On August 13, OpenAI published a report based on its own platform usage data showing that the gap between companies that use AI the most and those that use it the least has more than tripled in seven months — and the metric separating the two groups isn't whether a company "has AI," but whether it has moved from assistance to execution.
What changed
According to "From assistance to execution: how enterprises put AI to work," published by OpenAI itself, "frontier firms" — the top 10% of monthly AI usage among OpenAI's enterprise customers — now generate 8.3 times as many output tokens per active user as typical firms, up from 2.6 times in January. In parallel, Codex now accounts for 64% of combined Codex and ChatGPT output tokens among enterprise customers as of June, a sign that much of the usage is shifting from one-off assistance (drafting, summarizing) to delegated execution inside live systems — validating documents, routing approvals, triggering remediations, writing results straight back into records. At frontier firms, 21% of active users use Plugins weekly, compared with 9% at typical firms.
Worth noting: this is telemetry from OpenAI's own enterprise customer base, not independent market research — so it reflects the behavior of companies already on the platform, not the market as a whole. Still, the direction of the signal — the gap widening, not narrowing — is consistent with what other adoption research has shown throughout the year.
Why it matters
The most important number here isn't the "8.3x" itself, but the speed of the change: the gap more than tripled in seven months. That suggests the advantage held by companies that have already moved past the pilot stage into execution doesn't grow linearly — it compounds, because every agent operating inside a live system generates more data, more tested use cases, and more organizational trust for the next agent. Companies stuck at the assistance stage — using chat only to draft and summarize — aren't just "a bit behind." They're on a curve pulling away faster every month.
The impact for Brazil
OpenAI itself recommends a roadmap for closing that gap: measure depth of use, not just how many licenses were purchased; prioritize governance; invest in enablement; scale what already works; and move from chat-based assistance to agent-delegated work. It's essentially the inverse of a pattern we've flagged here repeatedly — including in the IDC/Cognizant research on the 88% of agent pilots that never reach production. For Brazilian leadership, the takeaway is direct: counting AI license seats as an adoption metric is measuring the wrong thing. What separates companies pulling ahead from those falling behind is whether the agent already has permission to act inside a real system, under defined governance, or whether it's still trapped inside a chat box.
Entercast's take
This report puts a number on something we've been pointing to here for weeks: the difference between pilot and scale isn't one more step on the roadmap — it's a compounding effect that punishes delay. Anyone reading this who still treats AI as an "individual productivity tool" rather than an execution layer with permissions, governance, and an audit trail inside their own systems is, according to OpenAI's own numbers, falling further behind every month that passes, not standing still.