AI image and video ads: from brief to approval

AI can generate image and video variants quickly. Nothing is published without explicit authorisation. A controlled creative process separates the creative brief, approved source materials, generation, rights and brand review, technical adaptation, approval and delivery. Every published file is tied to a specific version and a named approval. Speed comes from clear decisions, not from skipping checks.

The creative brief must be visually verifiable

A prompt is not a creative brief. The creative brief starts with the communication goal, target audience, channel, format and the one message the creative is meant to support. Visible requirements follow: product shape, colours, setting, perspective, text in the image, logo use, depiction of people and excluded elements. For video, duration, scene sequence, speaker role, subtitles, music and end card are added.

Formulate must-have and must-not criteria separately. “Modern and trustworthy” leaves a lot of room for interpretation. “Product label unchanged, no medical setting, no real person, logo only from the approved original” is more verifiable. Add references only once their use has been clarified. An image that is publicly available is not automatically approved source material.

The creative brief also names the status of the result: DRAFT, not APPROVED. A status label on its own does not prevent accidental publication. What keeps unapproved versions away from a live channel is the combination of explicit states, separate permissions and a dedicated release check. In the recommended project workflow, the approval ID and file hash are what ensure that only a reviewed version is delivered. Draft, approval and delivery remain three different states with different permissions.

Inventory the source materials before generating

Each source asset is assigned an ID, a record of its origin and a note on permitted use. This includes product photos, logos, typefaces, music, speaker voices, stock material, customer material and style references. The entry answers: who supplied the material? Which licence or consent is in place? For which channels, countries, periods and edits may it be used? Does it contain personal data or confidential information?

The Swiss Federal Institute of Intellectual Property (IPI) explains that with the training and use of AI, various technical acts and possible rights in inputs and outputs must be kept apart. Each creative workflow therefore needs its own rights review. Neither “AI-generated” nor “human-edited” is an approval on its own. If in doubt, the reference is removed or the open question is referred to a responsible specialist.

Start tests with synthetic material and generic placeholders. Do not use customer images, faces or unpublished products merely to test how a tool works. If real ad creatives are needed later, document their data flow separately: upload, processing, storage, possible reuse, export and deletion.

Generate variants to test a clear hypothesis

Ten random images are not a test strategy. The simplest controlled comparison changes one axis per variant family: image crop, background, information density, the order of a scene or visual emphasis. All other elements stay as stable as possible. Other planned and evaluable test designs remain permissible. That keeps it clear what differs between the variants, and later results can be attributed to that difference.

Record the prompt, tool and model version, settings, input IDs and time stamps. Any later manual edit is logged as its own step. For video, editing, voice-over, subtitles, music and format conversion all belong in the version chain. The final export must not carry the same file name as an unchecked draft.

Variants should not be synchronised automatically into an advertising channel. They first go into a review queue. A technical connector may transfer only the approved version and its permitted metadata. If the approval ID is missing or the file hash does not match, the transfer stops.

Check in four review steps

The subject-matter review checks whether the product, its application, the price statement, the text and the context shown are correct. A visually convincing image can suggest a wrong product function. For video, the sequence and the spoken statements are checked as well. Unsubstantiated performance or result promises are removed.

The brand review checks logo, colours, typography, tone, visual language and excluded motifs. It looks for unintentional similarity with competitors and whether the variations still belong to the same campaign idea. Brand conformity is a human decision based on a current guideline, not a general “brand score”.

The rights and risk review checks the inventory of source materials, people, ownership, licences, depictions and required notices. Synthetic people can depict real groups in a stereotyped or misleading way. Deliberately check anatomy, text fragments, symbols, uniforms, places and contexts. For identifiable persons and AI-supported processing of personal data, the Swiss data-protection perspective must be included. This article does not replace a case-by-case review.

The target-channel review checks format, file size, aspect ratio, text limits, subtitles, landing page, labelling and the current advertising policy. Google Ads points out that generated creatives are not automatically policy-compliant and must be checked by advertisers for accuracy, misleading content, policies and law. Meta describes its own AI information labels. Such mechanisms are platform-specific and can change; the current channel check takes place immediately before approval. The four reviews are dimensions, not four people: one qualified person can cover several without any review being dropped.

Do not invent people, brands or reality

An ad creative must not fake a real customer story, testimonial or usage situation that is not substantiated. Synthetic scenes are kept as such internally and visibly labelled where needed. Photorealistic depictions of real or supposedly real people require particular caution. A resemblance can be unintentional, so the review must check the depiction against any identifiable person, not only against celebrities or other public figures.

Even smaller changes can shift the meaning: an added object can suggest a product property, a new background can assert a location, a lip movement can make a person say something. So compare the result not only with the prompt but also with the approved factual and rights basis.

