AI creatives for ads: what that actually means

Four things that actually get made

Still images. Your product or service in scenes that would otherwise need a shoot: a different room, a different time of day, a different person in front of it. This is where AI is furthest along and least noticeable.

Video. Edited spots from existing material, plus movement that was not in the source. No shoot day, no crew, no scheduling four people.

Avatars. A person delivering your script, with no camera and no studio. Useful above all because the same message can come from five different characters without anyone facing a camera five times.

Voices. The same message in different tones, in several languages if needed. For a campaign in the German-speaking market that means the same ad once matter-of-fact and once casual, without a second recording session.

Which format tolerates how much AI

Not every ad format benefits equally. The overview below sums up where AI reliably helps in production today, where people remain indispensable and which risk is typical for each. It is based on our work with campaigns on Meta and Google, not on a study.

Why the argument is volume

The most common misunderstanding is that AI creatives are a quality promise. They are not. A good photographer makes a better picture than any model, and a good editor a better cut.

The point is a different one. Since the Andromeda change on Meta the motif decides who sees an ad, and the system needs many genuinely different motifs to do that. A shoot day delivers one. So the question is no longer who makes the prettiest motif, it is who can deliver twenty different ones before the old ones tire. What Andromeda changed exactly is here.

The key word is different. Twenty variants with a different background colour are roughly one motif to the system. Real differences come from another angle, another person, another situation or another format.

Put differently: AI does not move the ceiling on quality, it moves the floor on the cost per attempt. And when an attempt costs almost nothing, you can afford attempts that used to be uneconomical.

What the AI does not do

It does not know your offer. What your product means to a purchaser in mechanical engineering, which objection always comes up in the sales call, why customers walk away: none of that is in any model, it comes from your conversations.

Nor does it make the selection. Out of twenty generated motifs, five are regularly unusable, three are wrong in tone and two are factually off. Sorting those out is manual work, and it is the part that decides whether AI production is serious or not.

And it does not replace strategy. Which angle gets tested, in what order and against what benchmark, is a decision. A model that delivers twenty variants of the same thought does not help with that.

Where it works and where it does not

It works where the motif carries a mood rather than a specific object: situations, environments, people at work. Abstract subjects, for which a real photo never existed anyway, are a grateful case too.

It works badly on your specific product when it has details that need to be right. Lettering, controls, dimensions, packaging. Here a real shot is faster and cheaper than a third attempt at persuading a model to be accurate.

And it works badly for anything that draws its trust from being real. A testimonial with an invented face is not a testimonial, it is a risk.

How to spot bad AI ads

Sometimes you can tell. And then the motif does not go live. That is not a platitude, it is the reason the selection stays with people: anything that stands out costs more trust than it gains in reach. The typical warning signs:

• Hands and fingers: too many, too few, fused together, bent unnaturally.
• Text inside the image: letter soup on signs, labels and screens, misspelled brand names.
• Faces: skin that is too smooth, an empty gaze, teeth or eyes that are slightly off. In video, lips that lag behind the audio.
• Reflections, shadows, logos: light from two directions at once, mirror images that do not match, warped company marks.
• The generic look: glossy surfaces, perfect people in perfect offices, colours that all seem to come from the same tool. Not wrong, but it looks like your competitor's ad.

A few rules prevent most of it. Real photos as the starting point instead of text prompts alone. Text, logo and price added afterwards in the layout, never drawn by the AI. A short checklist for hands, text, reflections and logos that every motif passes before approval. And deliberately allowing some imperfection: a real office, a phone photo, a coffee cup that does not belong in the frame.

How a motif gets made here

The process from brief to live ad has six steps. Only one of them is the actual generation.

1. Brief and angle. The angle comes first, not the picture. We settle which claim gets tested: the price, the time saved, the objection from the sales call. Otherwise you end up with twenty pretty motifs making the same claim.
2. Set the brand frame. Reference images, colours, fonts, logo rules, tone and the things that must never appear. This frame goes into every generation and makes sure motif twelve still looks like the same brand as motif one.
3. Produce wide. Several versions per angle in different formats, because a motif in portrait works differently from a square one and both enter the shortlist separately.
4. Cull strictly. Anything off in the details, anything wrong in tone for you, anything that looks conspicuously machine-made does not run. What remains is less than what was produced, and that is the point of the exercise.
5. Check rights and labelling. Are real people, voices, third-party brands or music involved? Are there consents and licences for them? Does the motif need to be declared as AI content in the ad account?
6. Your approval. No motif goes live without your okay. You know your customers and catch mistakes we miss from the outside, such as a machine that was never built that way.

