Creating with AI: a method for retaining control of advertising
A campaign supported by artificial intelligence is a campaign: a promise to be understood, creative choices and evidence to be controlled. The gain is not measured by the number of proposals produced, but by the time it takes to achieve a fair, distinctive and diffused deliverable.
Article prepared with artificial intelligence assistance. The sources and method are indicated in the text.
Organize a pilot on an authorized real brief, measure full time and define validations before generation.
1. Write a brief that can be evaluated
Before opening a tool, specify the target, audience, value proposition, channel and expected result. Separate verified information from the assumptions to be tested. Add the mandatory elements: product name, exact offer, dates, formats and brand constraints. A vague statement produces above all an abundance of proposals difficult to compare.
Example of a useful criterion: a person who discovers the poster must understand the offer and its deadline. "Doing modern" does not, on its own, validate a creation.
2. Define what can be entrusted to the tool
Work on a copy of the brief. Remove confidential information that is not required for the task. Check the conditions applicable to the data and content used. For a first try, limit the scope: offer angles, condense a message, build a storyboard or compare page structures.
The NIST's profile on artificial generative intelligence provides a framework for risk management. Its interest in a communication team is to recall that the evaluation is about the whole process, not just the apparent quality of a result.
3. Produce few leads, but be able to explain them
Ask for three directions that meet different intentions. For each one, record the idea, the evidence put forward, the tone and the main limit. A spectacular visual that does not match the brand promise does not become a good campaign because it is quick to produce.
4. Organize validation in several ways
- Facts and offers: check each number, date, name and condition.
- Creation: examine visual details, integrated texts, hierarchy and readability.
- Make: control tone, distinctive signs and consistency with other media.
- Public: seek out cultural, linguistic or usage ambiguities.
- Rights and Dissemination: have the sensitive points checked by the competent persons.
5. Measure a result, not a volume of variants
Compare creations on a specific hypothesis: understanding the offer, attention, qualified visit or conversion. Keep the released version, sources, validations and the context of the test. Also measure the total time, including alterations. Produce 20 unused variants is not necessarily an increase in productivity.
To be noted: A useful process makes decisions clearer. Artificial intelligence offers and assists; the team sets the objective, verifies and assumes what is published.
The production file to be submitted to the customer
A useful file keeps the validated brief, the authorized sources, the selected versions and the elements still to be confirmed. It specifies the significant changes and the person who approved the deliverable. This follow-up does not need to be cumbersome: a stable nomenclature and a decision sheet may be sufficient for a limited project. It should allow to repeat a version without reconstructing all the history from the exchanged messages.
Practical case: one offer in two languages
A fictitious brand wants to announce a promotion in French and Arabic on poster, banner and short video. The brief sets a common offer, deadline and hierarchy of reading. Artificial intelligence can offer leads, but the team checks each adaptation in its final format. A version is not validated because its source text was: cropping, length and reading order can change understanding. This case illustrates a process, without assigning any result to a real campaign.
The four measurements of a pilot
Increase the preparation time, production time, retouching and validation time. Add the number of actually usable versions and the causes of rejection. Comparison with a usual method should be on an equivalent deliverable. A fast generation can mask a high correction work; conversely, a slower test can create a reusable method. The agency must explain this arbitration rather than promise a percentage of productivity without protocol.
The role of the creative director
Creative management chooses the track that best expresses the promise, not the one that seems the most technologically demonstrative. It must be able to defend the idea outside the tool used. This requirement protects the coherence of the project when the format changes, a result has to be taken over or a provider is replaced.
Source and method
NIST, Generative Artificial Intelligence Profile, AI 600-1, 26 July 2024The five-stage process is a methodological proposal of the Le Fils de Pub, not a formal NIST procedure.
Continue with a practical method
Suggested routes from the story to briefing, production or measurement. These features are separate from the original publication.