An AI creative agency produces brand, advertising and design work using generative tooling as part of the production process. In practice that means far more variations per concept, much faster iteration, and a review process that has to move from checking taste to checking provenance.
The output got cheap. The judgment about what is worth making did not, and that gap is the entire engagement.
This covers what genuinely changes in the work, what does not change at all, and the rights and disclosure questions that decide whether the work is usable.
What actually changes
Volume and iteration speed
The obvious one, and it is real. A concept that produced three executions now produces thirty. Testing that took a month takes a week.
The strategic consequence matters more than the cost saving. When variations are cheap, the constraint moves from production capacity to decision capacity — someone still has to judge which thirty are worth running, and that person becomes the bottleneck. Teams that add generative tooling without adding decision-making structure produce more work and ship no faster.
Where review moves
Traditional creative review checks whether work is good and on brand. That does not go away, but two new checks sit in front of it:
Provenance — where did the elements come from and are they cleared for this use?
Accuracy — does it assert anything about the product that is not true? Generative tooling produces confident, plausible, wrong detail, and it does so in copy more often than in image.
Reviewers used to catching taste problems are not automatically equipped to catch either. That is a training and process problem, and it is where most of the risk in AI creative sits.
What gets cheaper and what does not
Cheaper: variations on a settled concept, resizing and reformatting, exploratory directions, localisation, mockups and comps.
Not cheaper: the strategy, the concept itself, the judgment about what to make, the rights clearance, and the accountability for what ships. If anything, the last two got more expensive, because there is more work to clear and more surface to be wrong on.
Any agency pricing the second list as though it fell with the first is mispricing — and any buyer expecting it to has misunderstood where the cost actually sat.
What doesn't change
The brief. A vague brief produced bad work at three executions and produces bad work at thirty, faster. Generative tooling is an amplifier and it amplifies imprecision efficiently.
Knowing what is worth making. Volume raises the ceiling on testing and does nothing for the quality of the hypothesis. Thirty executions of a weak idea is thirty weak executions.
Accountability. Whatever the production method, a person is answerable for what ran. "The model generated it" is not a position anyone has successfully held.
Distinctiveness. Widely available tooling trained on overlapping data produces convergent output. The pressure on brands to look like each other has gone up, not down, and resisting that is a deliberate act now.
The provenance question
This is the operational core, and it is where an agency either has a process or does not. Five things belong in writing before work starts.
Which tools, and on what licence tier. Commercial rights vary by product and plan. A tool used on a personal or free tier may not grant the rights you need. Get the specific products and tiers named.
Training and input data. What was fed in? Work generated from a competitor's assets or from unlicensed reference is a problem regardless of how the output looks.
What is cleared for paid media. Organic and paid carry different exposure. Some licences and some likeness permissions do not extend to advertising.
Likeness and voice. If a face or voice appears, what is the release, does it cover synthetic modification, and does it cover the duration and channels you intend?
Disclosure. What is being disclosed, where, and to whose standard — the platform's or the regulator's.
Where disclosure is settled
If AI-generated or AI-modified creative is used in anything that functions as an endorsement, testimonial or review, the FTC's Endorsement Guides apply and they are not ambiguous. The Guides hold that endorsements "must be honest and not misleading," that an endorsement "must reflect the honest opinion of the endorser," and that it "can't be used to make a claim the marketer of the product couldn't legally make." The Guides also require that a material connection a significant minority of consumers would not expect be "disclosed clearly and conspicuously."
The practical reading for AI creative: a synthetic presenter is not automatically prohibited, but it must not be presented as a real customer's genuine experience, and it cannot carry a claim the advertiser could not make directly. This is SETTLED. Agencies uncertain on this point should not be producing testimonial-style creative for you.
The creative partner scorecard
Ten factors, rated 1 to 5. Same rubric we run internally.
| # | Factor | What a 5 looks like |
|---|---|---|
| 1 | Strategic input | They interrogate the brief before producing |
| 2 | Concept quality | Distinct ideas, not variations on a stock direction |
| 3 | Tool transparency | Named tools, named licence tiers |
| 4 | Rights process | Written clearance process, not assurances |
| 5 | Likeness handling | Releases cover synthetic modification and intended channels |
| 6 | Disclosure standard | They can state the FTC position without checking |
| 7 | Accuracy review | A named step where product claims are verified |
| 8 | Distinctiveness | Their work does not look like everyone else's |
| 9 | Volume with judgment | They recommend what to run, not just what exists |
| 10 | Asset ownership | Source files, prompts and rights transfer to you |
Score = 100 × sum(rating − 1) / 40.
Where a factor cannot be assessed, score the range — 1 as the lower bound, 5 as the upper — and treat the gap as the question to resolve before signing rather than splitting the difference.
The review workflow that actually catches things
Most creative review was designed for a world where producing an execution was expensive, so few executions reached review and each got real attention. At thirty variants per concept that model breaks, and the failure is silent — things get approved because nobody had time.
A workflow that holds at volume has four gates, and they run in this order for a reason.
