← Data Into Labor · All essays

Data Into Labor · Essay 04 · V2

The Next Ad

July 2026 · AJ Orbach

Hand a frontier model a prompt and it will produce beautiful media in a vacuum. Which is to say, it will produce the world's average taste: technically gorgeous, spiritually anonymous, could be anyone's ad for anyone's product. The industry calls this a quality problem. It isn't. The models are astonishing. It's a context problem, because a model that has never met your brand can only guess at it, and a guess with perfect lighting is still a guess.

People will tell you the deeper issue is that taste is subjective, and machines can't have it. In most fields that argument is hard to settle. Advertising is the exception, because advertising has a scoreboard. An ad either returned its spend or it didn't, and attribution knows which. Once taste has a scoreboard, taste becomes a thing you can study, build, and check. That's what this essay is about.

Creative is a job. Not a button, not a tab marked "generate": a job, with a desk, a deadline, and a client who knows what they like. The last essay built the media buyer, the person who decides where the next dollar goes. This one is about the twin desk: the studio that decides what the dollar buys when it lands. And the quiet scandal of performance marketing is that the bottleneck moved. Detecting a fatigued creative takes an hour; replacing it takes three weeks and an invoice. Every testing roadmap in ecommerce is gated not by ideas, and not by budget, but by production. So we asked the same blunt question we asked about the buyer: if you had to build a creative department from scratch, human or machine, what would you hand it on day one?^[1]

Briefs

Start where every real creative engagement starts: the brief. Most generative tools skip it, which is exactly why their output has the world's average taste. So the first instrument in the studio is not a model at all. It's everything the studio reads before it renders a single pixel. The brand's DNA — voice, palette, typography, visual motifs — held as structured context rather than a PDF nobody opens. The product catalog, so the thing in the frame is the thing in the warehouse. Your own top ads, ranked by attributed ROAS rather than by whoever argued loudest in the retro. Past generations, so the studio remembers what it already tried. And competitor creative, filtered by platform, format, and how long it has been running, which is the closest thing advertising has to a confession of what works.^[2]

The buyer's essay drew a line between trivia and instruments, and the same law holds here: a mood board is trivia. A reference set ranked by your own attribution is an instrument. Generation is a commodity. Context is not. This is why the studio lives inside the data platform instead of beside it — the system that already holds your brand, your catalog, your ad accounts, and your customer journeys is the only place a brief can be assembled without a human spending a week assembling it. The studio and the P&L finally share a building.

Engines

Now the part everyone fixates on, which is deliberately the second thing and not the first: the models. The studio keeps six engines on the bench, four for stills and two for motion, and treats none of them as a favorite.^[3] They have different temperaments — one is best when the instructions are dense and the edit is surgical, one for photoreal detail at 4K, one for fast cheap iteration, one for stylized looks that shouldn't look like a photograph at all — and the brief decides. Routing a job to the wrong engine is the new version of sending a product shot to a portrait photographer. The studio's job is to know the difference without being told.

Two capabilities matter more than any single engine. First, reference-carrying: a single render can hold up to sixteen brand references — logo, product, palette, past winners — which is the difference between "looks nice" and "looks ours." Second, range: aspect ratios from 1:1 to 21:9 and resolutions from quick 1K drafts to 4K masters on the image side; 480p proofs to 4K plates in eight, fifteen, or thirty second cuts on the video side. A creative department that can only make one shape of thing is a vendor. This one makes the whole package.

Desks

Watch the still desk work. A campaign starts with the anchor: the cinema-ratio hero that sits at the top of a landing page or a launch email, the frame that decides what the whole system feels like. From the anchor, the desk fans out into everything a launch actually ships. Paid social cuts with the headline set inside the model itself — type as part of the composition rather than a caption bolted on afterward — then re-read character by character in QA, because text is where image models still lie most fluently.^[4] Iteration sheets: the same product under four art directions, minutes apart, because taste is choosing and the desk exists to make choosing cheap. The expensive part of creative was never the winner; it was the three losers you had to shoot to find it. And story frames in 9:16, the same visual system turned vertical, because a campaign is not an image. It's a system of images that agree with each other.

