Injecting Creative Knowledge into AI Ad Workflows (Part 2)

In our tests, instructions like “be more creative” or tweaking parameters such as temperature helped only marginally. So how do we move forward? The approach described here is to build a system that compensates for the model's tendency toward the obvious.

This is achieved through a workflow built around content injection.

Inject Creative Input Where AI Falls Short

The logic is simple. If an AI model consistently defaults to bland ideas at specific stages of a creative workflow, we intervene at those exact points with high-quality, human-curated material that the model is unlikely to generate on its own.

We use the AI for what it does well, analysis and assembly, while feeding it the creative components it lacks.

• When the LLM offers weak marketing angles… we inject a pre-vetted list of original, compelling concepts it can choose from.
• When the LLM proposes generic visuals… we feed it a curated library of strong ad layouts and visual styles to guide its output.
• When the LLM writes generic copy for a target audience… we provide it with a curated set of specific persuasive angles and phrasings selected for that audience.

This approach does not create “pure” machine creativity. It works around the model's default tendencies to produce more diverse ad assets at scale. The quality of the result depends on the curated material, and the output still needs review.

Build the Knowledge Base

This injection system cannot draw from a vacuum. Everything starts with building a specialized knowledge base: a curated library of strong creative components.

Mapping every ad ever made is not realistic, but mapping a selection of strong ones is. The goal is to deconstruct and catalog the elements of effective advertising. Much of this material is publicly available:

• Landmark campaigns: Analyze the strategies behind iconic ads from brands like Apple, Nike and P&G.
• Public ad libraries: Browse resources like the Facebook and TikTok ad libraries for strong video clips and banners.
• Established frameworks: Catalog the main types of marketing angle, ad copy structure and visual storytelling technique.

This raw knowledge is the fuel for your creative engine. But it is of little use without a system to refine and inject it.

How to Inject the Knowledge

Here is where the real work begins. Simply handing all this material to the model at once is unlikely to work. What matters is how, when and where it is injected.

The general method of turning business knowledge into instructions and reference material, and of testing whether a workflow applies it correctly, is covered in Bringing AdTech Expertise into AI Workflows. This article focuses on its application to creative ideation.

Data Preparation

First, the raw knowledge must be carefully processed. Every creative concept, video script and copy layout needs to be assessed, filtered, classified and enriched with detailed tags and metadata. An idea for a sports brand needs to be tagged “sports”, “inspirational”, “high-energy” and so on. This turns a chaotic pile of data into a structured, searchable library.

Use AI for Selection, Not Creation

This is the key shift in thinking. In our experience, an LLM is a much stronger analyst and selector than it is an originator of ideas. We use this strength to navigate the knowledge base.

Here is how it works in a practical workflow:

1. Initial analysis: The AI is given a simple, non-creative analysis of the brand or product. For example: “Identify the industry for this new running shoe.” The AI would typically reply “Sports & Fitness.”
2. Retrieval: The “Sports & Fitness” tag is then used as a query to filter the prepared knowledge base. The system retrieves a pre-approved list of the ad concepts tagged for that industry.
3. Constrained “creativity”: The LLM is then instructed to build an ad campaign, but its creative options are no longer its own generic defaults. It works with the curated concepts it has just received.

The same process is repeated at each step where creative input is needed, from the overall marketing angle to the shot list for a video ad: the AI analyzes the context, then uses that analysis to select from the curated library.

In short, rather than asking AI to think outside the box, we build a better box filled with strong ideas and ask the AI to pick the most relevant ones. This helps produce more varied ads; their quality still depends on the curated material and on human review.