Why consumer-flavored AI design tools fall over on ERP screens
Try this: open v0, Bolt, or Lovable and ask for an inventory-management screen. Not a landing page for an inventory-management SaaS — the actual screen an operations manager stares at eight hours a day. Fourteen columns. Filters for supplier, warehouse, category, stock status, low-stock threshold, last-restocked date, plus a text search. A bulk-edit affordance that respects permissions — only the "restock" and "adjust price" actions surface for users with the right role. A cell-level warning icon on any SKU where projected stockouts fall in the next 14 days. What you get back is a beautiful landing page for a hypothetical inventory-management SaaS. Or a data table with 4 columns of Lorem Ipsum. Or a "dashboard" that's actually three metric cards over a chart. What you asked for isn't in there. This is the consistent failure mode of the 2025-2026 generation of AI design and app-generation tools when applied to ERP interfaces, admin dashboards, and other complex-software surfaces. It isn't a capability gap that gets fixed by a bigger model. It's a category error about what the tools were built for. ## What the training data looks like The consumer-flavored AI design tools that broke through in 2023-2024 — v0, Bolt, Lovable, Uizard's earliest versions — were trained (implicitly, via their prompt engineering and demo examples) on the shape of the internet's public-facing pages. Marketing sites. Landing pages. Consumer apps. Every "generate a Twitter clone" example in a demo v