Video is where generative tools usually stop being a studio and start being a slot machine: one prompt, one clip, pull the lever again. The motion desk refuses the premise. It works the way a director works, in treatments: ordered scenes with per-scene action and camera notes, negative prompts for what must not happen, image references for what must, and human-aware routing when there are people in the frame.^[5] There are three doors into a shot, and the desk uses all of them. Text-to-video, when the treatment starts from language. Image-to-video, when it starts from an approved still — and this matters more than it sounds, because it means the still campaign and the motion campaign are the same object at two frame rates, not two vendors' interpretations of a brand deck. And video-to-video, the door almost nobody has: hand the desk existing footage and a new art direction, and it re-renders the shot without a reshoot. Re-art-directing footage used to be a production line item. Now it's a revision note.

Post

Between the render and the placement sits the unglamorous work that decides whether any of it ships: the cut, the captions, the thumbnail, the eleven aspect-ratio exports nobody dreams about. The studio's rule for this floor is simple: the parts of post that must be exact are done in code, not sampled from a model.^[6] Playable motion pieces run on deterministic timelines — captions that appear when they're supposed to, kinetic type that lands on the beat every single time. Thumbnails and first frames are extracted programmatically. Formats are packaged per placement, not re-generated and hoped for.

The distinction sounds technical and is actually philosophical. A generative model samples, and sampling is where the magic and the misspellings both come from. An edit bay must not sample. When the timing, the type, or the packaging has to be identical on the thousandth export, the studio stops being generative and becomes deterministic, deliberately. Knowing which floor of the building you're standing on is most of what "production quality" means.

Verdicts

Here's the floor that separates a studio from a slot machine: the screening room. Nothing leaves the building unwatched. Every render, still or motion, is inspected before delivery and returns a verdict and a caveat, not a vibe.^[7] The inspector knows what generative work gets wrong because it has seen it get it wrong at volume: hands with ambitions, physics that forgot gravity, typography that dissolves on the second read, a colorway that quietly drifted off-brand between iterations. Discipline follows the verdict. A failed render gets one targeted retry with the defect named, and if it fails again the desk changes the route — a different door, a different engine, a different treatment — rather than pulling the lever until something pretty happens.

This is the answer to "machines can't have taste." Taste here is not a vibe. It's a checklist that runs every time, which is the only kind of taste that scales. An honest "the reflection on frame six is soft" is worth more than a confident gallery of unexamined output.

The loop

Now the instrument no standalone creative tool can have, no matter how good its renders get. Every other studio in the world ships the work and waits: for the agency report, for the quarterly deck, for someone to remember to tell them what happened. This studio ships into the same system that measures. The ad it made lands in an ad account the platform already tracks, spends dollars the platform already attributes, and shows up in creative performance by ad, in the same schema as the journeys and the orders. The distance between "we made it" and "we know what it did" collapses to zero, permanently.

Fold that back in and the loop compounds. Winners become references for the next render, so the studio's taste drifts toward what your market actually rewards rather than what a mood board suggested in January. Fatigue flags become briefs. Competitor moves become context. The reference set that made the first render "ours" instead of average is re-ranked by attribution every day the ads run. The studio learns the way a great creative team learns, except it never forgets and never leaves for a competitor.

Put this essay next to the buyer's and the shape of the thing appears. The media buyer and the creative department were never really separate jobs; they were one loop cut in half and mailed to different vendors. The buyer flags a fatigued ad at 9:00. The studio has on-brand variants in the screening room by 9:15. The governor asks a human to approve the swap. Attribution grades the winner by Friday, and the winner's DNA is in the next brief by Monday. No handoff, no agency queue, no three weeks, no invoice. The loop that used to cross four companies now crosses a hallway.

Notice, one last time, what the scarce part is. The engines will keep improving, and every improvement arrives on the bench the week it exists — that's the point of a router, and the reason the studio has no favorite model. But no frontier lab is going to hold your brand's DNA next to your attribution, rank your references by your own ROAS, keep a competitor shelf filtered by run time, or file a QA verdict on every render it ever made for you. That's not what labs build. It's what we build.

What comes out the other side is the thing every growth team has actually been asking for: more shots on goal, each one aimed. Volume without taste is spam. Taste without volume is a bottleneck with a portfolio. The studio delivers both at once, with the receipts attached — every asset labeled, inspected, and measured against the only critic that matters, the market. The buyer never runs out of money to place. Now the brand never runs out of ads worth placing it on.