-prototyping post reached ~6,000 impressions on a ~2,000-follower base with 62% of it out-of-network, because it solved a real problem, sampling friction, rather than showing off AI. Agha's most-engaged piece was an emotionally driven short made with a human director, not his usual commercial work. The pattern across all of it: performance follows the hook and the story, not the model choice.

The clearest before/after numbers in the dataset come from the people running alone.

Olena: one person, 190+ commercial product videos in about seven months, running 25 to 30 finished videos a month at peak, A/B variants included. Each of those used to be a studio day: product samples shipped, a set, a crew, hundreds of dollars per video minimum. Now a finished product video costs generation credits and a few hours. Her sharpest line is about what that actually changes: “testing 3–4 creative angles per product became normal instead of a luxury.”

Anu puts the agency-side number at roughly a 60% reduction in production time: teams that shipped a handful of ads over a month now create and test 10–20 variations in a week. One contributor's monthly tooling spend went from hundreds of dollars to under $100. None of these numbers is really about making one ad cheaper. They are about making the next ten ads testable.

Ask what no tool does well yet and the panel gives one answer at three altitudes. It is the same headline gap our filmmaking panel named, reached from the brand side instead of the story side.

Keep the face consistent. Everything today is workarounds: face passports, reference frames, the same image attached to every generation. Realism still falls apart the moment the character moves or smiles. Olena has the sharpest formulation in the dataset: “Today you have to choose: same face or real face. A tool that gives both would cut half my workflow.”

Keep the campaign consistent. Ritesh Parihar wants an entire campaign generated from a single creative brief with every character, product, environment and brand element held constant, plus single-shot edits that do not disturb the rest. Anirudh's version separates the controls: lock identity, body shape, framing, clothing and environment independently, then generate controlled variations without losing realism. Aryan Garg, co-founder of AI visual studio Oscar Studios, wants the same precision inside a single video: change a detail at one timestamp without regenerating the whole thing. Christophe extends it to feedback: a system that holds the brand world steady across image, video, sound and edit, and learns from campaign results “without gradually turning the brand into generic AI content.”

Keep the brand consistent. Anu goes up one more level: the real gap is that no tool truly understands a brand: its guidelines, past campaigns, personas, tone. Her ideal is “an AI marketing teammate rather than just a generation interface,” one that gets more valuable the longer it works with you. Evangelia's version, from fashion: “My ideal tool would think like a product developer, not just render like an artist.”

One smaller gap deserves its line. Charly wants real lip-sync control; the tech is nearly there, and the last few percent still ruins takes. Face, campaign, brand, and the fine control to fix any of them without starting over: any tool vendor reading this has just been handed a roadmap.

Ask this panel whether a studio should live inside one platform or assemble its own chain, and you get three incompatible answers, all of them currently paying rent.

The all-in-one loyalists. Charly runs everything through Higgsfield, full stop. Aryan backs the same platform on capability grounds while conceding it is expensive.

The specialised chain. Christophe Martin would rather wire specialised tools into a clear workflow than accept an all-in-one's compromises. Evangelia practises the same doctrine in fashion, building around the problem rather than the tool (“build it around the problem you're trying to solve”), routing fashion work to Caimera, editorial to Lumoo, and multi-model experiments through Flora.

Model-first, platform-agnostic. Anu treats the platform as an afterthought: “If a model performs well, the interface I access it through is secondary.” Her homework is model documentation and limits; the access point gets chosen per project.

Hold those three side by side. The platform two contributors call home is over-hyped to a third, and irrelevant to a fourth. We look at both layers ourselves: models blind-benchmarked, platforms evaluated on what models cannot show, meaning workflow, coverage and trust. On this panel's evidence, both camps are right about their own jobs.