With UGC-like visuals, it must remain clear whether they are genuine user-generated content. An artificially created “customer photo” must not appear as a real experience. For internal tests, clearly labelled fictional scenarios or generic placeholders are more suitable than invented social proof.

Use provenance, but do not overstretch it

C2PA Content Credentials can attach tamper-evident, signed provenance statements about origin and editing to an ad creative. That is useful context for a version chain. The specification expressly does not say whether the content is “good”, true or legally permissible. A valid provenance entry therefore replaces neither a fact check nor rights and brand approval. The linked version 2.2 is a valid versioned reference, but not the latest version published by the standards body.

Provenance data can be lost during export or on a platform. That is why you should also keep an internal origin register with file hash, sources, editing steps, approval and status. After every format conversion, check whether the metadata is preserved. Where a channel offers its own labels, these are set according to the current documentation; this still does not guarantee full compliance with all applicable obligations.

Build accessibility into the production process

An image needs context-related alternative text when it carries information. Decorative images are marked accordingly. Complex graphics also carry their key message as readable, selectable text where appropriate. Videos need accurate subtitles and, where necessary, a transcript or audio description. Automatically generated subtitles are checked by a person against the image and the sound.

Text must not exist only in pixels. Contrast, font size and safe zones are checked in the intended formats, not only in the master file. Accessibility is not a final export filter; it influences the creative brief, scene sequence and distribution of information.

Test formats and target audiences without stereotypes

A square, portrait or landscape format is not just a crop. An automatic reframe can cut a product, subtitle or person out of context. Check every file variant as its own ad creative and compare it with the approved message. For video, the beginning, the end and the transitions are checked; a single correct still frame is not enough.

Before the pilot, create a list of sensitive depictions for your brand and target audience. This can include age, disability, occupation, origin, family or financial situation. The review asks whether roles are distributed one-sidedly, people are objectified or life situations are misused as visual shorthand. This review is context-related and cannot be delegated to a blanket model score.

Deliberately test edge cases: several people, hands, text in the image, reflective products, regional symbols and translations. Record error types across variants. If a problem repeats, the creative brief, the reference or the tool is adjusted; it is not hidden by more random generations. A discarded variant remains in the review log, as far as retention and data protection permit, so that the team can learn from the error.

Separate approval and delivery technically

The approval references the creative ID, file hash, channel, country, period and the approving person. If the creative is changed in content or translated, the effect of the change is reviewed; an unchanged, technically faithful file variant can continue to rely on the unchanged factual and rights findings. An old approval does not automatically apply to a changed version. An explicit approval can permit publication by a commissioned party. The publishing or Ads access belongs to a separate role with minimal rights.

After delivery, rejections, complaints, rendering errors and necessary corrections are recorded as incidents. Performance figures do not decide rights or truth retrospectively. Whether a variant has an incremental effect is a measurement question and belongs in the analytics plan. A high click-through rate cannot legitimise a problematic claim.

This process is an editorial and operational decision aid; it does not describe a shipped cross-channel ad-creative system. Where an already approved creative is used in a Google Ads account, its use can be considered within the scope of a separate read-only audit. Production, rights review and delivery are not automatically included in the Google Ads product and require a separately agreed project.

How maitiq helps: derive variants from a clear test hypothesis

A creative brief defines the target audience, the offer, the claim that can be substantiated and the desired next step. From this, for example, three variants emerge: problem, product benefit and verifiable evidence for the stated benefit. Compare these concepts instead of changing the motif, target audience, offer and landing page at the same time. A single changed axis is the simplest controlled comparison; another test design is permissible if it is justified and can be evaluated.

The approval checks brand impact, factual accuracy and usage rights. In an agreed creative project, maitiq can connect a clear test brief with approval criteria and evaluation. Production, rights review and delivery are assigned within that project scope. Which metric is meaningful depends on the campaign goal: for demand goals, the contribution to qualified demand counts; for other goals, the metric agreed for the project. The number of images generated is in no case proof of success.

How a first pilot with maitiq begins

A focused initial assessment identifies a creative process with sufficient volume and results that can be evaluated. Within the agreed project scope, maitiq helps to define the use case, clarify the necessary data and control questions and prepare a pilot plan with criteria for a possible continuation.

Decision rule: Before generating, define the target audience, the statement, non-negotiable brand rules, exclusions, the test metric and the stop criterion.

Sources and how to interpret them

The list below brings together the sources used here. Platform documentation explains features and limits; authority sources provide the legal context. Publications by providers and associations must be classified accordingly, not as a general market price or proof of success. Information on how maitiq works is available at maitiq.com.

Have maitiq review your specific case.