A style guide for AI does not need to be long. One page with ten reference images, three colours, one typeface and a list of taboos is enough to prevent most outliers.

Rights and risks

The legal side of AI creatives is not a footnote. The points below are not legal advice, they are the questions we ask before every campaign. A binding assessment of your case belongs with a lawyer.

• Faces and voices of real people. Under Section 22 of the German Art Copyright Act (KUG), portraits may in principle only be distributed with the consent of the person shown. German case law also protects the voice as part of the general right of personality. An avatar resembling an employee, a customer or a well-known person therefore needs explicit consent, ideally in writing, that covers advertising use.
• Rights to the output. German copyright law only protects personal intellectual creations. An image generated without meaningful human creative input is therefore, in the prevailing view, not protected as a work. What you may do with it is governed mainly by the tool's terms of use.
• Training data. Whether and how models were allowed to train on protected works is currently before courts in several countries, and the question is not settled. What matters in practice for you as an advertiser is whether an output recognisably copies someone else's work, character or signature style. If it does, discard it.
• Third-party brands and stock. AI readily produces logos, product designs or packaging that resemble real brands. Those do not belong in an ad. Stock images and fonts you use as input need a licence that allows advertising and modification.
• Music. A song from your own playlist is not ad music. Use the platforms' licensed libraries or music with an explicit advertising licence. AI-generated music raises the same questions about terms of use and resemblance to known tracks.

Labelling: what the platforms and the EU require

Article 50 of the EU AI Act has applied since 2 August 2026. Among other things, it requires anyone using an AI system professionally to disclose deep fakes: AI-generated or manipulated images, audio or video that resemble real people, places or events and would falsely appear authentic. An avatar posing as a real customer is exactly that case.

Separately, the platforms label on their own. Meta adds an "AI info" label to ads created or significantly edited with generative AI, either under "About this ad" or next to "Sponsored", and also picks up signals from third-party tools. In July 2026 Google introduced a "How this ad was made" section: with Google's own AI tools the disclosure is set automatically, and for material made elsewhere there is a control you set yourself.

What exactly must be labelled, where the exceptions are and what that means for your ads is covered in detail in our piece Labelling AI content.

What it does not make cheaper

A misunderstanding we often hear in a first call: if production gets cheaper, the whole campaign should get cheaper. It does not.

Compared with a classic shoot, planning, location, crew and scheduling fall away, and new variants take hours instead of weeks. In return, steps are added that barely existed before: maintaining the brand frame, sorting rejects, checking rights and labelling. For a single hero motif a good shoot is often still the better choice. For ongoing tests with many variants, AI is.

So the cost block only shifts. What is saved in production moves into selection, setup and evaluation of the tests, and the biggest item remains the ad budget itself. We work from 1,500 euros of monthly ad spend, and AI does not change that number.

What changes is how many attempts you get for the same money. And since most attempts in advertising lose, that number is exactly what decides whether you find a winner.

What is different in B2B

Nobody buys on impulse in B2B. A managing director or purchaser checks whether a serious company stands behind the ad, and every detail that looks artificial counts against it. So the bar is higher here than in e-commerce.

• Real faces from the team. The managing director, the project lead, the contact person in sales. A short phone video of a real person often beats any perfect avatar in B2B. AI can cut that footage, subtitle it and turn it into variants, but it should not replace the person.
• Proof over imagery. Numbers, references, certificates, concrete results. They must be real and verifiable. AI helps present them well, it must not invent any.
• Technically correct details. A machine with the wrong construction or a dashboard showing nonsense values is spotted by experts immediately. Here we work from your real product photos and screenshots.

From more than 2,000 meetings booked since 2023 we have taken one thing above all: in B2B trust decides, and trust comes from being real. AI is the tool that turns that realness into enough variants. How we put this into campaigns is on our page for performance creatives for Meta and Google.

What you need to supply

Less than most people think. Product shots as they are. A few phone videos from everyday work, untidy is fine. Your website. And, this is the most important part, a conversation about what your customers actually ask before they buy.

Everything else grows out of that. What cannot grow out of it is a claim you never told us. If you want to see what that would look like for your offer, get in touch.

Frequently asked questions

Sources

1. EUR-Lex: Regulation (EU) 2024/1689 (AI Act), Art. 50 and Art. 113
2. Meta: How AI-generated images in ads are identified and labeled
3. Google: New AI labels for ads (July 2026)
4. Google My Ad Center: AI transparency in ads
5. Gesetze im Internet: Section 22 KunstUrhG (right to one's own image)
6. Gesetze im Internet: Section 2 UrhG (protected works)