Meera was buried in the marketing backlog at her Series B fintech. The CEO mandated 200 personalized landing-page variants — text, hero images, short video explainers. Her team of three had zero designers and a creative agency quote that would eat half the quarter's budget.
She spun up Midjourney for visuals, Claude for copy, Runway for clips. Forty-eight hours later she had the assets. They looked good. Until legal walked in.
“These images? Training data scraped from Shutterstock without licenses. And the video voiceover hallucinated a feature we don't even ship.” Meera's “2x faster” win became a three-week legal scrub. Output doubled. Trust dropped.
Generative AI solves real problems — speed, personalization, ideation. But it hits the AI PM Trinity harder than anything else: messy training data (IP landmines), non-deterministic models (brand tone drift), and UX that swings between delight and disaster.