Most of the worries brands have about AI imagery come down to one fear: something odd slipping through. A hand with six fingers. A product label that says something it shouldn’t. An image that looks like a stock photo from a slightly different universe.
Those worries are reasonable. They’re also avoidable with a process that’s fairly boring and very effective. This is the one our studio uses.
Write the brand rules down first
Brand guidelines are usually written for designers who already understand the brand. AI tools don’t. Before generating anything for a new client, we write a short, practical set of rules covering:
- The palette and the kind of light the brand lives in
- Settings that fit and settings that never would
- Who the customer is, and how they should be shown
- Things the brand never shows, says or implies
It sounds obvious, and it’s the step that prevents most problems. An image can be beautiful and still wrong for the brand, and it’s far easier to catch that against a written list than by gut feel at the end of a long day.
Never let AI draw the product
Image models invent details with complete confidence. Buttons that don’t exist, ingredients that aren’t on the label, a logo that’s nearly right. For an ad, nearly right is wrong.
Our rule is simple. The product comes from real photography or the client’s own renders. AI can build the scene around it, the bathroom, the beach, the kitchen, the lighting. A designer then composites the two and checks the product looks exactly as it does in real life.
Be careful with people
We don’t generate images of real people, and we don’t create anything that could be mistaken for a real customer giving a real testimonial. If an ad shows a person recommending a product, that person should exist.
For generated people in general scenes, designers check the usual giveaways: hands, teeth, ears, jewellery, text on clothing. They also check that the people shown reflect the client’s actual customers, not whatever the model tends to default to.
Claims need extra attention
Certain categories carry more risk. Beauty and skincare ads can’t imply results the product doesn’t deliver, which rules out a lot of generated before and after imagery. Health, finance and anything involving money or medicine has its own platform rules and, often, its own regulator.
In these categories the image goes through the same claims review as the copy. If a picture implies something the brand couldn’t say in words, it doesn’t run.
Check the platform rules each time
Meta and TikTok both have policies on AI-generated and digitally altered content, including when realistic AI imagery needs to be labelled. These rules have changed several times and will probably change again. We check the current version before each campaign rather than relying on what we read six months ago.
It’s also worth knowing that platforms review ads automatically, and AI imagery sometimes trips those systems in unexpected ways. Leaving a little time for rejections and appeals saves a stressful launch day.
Keep a record
For every generated asset we keep a note of the tool used, the source images and who approved it. Most of the time nobody looks at this again. When a client’s legal team asks how an image was made, the answer takes two minutes instead of two days.
We also only use tools whose terms allow commercial use of what they generate. It’s a dull detail and an important one.
The final check
Before any AI-assisted image goes to a client for approval, a designer checks it against a short list:
- Does the product look exactly like the real product?
- Does the image fit the written brand rules?
- Are there any visual errors in hands, faces or text?
- Does the image imply any claim the brand can’t make?
- Does it meet the current platform policy, including labelling?
Five questions, a couple of minutes per image. It’s not exciting, and it’s why we’re comfortable using AI as heavily as we do in our creative studio. The speed is only useful if you trust what comes out the other end.