How the Top 5% of Companies Are Building Their AI Edge

OpenAI unveiled B2B Signals, stating that enterprise AI advantage has begun to compound. This metric is an enterprise-focused extension of OpenAI Signals, tracking the depth of AI usage within companies and its spread across industries and functions, based on de-identified, aggregated enterprise usage data.

The core signal is depth. The 95th percentile, or top 5% of companies, demanded 3.5x more work from AI per employee than typical companies, widening from 2x in April 2025. Message volume alone explained only 36% of the gap, with the rest coming from usage patterns that delegate more complex tasks and provide richer context. Token generation was used as a proxy for the actual volume of work AI performed.

The gap was especially pronounced in delegation-style tools like Codex. Top 5% companies sent 16x more Codex messages per employee than typical companies, and a similar pattern appeared in ChatGPT Agent, Apps in ChatGPT, Deep Research, and GPTs. Companies are now moving beyond chat-based assistance toward delegating long-running tasks to AI across files and codebases.

AI is expanding from internal API usage into production workflows spanning coding tools and customer support. By function, IT and security focused on procedural guidance, software development and data science focused on coding, and finance focused on analysis and calculations.

• Cisco cut build times by about 20% using Codex, saved over 1,500 hours of engineering time per month, and increased defect resolution throughput by 10-15x.
• Travelers is automating initial accident reporting, insurance inquiries, information gathering, and internal system registration with its OpenAI-based AI Claim Assistant, planning to process about 100,000 cases in its first year.

The gap between leaders was greatest in the area of education and learning. OpenAI proposed measuring depth, building production-grade governance, treating adoption support as core infrastructure, scaling leading teams, and shifting from chat-centric use to AI-delegated tasks.