AI engineering productivity gains are not uniform; they cluster into three distinct tiers defined by organizational discipline, not model capability.

The first tier—distributing an AI IDE without process changes—yields a mean improvement of 20-46%.

Faros telemetry across 22,000 developers shows 66% faster epic completion but a 54% increase in bugs per developer. Google’s RCT and GitHub data put the number at 21-24%.

The frontier tier (Replit, NVIDIA, Amplitude, Anthropic) achieves 2.

5-3x gains by building an operating layer that orchestrates agents across tools like GitHub, Linear, and Slack, reducing human PR review time by 30% and complex support handling by 60%.

Replit’s internal agent outperformed a seven-figure SaaS tool at one-tenth the cost. The factory tier reaches 8x+ where agents operate as first-class organizational units.

Nubank achieved 8x engineering efficiency and 20x cost reduction using Devin for large-scale refactoring.

Goldman Sachs is piloting Devin alongside 12,000 developers and estimates agentic AI could deliver 3-4x the rate of prior tools. The gap is the operating discipline, not the model.