The Pandora spec ad: one studio instead of an agency
Earlier this year we produced a spec ad for Pandora: a 33-second cinematic product showcase built around light, skin and silver. It was a deliberate stress test of our own pipeline. Could an AI video studio deliver the polish a jewellery brand expects, without a crew?
Think about what a spot like that normally requires. Models and a make-up artist. A filmmaker and a DoP. A hired studio, lighting rigs, wardrobe — then editors and a colourist in post. Every one of those line items is sensible on its own. Together, they are why a thirty-second brand ad can swallow a quarter's marketing budget.
We produced ours alone, in-house. On this project, we cut the production cost by around 70% compared with what the traditional route would have run. Not by making the film look cheaper — by removing the parts of production that exist to move people and equipment around, and keeping the parts that make an audience feel something.
The finished piece holds up as a brand film: macro passes over the jewellery, controlled reflections, a model whose gestures were directed rather than found in a casting call. An AI product video, in other words, that behaves like a commercial — because it was made like one.
This article walks through the project in detail — the process, the tools, the maths — because the Pandora piece is the clearest case study we have for how AI brand ads get made when the goal is conversion, not novelty.
This is the pipeline behind our AI cinematic video ads service — and the photoreal product work comes from our AI CGI & VFX side.
What makes a video ad high-converting?
A high-converting video ad earns attention in the first two seconds, says one thing the right viewer already cares about, and exists in enough variants for the market to pick the winner. Production polish matters, but it is the fourth ingredient, not the first.
When we break down what actually moves results on paid social — and Meta ads creative is where most of our clients' spend lives — it keeps coming back to the same four levers:
- The hook. Most viewers decide whether to keep watching before the second scene arrives. An ad that opens slowly is invisible, however beautiful the rest of it is.
- Message–market fit. The promise in the ad has to match a desire the viewer already holds. A jewellery ad sells the moment of giving, not the metallurgy; e-commerce video ads that convert are usually the ones that name the customer's situation before they name the product.
- Volume of variants. Nobody knows in advance which opening, angle or line will convert — not us, not the platform, not the brand. The advertisers who win are the ones who can afford to find out.
- Speed of iteration. When the data shows something working, the ability to feed the account more of it within days — not next quarter — compounds every other advantage.
Notice that three of the four levers are about quantity and speed, not craft. That is precisely why AI production changes the economics of high-converting video ads: it makes variants and iteration cheap while leaving the craft untouched. The rest of this piece shows how that works in practice.
How much does a 30-second brand ad cost to produce?
Most of the cost of a traditional 30-second product spot pays for logistics, not for the finished film. Before a single frame is shot, the agency route books people, places and equipment — and every booking is a line on the invoice.
Take the Pandora brief. Produced conventionally, a jewellery ad at that level would typically book:
- One or two models, found through casting calls and paid day rates plus usage rights
- A director and DoP to design and execute the camera work
- A gaffer and lighting team, because jewellery is essentially photographed light
- A make-up artist and wardrobe stylist, so skin and styling survive a macro lens
- A hired studio with rigging, plus camera and lens rental for the shoot days
- Editors and a colourist in post-production, often at a separate facility on their own timeline
None of those roles is padding. Each exists because physical production genuinely needs it. But when the set becomes a generation pipeline, most of those line items have no equivalent — the work they represent collapses into direction decisions made at a desk.
Here is how the two routes compare for a 30-second product spot, based on how we actually produced the Pandora film:
| Line item | Traditional agency production | Pelisson AI pipeline |
|---|---|---|
| Team size | Ten or more people across departments | One studio, working in-house |
| Casting and talent | Casting calls, model fees, usage rights | Directed AI performers and AI avatars |
| Studio and equipment | Hired stage, lighting rigs, camera package | Each shot briefed and generated at a desk |
| Shoot days | One to two days on set, plus prep and wrap | None |
| Edit and grade | External editors and colourist | Cut, mixed and graded in-house |
| Timeline | Weeks from brief to master | Days from brief to master |
| Relative cost | The benchmark | Around 70% less on our Pandora spec ad |
A note on that figure, because we are careful with claims. The ~70% saving is our own number, from this project: what our in-house route cost against a realistic costing of the traditional one. It is not an industry statistic, and a different brief may land differently. But the structural point holds for any e-commerce video ad — once casting, sets, lighting and wardrobe become generation decisions, the AI ad production cost is dominated by skilled hours rather than logistics.
For the wider comparison, see how AI video costs compare to traditional production in our FAQ.
How we work: from brief to 4K master
The Pandora piece followed the same process every client project does. AI sits in the middle of it, but the beginning and the end are stubbornly human.
Brief, script and story
We start where any good agency does: what is this ad for, who is it for, and what should they do after watching it? From there we write the script and build a shot list — hook, story beats, product moments, end card. On Pandora, that meant treating the jewellery as the protagonist and writing every shot around how light moves across it.
This stage is also where conversion gets designed, not just story. The shot list is written with the platform cuts already in mind: which moments must survive a vertical crop, where the product needs to be by second three, and what the end card asks the viewer to do.
AI production, human-directed
Then we generate. Each shot is briefed like a real setup: lens, camera movement, lighting, mood. We iterate takes the way a director shoots coverage — most get discarded, the best get pushed further. Nothing lands in the timeline because a model produced it; it lands because it serves the story.
The models generate footage. They don't make decisions. Taste, story and restraint are still the job — and that job is ours.
