How to get your marketing team using AI creative tools
Start with one use case, a small pilot group and written rules, and run it for 30 days before you widen it. Name one person who approves every image, clear legal questions before anything is published, and measure renders per keeper, hours per keeper and live results against your current creative. If you work in the EU, the AI Act already requires companies that use AI systems to take steps to build their staff’s AI literacy.
A 30-day rollout plan
| Days | What happens | Who | Done when |
|---|---|---|---|
| 1 to 3 | Pick one use case and one product. Write down today’s cost and time per image for it | Head of marketing, creative lead | A one-line goal, such as "10 lifestyle ads for one product" |
| 4 to 7 | Write the AI use policy, choose the tool, set spend caps, collect reference photos | Creative lead, legal, finance | Policy signed; every key capped |
| 8 to 10 | Train the pilot group on briefs, references, review and cost tracking | Creative lead | Each person has made and reviewed a test set |
| 11 to 20 | Make the pilot set in review rounds; log renders, hours and rejections | Pilot group, named reviewer | Keepers approved and legally checked |
| 21 to 30 | Run the keepers against current creative in an A/B test of at least 7 days, then write up the numbers | Performance marketer | A decision: stop, change or widen |
Pick the first use case
A good first use case happens often, carries little risk, can be measured, and uses products you already have good photos of.
- Good first uses: lifestyle and seasonal images of products you already sell; new feed ad concepts to test against your current ads; backgrounds and scenes for social posts.
- Leave for later: main marketplace images, on-model shots where fit decides returns, images that carry health, beauty or performance claims, and anything with a real person or a lookalike.
Write an AI use policy
1. Allowed - Lifestyle, seasonal and ad images of our own products, made from our own reference photos. - Drafts and concepts for internal review. 2. Needs sign-off from [legal owner] first - Images of real, identifiable people, or lookalikes of them. - Other brands’ logos, products or packaging in the frame. - Health, beauty or performance claims shown in images. - Main marketplace images that would replace real product photos. 3. References - Only photos we own or have licensed. No images from search results, competitors or other creators as references. 4. Product accuracy - Every image is checked against the real product (label, color, shape, parts) before use. Images that change the product are rejected. 5. Disclosure - We follow each platform’s AI labeling rules and the law where we advertise. [Owner] keeps the current list. 6. Data - Business or paid accounts only. No unreleased products, customer data or confidential files in free consumer tools. 7. Records - For each published image, keep the prompt, the model, the reference photos and who approved it. 8. Approval - [Name] approves every image before it is published. [Name] reviews anything in section 2. 9. Budget - Each person uses their own API key, capped at $[amount] a month. Owner: [name]. Next review: [date, at least every 6 months].
The data rule follows from how terms differ by plan. OpenAI says data from ChatGPT Business, ChatGPT Enterprise and its API is not used to train its models by default, while on its Free, Go, Plus and Pro plans content can be used for training unless you opt out. Google’s Gemini API terms say paid-tier prompts and responses are not used to improve its products; free-tier content can be, and human reviewers may read it.
For the rules behind sections 2 and 3, see AI images of real people and using other people’s photos as AI references.
Who owns review
| Question | Who decides |
|---|---|
| Is it on brand and good enough to publish? | One named creative lead |
| Is the product shown accurately? | The product or ecommerce owner |
| Could it break a law or platform rule (claims, likeness, disclosure)? | Legal or compliance |
| Was the test fair, and did it win? | The performance marketer |
| Is spend within the cap? | The budget owner |
One person makes the final call on each question, and the decision is written down with a reason. How to review AI images as a team covers verdicts and notes.
Training that sticks
- Set expectations with real numbers. Our agency’s 2026 campaigns needed about 8 renders per usable photo: about 3 for a product alone, about 14 for one model across many scenes.
- Teach briefs and references: what each image is for, and which reference photo goes with which shot.
- Teach review with the AI image QA checklist, and how to write a note a model can act on, such as "move the light to camera left".
- Teach cost: how to read the spend page and the cap on your own key.
- Pair each person with the reviewer for their first set.
In the EU, this training is part of how you meet Article 4 of the AI Act, which has required providers and deployers of AI systems to take measures for their staff’s AI literacy since 2 February 2025. The Digital Omnibus amendments, in force since mid-July 2026, kept that duty but dropped any fixed "sufficient" level, and national authorities supervise it from August 2026. The European Commission’s Q&A (updated 27 July 2026) says no certificate is needed, and that organizations can keep an internal record of their trainings. EU AI Act rules for marketing images covers disclosure.
Common pushback, and plain answers
| What people say | An answer you can give |
|---|---|
| "It will look fake." | Some renders will. Only reviewed keepers go out, and the pilot counts how many renders it takes to get one. |
| "It will change our product." | Image models redraw the product, so every keeper is checked against the real one, and main store photos stay real photos. |
| "Is it legal?" | The policy lists what needs legal sign-off before publishing. Disclosure and likeness rules vary by country and platform. |
| "Will it replace our photographer?" | The pilot tests one use case. Shoots still cover what AI gets wrong, such as exact fit and reflective products. |
| "Our data will train their model." | Use business or paid API accounts: OpenAI and Google both say those are not used to train or improve their models by default. |
| "It costs too much to learn." | Model fees for a 20-image pilot at 8 renders each are about $11 to $34 at September 2026 prices. Most of the cost is the team’s time, which the pilot measures. |
What to measure in the pilot
- Renders per keeper, with our agency’s 8 as a reference point.
- Hours per keeper, from brief to approval.
- Model fees per keeper.
- Share of renders rejected for product accuracy.
- Live results against current creative, in a fair A/B test; see how to tell if AI creative is working.
- What the team would change in the policy before rollout.
Questions people also ask
- How many people should be in the pilot?
- Enough to cover each role in the review table and no more: a creative lead, one or two people making images, and a performance marketer. A small group keeps the rules consistent while they are new.
- Do we need a lawyer to write the AI policy?
- Have a lawyer review the parts on likeness, claims, disclosure and data for the countries where you sell. The rest is creative process that your own team can write and update.
- Should we tell customers we use AI?
- Follow each platform’s labeling rules and the law where you advertise; beyond that it is a brand decision. See do you need to label AI images in ads.
Where Overs fits
In a shared Overs workspace, the work is shared and each decision is recorded: who approved, asked for changes or rejected a photo, and why, with comments beside the photo. Each person pays for the AI on their own OpenRouter key (OpenRouter is a service that bills many AI models in one account). The Team plan is $99 a month for 5 seats.
Free for 40 photos a month. The AI that makes the photos is billed separately, on your own key, with no markup from Overs.
Sources
- European Commission: AI literacy, questions and answers (updated 27 July 2026)
- OpenAI: Enterprise privacy
- ChatGPT: Pricing (plan comparison, model training setting)
- Google AI for Developers: Gemini API additional terms of service
- Meta Business Help Center: Best practices for A/B testing
- Google AI for Developers: Gemini API pricing
- OpenAI: Image generation guide, cost calculator