OpenAI published the report “Enterprise Signals” and the working paper “How Organizations Use AI: Evidence from ChatGPT,” based on telemetry from more than 10 million messages. The data records a structural shift: enterprise clients are moving AI from the role of an assistant to a mode of autonomous task execution through Codex, plugins, and skills. By June 2026, Codex generated 64% of all output tokens among enterprise clients.

image
image

What happened

OpenAI published the article “From assistance to execution: How enterprises put AI to work” along with two documents — “Enterprise Signals” and the working paper “How Organizations Use AI: Evidence from ChatGPT.” The analysis is based on telemetry from more than 10 million messages. By June 2026, Codex generated 64% of all output tokens among OpenAI enterprise clients (by the sum of Codex and ChatGPT). Weekly active users have grown since February: legal department — 108 times, sales — 41 times, recruiting — 41 times, marketing — 26 times, engineering — 5 times. Frontier companies (top 10% by tokens per active user) generate 8.3 times more tokens per user than typical companies (45th-55th percentile), compared to 2.6 times in January. Frontier companies use advanced features more actively: plugins — 21% versus 9% for the rest, skills — 19% versus 3%.

Context

The AI industry has observed a gradual transition from chat interfaces to agent architectures during 2025-2026, and the OpenAI report became the first public confirmation of this trend at the level of a major provider. The shift to agentic execution means that AI gains access to tools, external data, and permissions, rather than simply answering questions in a dialogue. A similar structural shift occurred when the industry transitioned from on-premise to cloud infrastructure in the 2010s — companies that moved to the cloud earlier gained a significant competitive advantage.

Why this matters for the industry

The report provides a signal for product strategy: the focus is shifting to agentic workflows, plugins, and skills, especially for non-technical departments — legal, commercial, and marketing. The need for eval methodologies for multi-step agent tasks (rather than single-turn QA) is becoming critical, as current benchmarks MMLU and GSM8K do not measure real enterprise workloads. The change in load profile requires a review of capacity planning: inference costs are higher, context windows are larger, and latency requirements are different. Infrastructure is adapting to specialized serving stacks for agent inference and the market for agent observability solutions.

Why this matters for users

If you work in a company implementing AI, the report provides specific benchmarks: the greatest growth in adoption is not in engineering (5x), but in legal (108x), sales (41x), and recruiting (41x). The 8.3x gap between frontier and typical companies shows that the use of plugins and skills is a key factor in lagging or leading. However, it is worth noting: the data was obtained exclusively from OpenAI product telemetry, and there is no independent verification of the effectiveness of these practices.

What is still unknown / limitations

OpenAI itself acknowledges that tokens are a proxy for depth of use, not a direct measure of business value. The 8.3x gap between frontier and typical companies does not prove a causal relationship with business outcomes: more tokens may mean more tasks, more iterations, or more hallucinations and repeated requests. The data covers only users of OpenAI products and is not representative of the entire industry. The report does not contain data on generation quality, error rate, or task completion rate.

Sources

Author

Look at AI, editorial team