The same discipline applies when a brief needs a presenter rather than a product hero. For UGC-style campaigns we direct AI avatars the way we would direct performers — cast for the audience, scripted to the hook, lit to match the brand world — then cut them like any other footage.
Edit, sound and grade
From there it is a normal post-production job. We cut the film, design the sound, and grade it so every shot speaks one colour language — the grade is usually what separates an AI clip from a brand ad. On Pandora, it pulled a dozen separately generated shots into a single world of warm skin and cool metal.
Delivery: one 4K master, every platform cut
Every project leaves the studio as a finished 4K master plus platform cuts in 16:9, 9:16 and 1:1. One story, ready for YouTube, Meta, TikTok and wherever else it needs to earn its keep.
Which AI tools do we use — and how do we choose them?
We use whichever model wins the shot — right now that means Kling 3.0 and Seedance 2.0 for motion, with ChatGPT Image 2 and Nano Banana Pro supplying frames and product-true stills. There is no house model, and there never will be.
Kling 3.0
Kling 3.0 carries most of our photoreal motion. Skin, fabric, water, the way a chain settles against a collarbone — it renders physical behaviour convincingly enough to survive a macro-style close-up, which is why it did the heavy lifting on the Pandora jewellery passes.
Seedance 2.0
Seedance 2.0 is remarkably strong on multi-shot sequences and camera blocking. When a scene needs several cuts that feel like one director shot them — matched eyelines, consistent geography, deliberate camera moves — it is usually the first model we brief. It thinks in coverage, which is rare.
ChatGPT Image 2
ChatGPT Image 2 is our art department. We use it to explore looks, lock compositions and design end cards before any video is generated. Getting the frame right as a still is far cheaper than discovering a layout problem in motion, so a surprising amount of directing happens here first.
Nano Banana Pro
Nano Banana Pro is what we reach for when the product must be exactly the product. It handles product photography with AI at a fidelity that keeps clasps, engravings and logos true to the real item — and those stills then drive the video generation whenever a shot demands exact control.
We stay deliberately tool-agnostic. These models leapfrog each other every few months, and loyalty to any one of them would mean shipping last quarter's quality. Our job is to know what each is best at right now, and to pick per shot — the way a DoP picks lenses.
From one master to every platform
A finished master is the start of the media plan, not the end of the project. Every delivery includes the three video ad formats paid social actually runs on:
- 16:9 for YouTube, website embeds and connected TV
- 9:16 for Reels, TikTok and Stories, reframed shot by shot rather than centre-cropped
- 1:1 for feeds, where the square has to protect both the product and the performance
Each cut is rebuilt, not resized. A composition designed for a cinema-wide frame rarely survives a vertical crop by accident, so the 9:16 version gets its own framing decisions — and in our pipeline, that sometimes means regenerating a shot natively for the format instead of fighting the crop.
Hooks first
High-converting video ads are usually won or lost in the first two seconds, before the story has a chance to work. So we treat hooks as their own craft: several distinct openings on the same body of the ad, each making a different promise to a different kind of viewer. For Pandora, the same 33 seconds could open on the clasp, the glance or the gift box — three ads for the price of one edit.
A steady creative testing cadence
Creative testing works best as a rhythm, not a rescue mission. The pattern we build for clients is simple: launch a small family of variants, let the platform spend against them, read the results, keep the winners and replace the losers. Because new variants cost days rather than shoots, that loop can run continuously instead of once a quarter.
Refresh before creative fatigue
Every ad wears out; audiences simply stop noticing it. Traditionally, replacing tired creative meant another shoot and another quarter. With our pipeline, we refresh in days — new hooks, new scenes, same brand world. Staying ahead of creative fatigue is one of the most reliable ways we know to keep cost-per-result healthy.
What does this mean for your ad budget?
It means the same production budget buys more attempts at finding the ad that works. That is the whole argument, and it is worth stating plainly.
A traditional budget typically produces one hero spot and a couple of cutdowns, then rides them until they fatigue. The same spend through our pipeline produces a cinematic master plus a family of hooks, scenes and format-native variants — which changes what your media buying can do, because an algorithm can only optimise among the creatives it is given. More genuinely different creatives means more chances for one of them to be the outlier that carries the account.
We won't promise a guaranteed return multiple, because nobody honestly can — results depend on offer, price, audience and a dozen things outside the edit. What we can say from our own work is narrower and more useful: production stops being the bottleneck, testing becomes affordable, and fresh creative is always days away rather than a quarter away. In paid social, those three conditions are what a healthy cost-per-result tends to grow out of.
The short version
- Our Pandora spec ad — a 33-second cinematic jewellery ad — was produced entirely in-house, for around 70% less than the traditional route we costed. That figure is our own experience, not an industry claim.
- The roles didn't vanish; they collapsed into direction. Casting, sets, lighting and wardrobe became generation decisions, while script, edit, sound and grade stayed hands-on craft.
- We pick models per shot — Kling 3.0 for photoreal motion, Seedance 2.0 for multi-shot blocking, ChatGPT Image 2 and Nano Banana Pro for frames and product-true stills — and stay tool-agnostic on principle.
- Every delivery is a 4K master plus rebuilt 16:9, 9:16 and 1:1 cuts, with multiple hooks on the same body of the ad.
- The conversion story is volume and speed: more variants tested for the same spend, refreshed before creative fatigue sets in — with no guaranteed multiples promised, ever.