Review and approval · rules checked September 2026
How to review and approve AI creative from an agency
Review AI creative in two passes. First check the facts against the real product (label text, logo, colors, parts, hands and the approved model's face), then judge brand fit and how each placement crops. Give every file one verdict, approve, change or reject, with a reason specific enough that the next render can act on it.
100 Creatives makes its images in Overs, an AI photo tool built by our founder, Abhi Chawla. Both are brands of 2587689 Alberta Ltd. Where this page describes a review feature, it is an Overs feature.
The short version
Ask for three things with every batch: the clean version and the version with type, every placement size, and the references and model behind each photo. Check facts at 100% zoom beside the product, then check the brand and the crops, and write notes that name the defect and the fix. In our own hand-built campaigns about 1 render in 8 became a keeper, so a batch from an agency should already be the selected few.
What to ask for with every batch
- The clean version and the version with type. Type set in a separate pass can be fixed without touching the photo. Overs keeps the clean version when it sets a headline, subline and button.
- Every placement size. Judge each crop on its own. Meta's feed placements call for 4:5 at 1440 x 1800, and Stories and Reels for 9:16 at 1440 x 2560 (Meta Ads Guide, checked September 2026).
- The references and model behind each photo. When a label comes back wrong, the first question is which reference the model saw. Overs exports each photo with a record of its prompt, model and reference pictures.
- The approved character sheet for any photo with a person, so you can check faces against it.
Check the facts against the product
| What to check | How to check it | Why it goes wrong |
|---|---|---|
| Label and package text | Read every word beside the real pack | Every render redraws the pack, so letters can shift or change |
| Logo | Compare shape, spacing and position with the reference | It is redrawn in every render too |
| Colors | Eyedropper the product and background against your hex codes | A hex code in the prompt is a target the model aims for, so measure it |
| Parts and details | Count buttons, pockets, closures and seams | A render can add details the product doesn't have. Our trousers prompts spelled out "a single dark button at the waist closure, no rivets" and "nothing added and nothing removed" |
| Hands and bodies | Count fingers; check grips and proportions | Our GPT Image prompts asked for "natural believable hands and body proportions" in so many words |
| The person | Compare face, hair and build with the approved character sheet | Faces can drift between renders, which is why the approved sheet goes in as the first reference |
| What the image claims | Ask what a viewer would believe the product does | A picture can show a result the product can't deliver |
Start with the label, because it is the detail shoppers compare with the box in their hands.
Check the brand and every crop
- Palette. Do the background and props sit inside your hex colors?
- Light. Does the light come from the same side, with the same hardness, as your approved photos?
- Lens and grain. Does the photo match your focal length and grain? On our brownie job, nine of the ten kept prompts carried the same hand-added sentence: "Heavy visible film grain throughout."
- The crop. Does the product survive the 9:16 and 4:5 cuts, with room left for the headline?
- The type. Can you read the headline on a phone at arm's length?
Write notes a model can act on
Every note ends up as words in the next prompt, so write the defect and the fix. Give one verdict per file (approve, change or reject) and a reason for every change or reject.
| Vague note | Note the next render can act on |
|---|---|
| "Make it pop" | "Raise contrast on the product only. Keep the background at #F2F5F8." |
| "The label looks off" | "The label spelling doesn't match reference 2. Match its wording and layout exactly." |
| "Wrong vibe" | "Light should come from the upper left at 45 degrees, as in the approved hero. This one is lit from the right." |
| "The model looks different" | "The jaw and hairline drift from the approved character sheet. Send the sheet as the first reference." |
| "Too busy" | "Remove the second plant and the book. Keep one prop and the contact shadow." |
OpenAI's prompting guide for its GPT Image models (April 2026) recommends the same habit for edits: say what to change, say to keep everything else the same, and repeat the list of what to keep on every round to stop drift.
In Overs, approvals and rejections with reasons become brand memory for later photos, and a note that fixed an approved photo is remembered. With any agency, the reason is what the person writing the next prompt needs from you.
Rounds, sign-off and records
- One owner per verdict. Name who approves brand fit and, in regulated categories, who approves claims.
- Keep the approved export beside its prompt and references, instead of a screenshot.