The most-nominated target is not a product but a promise. Christophe named the one-click AI ad generator as a category: efficient at volume, but it tends to “flatten every brand into the same visual language.” Friction removed, taste not included. Anirudh landed on the same verdict independently: fully automated ad platforms strip out too much human judgement (casting, art direction, pacing, emotional tone) and the output reads generic as a result.

One contributor, published anonymously at their request, went a layer deeper: many AI platforms are “simply wrappers around the same underlying models,” and the new agentic creative features are overhyped for the same reason: a single-line prompt does not produce a polished commercial, and pretending otherwise is a demo, not a workflow.

The rest of the list is specific. Aqib finds Grok hyped beyond its results. The UGC-avatar lane took a hit from inside. Charly, who works UGC-adjacent DTC, found Arcads over-praised, citing skin-tone rendering, though he flags he has not used it much. And Seedance 2.0 got both barrels and a bouquet in the same dataset: Aryan says it blurs motion on VFX work; Olena watched it hold one consistent face across a four-shot drama scene in a single generation. Both are right. On this panel's evidence Seedance is a consistency instrument, not a VFX instrument. File the complaint under wrong-job-for-the-tool.

And one contributor called out a popular all-in-one video platform as slow, not genuinely free, and heavy on hype-driven features that underdeliver. That answer is anonymous at the contributor's request, and it cuts against three other contributors here who run the same platform as their primary or recommend it to beginners. Treat it as a live split in the market, not a verdict.

Three bets dominate the panel's view of the next 6–12 months, from July 2026.

1. Agentic campaigns, with an asterisk. Agha and Christophe both expect campaign systems that plan, optimise and extend a creative idea with minimal human input. In Christophe's version, persistent systems that remember a brand's visual world and connect creative decisions to performance data. Anirudh reaches the same destination by a different road: owned digital characters and reusable libraries of faces, environments, voices and visual identities, deployed across dozens of ads instead of regenerated for each one. The asterisk belongs to the anonymous contributor above, for whom today's agentic features fail exactly where ads are won: judgment, pacing, iteration. Read together, the disagreement is not whether agents come. It is who owns them.

2. The backlash against AI slop. Agha's bet, and his lane: “Brands that resist synthetic sameness are winning attention.” He is building cinematic, character-consistent work with real direction behind it while AI-heavy campaigns take visible public heat. Olena supplies the economics underneath: once everyone can generate a hundred videos, volume is worthless; survival goes to the people who “know what a good frame looks like before they generate it.”

3. Away from fake-authentic, toward openly cinematic. The most contrarian bet is Anu's, and it argues against her own best result: her one hard ROAS win was a UGC-style ad. Consumers, she believes, are getting skeptical of content that imitates real human reviews, and are more comfortable with clearly creative, cinematic advertising where AI is a production tool rather than a disguise. The lane that works today, on her read, is the lane that burns out.

All three point the same way as the pitch philosophy in section 04: AI stops being the selling point and becomes invisible infrastructure behind elite creative.

What the panel would tell you

If you are starting out

The panel's beginner advice is its most practical material. The consensus on-ramp is a single friction-free environment. Google Flow was named twice (Nano Banana for images, Veo for video, one clean interface), with Higgsfield-plus-CapCut and ChatGPT-plus-Magnific as alternates. But two contributors would have you start somewhere else entirely. Evangelia starts before the tools: learn to think with ChatGPT and Claude, better questions and sharper ideas, and only then pick up a specialised platform. Anirudh starts after them: ChatGPT plus CapCut, because the editor is where you learn pacing, captions and how a short-form ad is actually built. Learn the cut, then add one generation tool. Not ten.

And the craft to practise first is the image, not the video. Video, in Olena's telling, is not the hard part. Getting one good image and knowing how to describe what you want is, and it transfers: “video tools change every few months, but a good image and a clear prompt work everywhere.” Aryan's beginner picks are the tested-and-reliable pair, Nano Banana Pro for images and Kling for video. Boring on purpose, which after ten sections of this report should sound like the highest compliment a tool can earn.