Reflections on Building an AI Ad-Generation System

Naive prompt loop for generating a publishable banner. Image by author A good-looking ad is easy to generate. A publishable one is a different challenge. A few months ago, I left the marketing consulting firm where I had led and developed an AI-powered ad-generation system. I was there for only a short time, but the project became one of the most formative experiences of my professional career. I worked closely with product, marketing, creative, and client teams to validate its outputs and integrate it into their workflows. The system eventually contributed to work that won two Effie Awards . Yet the awards are not the most important part of the story. The real challenge was turning an impressive AI demo into a system people could actually trust: one that respected brand rules, legal constraints, platform requirements, and cultural trends, while fitting into the workflows of the people responsible for publishing the final result. Referential banners generated by the system. Text, logos, and legal elements were removed for publication. Image by author. I cannot share the project’s technical details. But I can share what the experience taught me about building AI products, where “almost correct” still means not publishable. A short story about the project ( Trend2Banner system) The project started with a time-sensitive opportunity: turning emerging trends into advertising campaigns while they were still relevant. The system researched each trend, evaluated its relevance and brand safety, discarded topics that were inappropriate for the business, developed an advertising concept, and generated banners adapted to different platforms and formats. Generating an attractive image was relatively easy. Generating something the client could actually publish was a different problem. What made a banner publishable? A visually appealing banner was not necessarily publishable. For this project, publishable meant ready for production and compliant with the following business requirements: Brand alignment: Reflect the brand’s identity, values, and messaging. Brand guidelines: Follow its visual system, including colors, typography, logos, and composition. Brand safety: Avoid sensitive, inappropriate, or controversial topics, even when they were highly relevant online. Competitive boundaries: Avoid concepts, phrases, and references associated with competing brands. Platform requirements: Be adapted to the specifications of each advertising platform and format. Text accuracy: Include clear, grammatically correct copy without spelling errors. Legal compliance: Display every required disclaimer accurately and in the correct position. Timing: Be generated while the trend was still relevant. Some trends were popular and culturally relevant but still unsuitable for the brand. Political impersonations, sensitive news, crime-related topics, and references associated with competitors had to be detected and rejected before any creative asset was generated. A reliable ad-generation system must know not only what to create, but also when not to create anything. These requirements transformed image generation into a broader engineering problem. Creativity could be probabilistic, but legal text, spelling, dimensions, and positioning could not. How extensible is this project for other brands? How much of the system is reusable? The core generation engine was largely brand-agnostic, but the complete system was not plug-and-play. Adapting it to another brand would require updating its business context, visual guidelines, legal constraints, platform requirements, and relevance criteria. The reusable part was the underlying workflow: detecting a trend, evaluating it, generating a concept, producing the creative assets, and validating the result. The brand-specific knowledge and rules would still need to be configured and tested for every new implementation. The results (The numbers at the “end” of the project) During the first two-week pilot, banners generated by the system and published through Meta Ads drove more than 20,000 recharges totaling S/ 726,000, helping the company capture incremental share of wallet from competing alternatives. Over the following three months: More than 150 banners were published . Around 20 banners could be generated per day . Each banner required approximately four model API calls . Generation cost remained below US$1 per banner . The system used more than 6,000 words of business context to guide generation and validation. Some sample results. Text, logos, and legal elements were removed for publication. Image by author These numbers showed that the system could create business value, but they do not explain what made it difficult to build. The hardest lessons came from the decisions behind those results: what to validate first, where to use generative models, how to measure reliability, and how to integrate the system into the client’s existing processes. Those decisions changed how I think about building AI products. The following reflections are the lessons I would carry into my next project. What building this system taught me R1. Solve the generation problem before committing to the product Before telling my project manager that we could build the product, I asked a more fundamental question: How can we generate publishable ads with minimal human intervention while meeting every brand, legal, and platform requirement? Problem: Generate timely, publishable ads that comply with brand guidelines, legal requirements, and platform specifications. Product: A platform where users can review, approve, and publish generated ads with a single click. Both the problem and the product matter, but AI-native products require a particular balance. When the entire experience depends on autonomous generation, it is important to validate the core generation problem before committing to a specific product. Otherwise, we risk designing workflows around an engine that cannot yet produce reliable results. Once the image-generation engine was partially validated, building the product became a different challenge. It required talking to users, understanding their existing processes, defining the right workflows, and continuously implementing and iterating. The product was not simply an interface around the model. It was a new operational layer built on top of an already tested, but still evolving, generation engine. Problem and product relationship. Image by author Partially validating the generation engine showed that the idea was technically viable. It did not show how to make its outputs consistently reliable. That required much more than improving the prompts. R2. Prompting is only one layer of the system When people think about AI engineering, they often think about prompts, API calls, temperature settings, and chatbots. Prompting matters, but it is only one layer of the system. A good prompt can guide the model toward the desired output, but it cannot guarantee reliability. The harder work is building the harness around the model: verifiers, guardrails, feedback loops, evaluations, and recovery paths. Together, these mechanisms make outputs reliable enough for production. Building these mechanisms had another advantage: they reduced our dependence on any particular model. The generation engine could evolve without rebuilding the entire product around it. Terms related to the project. Image by author R3. Build systems that improve as new models come out I first heard this idea on the AI Engineer channel , and it changed how I think about AI products. The model is only one part of the system. Reliable evaluations, memory, guardrails, and observability are what turn its probabilistic outputs into a dependable product. Models will improve, and benchmarks will continue to be beaten, but the surrounding system will remain. When a new model or prompt is introduced, that system should tell you whether the product actually improved, not just whether its outputs look better. But technical reliability was only part of the challenge. Even a well-evaluated system creates no value if people do not trust it or cannot integrate it into their daily work. R4. AI adoption is more about culture and processes than tools The technology may already exist, but the culture and processes required to adopt it often do not. People were not necessarily afraid of AI. They were skeptical about its reliability and how it would fit into their existing workflows. Building trust required clear responsibilities, human oversight, and evidence of real value. This project did not replace the existing campaign processes. It enabled a new category of trend-driven campaigns with less effort, complementing the team’s work and generating incremental revenue for the company. Closing Thoughts The hardest part of this project was not generating an attractive image. It was building a system that could produce something timely, compliant, measurable, and useful within a real marketing workflow. Turning a compelling demo into a production-ready system required two things: trust from the people using it and a reliable harness around the model. The harness provided verification and consistency, while the product and its workflows made those capabilities useful to the business. The two Effie Awards and the campaign results were meaningful outcomes, but the most valuable result for me was a different understanding of AI engineering. A successful AI product is not defined by what a model can generate in a demo. It is defined by how reliably the complete system solves a real problem for its users. Models will continue to improve. The harder and more durable work is building the product, processes, and feedback loops that allow those improvements to create real value. Found this helpful? Share your thoughts or questions in the comments. You can also connect with me on LinkedIn . If you are interested in another practical AI use case, read my previous article on structured video analysis with Gemini . Reflections on Building an AI Ad-Generation System was originally published in Towards AI on Medium, where people are continuing the conversation by highlighting and responding to this story.