- Record who decided. In a shared Overs workspace, each approve, change or reject records who made it, and comments sit beside the photo, so a creative director's no and a client's no stay separate.
- Know the turnaround. At 100 Creatives, revisions come back within 48 hours.
The UK's IPA and ISBA published twelve principles for generative AI in advertising in November 2023, and they ask advertisers and agencies to be open with each other about their use of AI (IPA). Ask your agency which of your images are AI-made, and get the answer in writing.
Does the image need an AI label?
Check this before sign-off, because the rules differ by platform and place. All of these were checked in September 2026:
- Meta requires advertisers to disclose AI-made or altered photorealistic media only in ads about social issues, elections or politics. For other ads, Meta adds an "AI info" label itself when its own tools created or significantly edited the media, or when it detects signals such as C2PA metadata.
- New York has required anyone producing a commercial ad to conspicuously disclose a synthetic performer they know is in it since June 9, 2026. Fines are $1,000 for a first violation and $5,000 after that.
- The EU AI Act's Article 50 transparency rules apply from August 2, 2026, including the duty to disclose deep fakes.
- Amazon requires the metadata keyword "contains-synthetic-performer" on main images that show photorealistic AI-generated people (Amazon product image guide).
- The FTC's Endorsement Guides say an endorser presented as a user must have used the product, so an AI person can't pose as a customer.
Treat this as a checklist and ask counsel about your own ads. For a platform-by-platform rundown, see Overs' guide to labeling AI images in ads.
Review built into the tool
If you run images yourself, Overs builds this review into every run. It shows the shot plan and an estimated cost before anything renders, and for anything worn or held you approve the character sheet first. Each photo gets approve, changes or reject with a reason, the reasons become brand memory, and in a shared workspace every decision has a name on it. On Pro and Team, a reviewer checks each photo against the references before you see it. An agency can send its client a review file, a single HTML file, and the client's decisions come back into Overs.
It's the tool our team makes its images in today. At our average of 8 tries a keeper, a usable photo costs $0.27 to $1.69 in AI fees. Overs is our sister product, made by 100 Creatives' founder. Try Overs free at www.overs.studio. The free plan makes 40 photos a month, plus a few cents a photo in AI fees on your own OpenRouter key, with no markup from Overs.
Read next
Frequently asked
questions
How do you review AI-generated ad creative?
In two passes. First check the facts against the real product at 100% zoom: label text, logo, colors, parts, hands, and the face against the approved character sheet. Then check brand fit and each placement's crop. Give every file one verdict, approve, change or reject, with a reason the next render can act on.
What should I check first in an AI product photo?
Check the label and package text first, word by word beside the real product, because every render redraws the pack and letters can shift. Then check the logo, the colors against your hex codes, and parts such as buttons, closures and seams. Our own apparel prompts spelled out details as small as a single dark button at the waist closure.
How do I give an agency feedback on AI images?
Name the defect and the fix in words a prompt can use. "The label looks off" gives the next render nothing to act on. "The label spelling doesn't match reference 2, match its wording and layout exactly" does. OpenAI's prompting guide for its image models recommends the same habit: say what to change and list what must stay the same.
How fast should an agency turn around AI revisions?
Ask before you sign, and get the answer in the quote. At 100 Creatives, creative batches come back in 48 hours and revisions within 48 hours. Keeping type in a separate pass also helps, because a headline edit then leaves the photo underneath alone.
Do AI images in ads need a label?
It depends on the platform and the place. Meta requires disclosure only for ads about social issues, elections or politics, and labels some other AI media itself. New York has required a conspicuous disclosure for synthetic performers in ads since June 9, 2026, and the EU AI Act's transparency rules apply from August 2, 2026. Check each market you run in.
Who should approve AI creative at a brand?
One named owner for brand fit and, in regulated categories such as supplements or skincare, a second reviewer for claims. Record who approved which file. The UK's IPA and ISBA principles for generative AI in advertising also ask agencies and advertisers to be open with each other about AI use, so ask which images are AI-made.
Want batches that arrive
already checked?
100 Creatives makes static and video ads and the imagery behind them, with batches back in 48 hours and revisions within 48 hours. Every engagement starts with a written quote, from $4,000.