Seven channels.
Seven format contracts.
Every limit enforced in code, not requested in a prompt.
One of these is included. The other is a hard charge.
A lot of platforms blur the studio and the asset factory into one word. They are kept apart here because they are charged differently and they fail differently, and a buyer who does not know which is which cannot audit either.
The studio
A brief becomes channel-correct copy and structure: headlines inside the real character limits, audio scripts with read direction, DOOH copy rendered to genuine SVG, storyboards carrying the trafficking metadata the channel will ask for. Then the clearance pre-check runs over it, and you can revise any of it in your own words. None of that is charged for. It sits inside the 10% media fee.
Asset production
Turning that copy into finished video, images, voiceover, music and motion loops costs a paid provider real money at the moment you press submit. So it is a hard charge on its own statement line, debited from the wallet at submit, never blended into the media fee. You can also skip it entirely, bring your own finished assets, and use the studio for the channel adaptation and the pre-check alone.
What each channel will actually accept
The creative studio produces the right shape of creative for each channel from one sentence: responsive search copy inside Google's real character limits, paid social copy bounded to a portable envelope that is valid on both Meta and TikTok, fifteen and thirty second audio scripts with voiceover and SFX direction, DOOH copy rendered to genuine SVG at UK screen resolutions, CTV and cinema storyboards with the trafficking metadata attached, and TV scripts carrying Clearcast substantiation notes. Every one of those limits is a schema constraint, so copy that cannot fit fails at the brief boundary rather than arriving truncated at a provider.
"Keep every headline under 30 characters."
A limit a model is asked to respect is a suggestion. It produces thirty-one often enough to matter, and the over-long headline travels all the way to a provider before anyone finds out — where it is truncated mid-word, or rejected, and either way you learn about it late and in someone else’s error format.
headlines: z.array(z.string().min(1).max(30)).min(3).max(15)
The same rule as a validator on the brief boundary. Thirty-one characters is a 400, every time, on every path — and so is a second headline where the format needs three, because a minimum is as enforceable as a maximum once the limit is a type rather than a request.
Paid search
Responsive search ads
3 to 15 headlines at 30 characters each and 2 to 4 descriptions at 90 characters, as schema minimums and maximums. Fewer than three headlines is as invalid as a thirty-first character.
Paid social
Bounded social copy
Headline 40 characters, description 30. Conservative portable limits chosen to sit inside TikTok's tighter 100-character text envelope, so one creative is valid on Meta and TikTok rather than needing two.
Podcast & audio
15 and 30 second scripts
Written with voiceover direction and SFX notes rather than as bare copy, because a script with no read direction is half a brief and the read is most of the ad.
DOOH
Real SVG creatives at UK screen resolutions
1080×1920, 1920×1080 and 1400×400, rendered rather than mocked up. Headline capped at seven words. Legal copy is held separately from the headline, subline and call to action, because those three are word-clamped and legal text is not.
CTV & streaming
Storyboards with VAST metadata
16:9 ProRes 422 HQ mezzanine at 1920×1080, 25fps for the UK, at least 20Mbps. Supers inside 90% title-safe, loudness to EBU R128.
Cinema
Storyboards and a DCP delivery spec
2K flat at 1998×1080, 24fps, 5.1 audio, 60 seconds maximum, routed via DCM Cinemapper with a CAA clearance certificate before print traffic.
Addressable & linear TV
TV scripts with substantiation notes
Every superlative or market-leadership claim is flagged with the substantiation Clearcast will ask for, on the record, before it reaches a clearance queue.
Legal and mandatory lines are held apart from every word-clamped field, so a “Ts&Cs apply” that will not fit inside a seven-word DOOH headline fails as its own bounded field rather than being quietly shortened into non-compliance.
Bring a brief. We will run it through the table above.
Every limit in that table is enforced at the boundary rather than downstream, and the quickest way to believe that is to watch a real brief hit one.
Book a working session45 minutes. Bring a real brief and we compile it live.
Clearance, predicted before it costs you a week
A pre-check runs every creative against the UK codes that actually decide whether it airs, and returns the findings rather than only the verdict: a blocker fails it, a warning flags it, and each finding names the rule it came from. Its rule checks are deterministic; where a model reads the copy, the finding says which path produced it. It is a prediction of what Clearcast, the CAA and publishers will flag, produced before you commission the asset.
