• by Laurie Sullivan , Staff Writer @lauriesullivan, March 4, 2025

Lured by the potential of generative AI (GAI), many marketers struggle to adopt the technology and get the desired results within a budget. The models need to understand a company’s products and services, and what makes them appealing to specific consumers or businesses.

GAI tools can mimic a brand's style and generate content from prompts, but often require manual refinement to become campaign-ready assets. There is a process to do this for any company interested.

Gartner analyst Andrew Frank spoke with MediaPost about how to teach and train GAI about a company’s brands. Marketing teams face challenges in integrating GAI into content, according to Gartner’s 2024 survey, which found that 77% of marketers are exploring GAI for creative development, but only 44% report significant benefits.

GAI promised to reduce campaign development times and costs, but many marketers have not found that magic process. One of the biggest hurdles is that many AI tools have been trained on databases full of copyrighted content that has not been licensed.

Frank said most organizations express reluctance to use the content without proper representation. The tools cannot really produce accurate representation, which is important in commercial representation such as products and spokespeople, so the tools are left to fill backgrounds.

AI compute costs, although less expensive than a photoshoot, can climb higher than expected. Most companies are licensing the tools on a per-seat basis and have a quota of queries or tokens they can use. Exceeding the token allocations can easily reach hundreds of thousands of dollars, he said.

Gartner has tools to calculate the cost of exceeding license limits. Some large CPGs in discussion with companies like Microsoft and Google were warned that costs could soar “sky high” if trying to generate personalized campaigns using GAI.

Dedicated media models like Adobe Firefly have solved this problem because, in this instance, the prompts were trained on its own stock photos.

While the models evolved, they are not easy to train out of the box. “You need to go down to the developer-level models that can change, train and create new models with knowledge and extra layers,” he said.

Training the AI models for a specific brand requires a vocabulary filled with words and phrases, as well as compiling sets of images that represent the phrases. Then you need to capture them, he said, so they represent different images, angles, and lighting.

Once a training package is developed a LoRA (low rank adaptation) can be used to gain better-quality characters, scenes, or concepts other than what the base model could create.

It is part of an enhanced AI workflow version, Frank said, which includes compliance pre-checks for things like brand information, perception checks with ways to simulate an audience to ensure it is creating the desired feeling, and certification checks.