AI-First Studios Hit 20x Productivity. Everyone Else Stalled.

Wharton Generative AI Labs interviewed 20 game studios, from AAA publishers to indie shops. Only three of them were getting dramatic value from AI. We're talking 4-20x productivity gains. These three shared one trait: they were built around AI from day one, not trying to bolt it onto existing structures.

The research, authored by Zimran Ahmed, lays out four stages of AI adoption. Most studios start at Stage 1: copy-and-paste. They provision something like ChatGPT Enterprise, let people experiment, and call it progress. Individuals get faster at certain tasks. The org chart stays the same.

When studios try Stage 2 (workflow pilots), they hit a wall. Multi-person workflows depend on tacit knowledge. The unwritten rules living in people's heads. A senior artist knows exactly how to prep files for the engine, but has never written it down. When someone asks them to document it for an AI pipeline, they push back. Nobody wants to codify all that. Some employees resist the process entirely. Ahmed notes this resistance is "understandable."

Stage 3 is where individual contributors start crossing boundaries. A product manager writes data queries that used to require an engineer. One PM cut query failure rates from 70-80% down to 5% over several months. Knowledge compounds in shared documents. However, even these automated workflows can face reliability issues if they are built on code that isn't designed for agents.

But the real leap happens at Stage 4, the AI-first studio. These places run with small generalist teams organized around business outcomes, not technical domains. One studio operated with five technical people total. Another had roughly 22. They feed game design documents written in markdown straight into AI production pipelines. The design docs read like detailed recipes: character specs, animation triggers, physics parameters, all structured so AI can parse them without human handoffs. A vertical slice that took four months now takes four weeks. UI icons that took weeks from contractors get done same day.

Asking "what can we automate?" only gets you so far. Ahmed's research suggests the better question is "how would we build if we designed around AI from day one?" Traditional org structures built on specialization may actually work against you. The AI-first studios bypassed the tacit knowledge problem entirely because their processes were documented from the start.