BackPortfolio / AI integration
I rebuilt the process so the first pass of concept development goes to AI, and human work starts where it's actually needed.
When I joined the organisation, content production looked like a closed loop: task, brainstorm of ideas and formats, prototype, edits, a change of concept, prototype again, edits again. Every cycle ate up time and resources, and the result still needed rework.
I proposed a different setup: the first pass of concept development goes to AI, and human work begins where it's genuinely needed.
At first I took on the role of translator between the marketers and the AI. We discussed the goals and tasks the final product had to serve, then I wrote the brief for the model. That kept us from burning hours on empty discussions and showed the direction right away.
The model offered a few concepts in broad strokes. I brought them to a working state by hand and started production — instead of guessing at options blind.
Stock photography was a story of its own. Digging through thousands of unsuitable images took no less time than generating them. I moved the process onto a different plane: the right image is easier to create than to find.
Instead of vague phrases like “make it more alive, prettier, more realistic” and a full redo of the image, I gave targeted notes that set a clear direction for generation.
One example — people at work. Standard generation produced people who were too smooth and plastic, people you don't believe in. I filled those in with accent details from the real world.
Gradually a pool of base additions for prompts took shape. It made it quick to generate images with people where the mood and tone of the scene came through clearly.
AI results without a human who has an eye and a clear vision of the outcome give you AI slop right away. You can see it at a glance — this is what the attempts looked like before the process was tuned.
This is the result after the prompts were tuned and the best picks were made: real people in a working environment, with the mood and tone of the scene coming through.
Besides the photos — a set of illustrations in one style: for banners, presentations and diagrams.
Instead of a hundred images that had to be reworked, burning tokens and grinding over them for three days, we reached a strong result in a day with two or three iterations and a minimal number of generations.
AI takes on the volume and the rough work, but without a human who has an eye and a vision of the outcome it produces AI slop. The pairing is what works: the model suggests — the designer selects and refines it with targeted edits.