Objective and superlative claims — best, number one, market-leading — need documentary substantiation. The pre-check finds the claim and names it.
Medicines, medical devices, treatments and health claims. Clearcast requires evidence and may block outright; an unsubstantiated health efficacy claim is routinely rejected.
Cinema copy containing strong language restricts pairing to 15 and 18 certificate films, which is a media constraint as much as a creative one.
Four more in the same catalogue
- BCAP 14
- Financial promotions. Guaranteed returns, risk-free and 100% safe are blockers outright; investment, credit, loan and APR language draws a warning for the risk statement, the representative APR and the FCA-authorisation detail.
- CAP 18 + venue rules
- Alcohol on digital out-of-home. Media owners require screens within 200m of a school to be excluded, so the finding is a media constraint before it is a copy one.
- HFSS restrictions
- Less-healthy food and drink. No targeting of under-16s, and the Transport for London estate prohibits HFSS creative outright, which is a media exclusion as much as it is a copy finding.
- CAP 3.17
- Significant conditions on a price claim. A price, a percentage off or a free offer with no visible conditions is flagged, and the check reads every UK house style of the qualifier rather than one spelling of it.
A health efficacy claim with no evidence, or guaranteed-returns language.
A superlative with nothing on file, a price with no conditions, alcohol on DOOH.
Recorded and carried forward, including a claim relying on evidence you have declared.
One verdict rule, recomputed from the findings on both the model path and the deterministic one: any blocker is a fail, any warning is a flag, otherwise it passes. A model cannot return a clean verdict alongside a blocking finding, because it is not the thing that decides the verdict.
It cannot read your evidence
The pre-check cannot read a substantiation document and does not pretend to. Where you declare that you hold evidence, it records the declaration alongside the claim and says explicitly that it does not assess evidence quality — Clearcast will require the documents themselves before clearance. Both the claim and the declaration stay on the record, which is the useful thing to hand the body that actually decides.
Finished assets, not just copy
Five modalities, generated through the same provider stack, priced from the configured provider cost bases at run time and charged as their own statement line.
Video
fal.aiThree tiers: Wan for drafts, Kling v2.5 Turbo Pro for standard, Kling v3 Pro for premium. Five to ten second clips, 16:9, 9:16 or 1:1.
Image
FLUX via fal.aiStills for social, display and DOOH backplates.
Voiceover
ElevenLabsUK male, female and neutral voices, at slow, normal or fast pace.
Music
stable-audioBeds from 5 to 180 seconds.
Motion
Programmatic or AI-rendered5 to 15 second loops. Programmatic motion is generated SVG and effectively instant; rendered motion is an AI animated-text loop of 5 to 10 seconds. Both output at 1080×1920, 1920×1080 or 1400×400.
Without provider keys, every modality returns a deterministic placeholder rather than an error, so the whole creative loop runs end to end before anyone pays for a render.
The published ladder
retail, per asset- Image (FLUX)at the 50p floor
- 50p
- Voiceover (ElevenLabs)UK voices, male, female and neutral
- £1.25
- Music trackat the 50p floor
- 50p
- Programmatic motion loop5–15s SVG, generated instantly
- 50p
- Rendered motion loop, 10sAI-rendered animated text
- £2.00
- Video, 10s standardKling v2.5 Turbo Pro
- £6.00
- Video, 10s premiumKling v3 Pro
- £15.60
Priced at 80% gross margin on raw provider cost — retail is cost divided by one minus the margin, so 5× cost, with a 50p floor — debited from the wallet at submit. A distinct hard charge with its own statement line, never blended into the 10% media fee.
The 80% is a pricing rule against raw provider cost, not an achieved gross margin: it excludes storage, compute, egress, support and overhead. The cost bases behind the ladder are also explicitly unverified against a live provider invoice, and and we must reconcile them before charging real customers. Prices reprice from configuration, so read the ladder as a current default.
A definite provider failure refunds the charge automatically. An ambiguous outcome does not: the charge is retained and retry is disabled, so a duplicate render cannot be produced and billed, and it is cleared by a named human with reconciliation rights.
