OpenAI says the enterprise AI gap is organizational. Note who is holding the ruler.

OpenAI published two studies on August 12 arguing that the enterprise AI divide is increasingly set by how companies organize work around agents rather than by access to a stronger model ^[1]. Some of the numbers are worth acting on and the framing is worth discounting, because this is OpenAI's analysis of its own customer telemetry rather than an independent or cross-vendor study ^[2]. The reports present ChatGPT Work, Codex, Plugins, skills and connected apps as an operating system for enterprise agents ^[3]; runtimewire.com reads the exercise as OpenAI using its telemetry to tell executives what to buy and deployment teams what to build ^[4].

Start with the cleanest figures. Among weekly active users, 19% at what OpenAI calls frontier firms used skills, the feature that packages reusable instructions for recurring work, against 3% at typical firms ^[5]. That is roughly 6.3 to 1 ^[6]. Plugins ran 21% against 9% ^[7], about 2.3 to 1 ^[8], which is where OpenAI's "six times" and "over twice the rate" summary comes from ^[9]. Both gaps describe the same thing: whether instructions and tool access sit in one person's prompt history or in a shared process. OpenAI's own example is a sales Plugin that pairs the company playbook with CRM access so an agent can retrieve account history, consult prior proposals and prepare a response for a person to review ^[10].

Now the number doing the promotional work. OpenAI defines frontier firms as the top 10% each month by output tokens per active user, and typical firms as the 45th to 55th percentile ^[11]. On that basis frontier firms generated 8.3 times as many output tokens per active user in June, up from 2.6 times in January ^[12], a 3.2-fold widening in five months ^[13]. The cohort is selected on the metric being reported, so what this shows is heavy users pulling away on the yardstick used to identify them, not better work or higher returns ^[14]. OpenAI concedes the proxy problem: a long response can be low value, a short one can settle an expensive question, and longer agent assignments naturally produce more tokens ^[15]. Membership is also recomputed monthly, so the January and June groups need not be the same firms ^[16].

The Codex spread is the more useful signal for staffing. Codex accounted for 64% of combined Codex and ChatGPT output tokens among enterprise customers in June ^[17], a figure OpenAI says reflects longer agent tasks as well as adoption and cannot be read as a share of users or assignments ^[18]. Since February, weekly active enterprise Codex users grew 108 times in legal, 41 times in sales, 41 times in recruiting and 26 times in marketing, against five times in engineering from a more established base ^[19]. Across more than 10 million messages, coding plus system and agent operations account for nearly three-quarters of agentic messages ^[20], with system and agent operations at 32% in recruiting, 26% in sales, 25% in policy and 24% in communications ^[21]. Multiples off a near-zero February base are cheap; the composition finding is the part that says non-engineers are now doing operations work.

The internal benchmark is the tell. OpenAI reports that 95% of its own active employees use Plugins weekly ^[22], roughly 4.5 times the rate at its best-performing customers ^[23]. That is a target the vendor is setting, not evidence of what the target earns. The flip side is the actionable half: 97% of weekly active users at typical firms used no skills at all ^[24].