Gate one — provenance, before creative review. Which tools produced this, on what licence tier, from what inputs. Anything failing here stops immediately, because reviewing the aesthetics of an asset you cannot legally run is wasted time. This gate is a checklist, not a judgment, and it can be delegated.
Gate two — accuracy. Does the asset assert anything about the product, price, availability or performance that is not true? Generative tooling produces confident, plausible, wrong detail, and it does so in copy far more often than in image. This gate needs someone with product knowledge, not design knowledge — which is usually a different person than the one reviewing creative, and is the gate most often skipped.
Gate three — brand and craft. The traditional review. Is it good, is it on brand, does it work in the placement.
Gate four — sampled re-check after approval. Pull a random five percent of approved assets and re-run gates one and two. This is the only thing that tells you whether the earlier gates are actually being applied or have quietly become a formality under volume pressure. In practice this is where the misses surface.
Two structural notes. Gates one and two should be checklists with named owners, because judgment does not scale and checklists do. And the review capacity has to grow with output — a team that triples production and keeps one reviewer has not tripled its throughput, it has tripled its risk.
Briefing AI creative
The brief matters more than it used to, not less, and briefs written for a three-execution world under-specify for a thirty-execution one.
Specify the invariants. What must be true in every execution — product representation, claim language, logo treatment, the thing that makes it recognisably yours. At volume, drift is the default, and undefined elements drift first.
Specify the variables deliberately. What you actually want explored: setting, framing, message emphasis, demographic representation. Undirected variation produces thirty versions of the same idea with different backgrounds, which is volume without learning.
State the claim boundaries. What can and cannot be said about the product. This is the single most useful line in an AI creative brief, because it is the thing generative tooling will invent if left open.
Name the test. What are these thirty executions supposed to teach you? If the answer is "which one performs best," that is fine — but the variants need to differ along one axis at a time, or the result teaches you nothing you can apply.
Reviewing a portfolio when anyone can generate one
Portfolios have become weak evidence. Anyone can produce a striking image; the question is whether they can produce the right one for a brief, repeatedly.
Three things to ask for instead:
The brief behind the work. What was the problem, what was the constraint, why this solution? A generated image has no answer to this. A designed one does.
The rejected directions. Strong teams have a graveyard and can explain it. It is the clearest evidence of judgment in the room.
Work in a category adjacent to yours, with the performance context. Not necessarily your sector — the reasoning transfers, the aesthetic often does not.
Two methods, two answers
There is a live and consequential dispute in AI creative, and where a partner lands on it shapes everything they produce: is AI creative best used for volume or for exploration?
Position A — volume. Once a concept is validated, generate as many executions as the channel will absorb. The argument sits on how modern ad platforms work: delivery systems optimise across a creative pool, and a larger pool of viable variants gives the algorithm more to find a winner in. On this view, the correct use is industrial — settle the concept, then mass-produce.
Position B — exploration. Use generative tooling at the front of the process, to explore more distinct territories cheaply before committing, then produce the winners properly. The argument is that volume optimises within a concept while exploration finds a better concept, and concept choice has far more leverage on results than execution count.
CONTESTED, and both are practised seriously by good teams.
Our read. They solve different problems and the mistake is running one when you need the other. If you have a proven concept and flat performance from creative fatigue, volume is correct — you need fresh variants, fast, and exploration wastes the window. If your performance problem is that nothing has worked, volume amplifies a failing hypothesis and exploration is the only route out. Diagnose which problem you have before choosing a production model, because the two look identical on a media report and require opposite responses.
The practical test: look at your best and worst performing creative from the last quarter. If the spread is wide, you have a concept that works and volume will help. If everything clusters at the same mediocre level, you have a concept problem and no amount of variants will fix it.
What would change our position: platform-published data on how creative pool size interacts with concept diversity in delivery. The platforms hold it. None of them have released it.
What to do in the next ninety days
First thirty days. Settle the rights and disclosure standard in writing, with whoever produces your creative. This is the cheapest risk reduction available and it is almost always undocumented.
Days thirty to ninety. Diagnose volume versus exploration using the spread test above, then run one production cycle against that diagnosis with a defined win rate before scaling the model.
Beyond ninety days. Build the review capability internally. The provenance and accuracy checks are becoming a standing function, not a project, and they should not live entirely outside the building.
Where this goes next
Two shifts worth planning for.
Distinctiveness becomes the scarce input. When everyone has the same tooling trained on overlapping data, output converges. The brands that stand out over the next two years will be the ones that deliberately spent on a point of view — art direction, casting, photography, a visual system with rules — while competitors optimised for throughput. That is a strategic choice available now and harder to retrofit later.
Disclosure expectations will tighten. Platform policies and regulator attention are both moving in one direction, and the direction is more disclosure, not less. Agencies and brands that built clearance and disclosure into their process early will absorb that quietly. Those that did not will be retrofitting it across a back catalogue.
If you are reviewing creative partners, ask for the provenance answers in writing. We will put ours in writing first.
If you want the spread test run against your own last quarter — to establish whether you have a volume problem or a concept problem before you buy either — that is where an AI creative engagement begins. WKND has been producing and running performance creative since 2018 across more than 400 brands, with 20+ years of collective DTC and B2B experience behind the work.