Nothing is charged below 50p, which is why an image, a music track and a programmatic motion loop all share the bottom rung. The SVG loop is generated in this process and costs a provider nothing at all, and the raw provider cost of an image or a track is small enough that the margin rule lands under the floor as well.
Where the model stops
The copy above is written by a model. None of the numbers, identifiers or limits are, and that split is not a matter of style. It is the reason the output of this studio can be handed a budget, and it is worth being exact about where the line falls.
The model writes. Code does the arithmetic.
Plan generation is grounded in adapter forecasts and every number in a plan is computed deterministically. A language model is not asked to add up a budget, because a model that is wrong about arithmetic is wrong in a way that looks right.
The model never emits a vendor id
It extracts semantics; registry tables emit segment identifiers. A hallucinated segment cannot reach a buy because the model was never holding the pen at that point.
Limits are schemas, not instructions
A prompt asking for thirty characters produces thirty-one often enough to matter. A schema rejecting thirty-one produces a 400 at the edge, every time, on every path.
There is a full deterministic fallback
With no API key the persona extractor, plan generator, pre-check, optimiser and fidelity narrator all run on heuristics instead. The product works without the model; the model makes it better rather than making it possible. AI-produced output is visibly distinguished from deterministic fallback output wherever it appears.
The AI layer runs on Claude — claude-opus-5 by default, overridable — for persona extraction, planning, creative pre-check, optimisation and fidelity narration.
Each AI operation returns a record of how it was produced: an execution mode of model, deterministic or mixed; the model that was attempted; and, where it fell back, which of four typed reasons applied — a missing API key, a provider error, a refusal, or structured output that failed its own schema. A deterministic fallback is never presented as live model output.
Questions about the creative
Rejection risk, whether you have to use any of it, why the price is visible rather than bundled, what happens when a provider fails halfway, and where AI disclosure is enforced for you.
Will the AI write ads that get rejected?
It will write ads that are structurally valid, because the format limits are schema constraints rather than instructions, and it will flag the claims most likely to be challenged before you spend anything on production. What it cannot do is guarantee clearance: the pre-check predicts what Clearcast, the CAA and publishers will flag, and those bodies decide. Where you declare that you hold substantiation for a claim, the system records the declaration and states plainly that it has not assessed the evidence.
Do I have to use the AI creative?
No. You can bring finished assets and use the studio only for the channel adaptation and the pre-check, or use it for structure and produce the assets elsewhere. The asset charge only applies to what you actually generate here.
Why is ten seconds of premium video £15.60 when competitors include AI creative free?
Because free creative is not free, it is bundled, and a bundled cost is one you cannot audit. £15.60 is the top of the ladder rather than the price of video: ten seconds of standard video is £6.00. It is a genuinely harder sell — Universal Ads and Global AdPower both include AI creative — and the honest case is that a visible margin on a small render is better for you than an invisible one inside a media percentage. What this page will not do is quote a "versus traditional production" saving to make the number look generous: we deleted our own such comparison as unsourced and added a test to stop it coming back.
What happens if the provider fails halfway through?
It depends on whether the failure is definite. A definite failure refunds the charge automatically to the wallet, as its own ledger entry referencing the failed job. An ambiguous outcome — where the provider may or may not have produced the asset — retains the charge and disables retry, precisely so a second render cannot be produced and billed, and it is resolved by a named human with reconciliation rights. There is no warn-only path for asset charges at all: a real provider cost is incurred at submit, so the wallet check is a hard gate with no override.
Who owns the AI-generated assets, and are they indemnified?
You get the asset and you can use it commercially, and two things are weaker than you might assume. The provider's terms do not grant ownership of generated output and note it may not be unique across customers. And there is no intellectual-property indemnity: under those terms the customer indemnifies the provider rather than the reverse, and output is not warranted original or non-infringing. The default image model, FLUX.1 [schnell], is Apache-2.0 and openly licensed for commercial use; the closed-weight video and audio models rest on the host's documentation rather than a contractual warranty. Each model's position is served next to its price, and an unlisted model reports as unverified rather than permitted.
Does AI-generated video need to be disclosed?
On TikTok, yes, and the platform enforces it rather than leaving it to you. Video produced here is submitted with a self-disclosure provenance declaration, missing provenance blocks the call before any provider request is made, and the provider readback must still carry the same immutable declaration at resume.