Enterprise AI Adoption’s 8.3× Usage Gap

Among OpenAI enterprise customers, firms in the top tenth of monthly usage generated 8.3× as many output tokens per active user as typical firms in June 2026, according to OpenAI’s enterprise analysis. That is a gap in generated output among people who used the tools, not a measured gap in completed work or business performance.

Organizations use ChatGPT Enterprise for writing, research, technical work, communication and analysis across many roles. Adoption deepens when people return to those tasks and incorporate AI into workflows; access alone does not determine how often they use it or how much they ask it to produce. The evidence separates three questions that a rollout can blur: who participates, what they do, and whether the resulting work helps the business.

What the 8.3× comparison measures

The comparison ranks enterprise customers each month by output tokens per active user. Its frontier group is the highest-usage tenth, while its typical group sits around the middle of the distribution. Dividing by active users removes the simple effect of having more users, but it does not make the firms’ tasks, tool settings or work practices equivalent. Membership can also change from month to month, so the comparison is not a fixed group of companies pulling away over time.

An output token is a unit of generated content, not an accepted deliverable. A long agent task that searches, drafts and revises may produce far more tokens than a short answer to a valuable question. The ratio therefore captures depth of measured use among active users. It cannot tell whether an answer was accurate, whether a worker used it, or whether the employer saved time. It also says nothing by itself about the share of each firm’s entire workforce that participates.

How use spreads across firms and roles

The companion ChatGPT Enterprise working paper links account activity, inferred roles and classified tasks through March 2026; its worker sample at six months after adoption contains 1,764 organizations and more than 17 million messages. It describes growth from both newly adopting firms and greater use within existing customer cohorts. This is a different sample and period from the June enterprise-customer comparison behind the headline ratio.

Within adopting organizations, active users appear in engineering, finance, marketing, sales and leadership. Managers and directors form a substantial share of observed active users, but the busiest groups are not necessarily the largest. Analysts and marketing and communications workers send more messages than the average active user at their own firms; executives send fewer. Early-career workers and trainees are especially intensive users once active. These are comparisons of message activity, not rankings of the economic importance of each role.

The role findings have a denominator problem. Administrative job titles are incomplete, and the researchers do not observe every employee in each role. A share of active users cannot therefore be converted into the percentage of engineers, managers or trainees who adopted the product. The paper’s separate U.S. public-company analysis finds adoption more common among larger firms and firms with stronger prior revenue per employee and investment in organizational or technical capabilities. Those characteristics predate or accompany adoption; they are plausible reasons firms may differ in their ability to deploy AI, not benefits shown to have resulted from it.

Which tasks account for use

OpenAI’s early-adoption analysis identifies writing, research, programming and analysis as major task groups during departments’ first three months with ChatGPT Enterprise. Engineering users lean toward programming while also seeking research and documentation help. Marketing and other customer-facing teams use it more for writing, research and creative work. The pattern is role-specific use of a general tool, rather than a single workflow shared by every department.

The working paper distinguishes a task’s reach from its volume. Reach asks whether an active user performs a task during a week; volume asks how many messages that task generates. Documentation and technical writing, technical digital work, and message drafting account for substantial message traffic. Topic overviews, market research, and legal or financial questions can reach many users without producing comparably large message counts. A brief exchange may resolve a useful question, while a long exchange may end in discarded work, so neither count alone establishes value.

Why intensive firms produce more output

OpenAI’s Enterprise Signals data reports that Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers in June 2026; weekly Plugin use reached 21% of active users at frontier firms versus 9% at typical firms. The source also finds more use of skills at frontier firms. These capabilities can give agents reusable instructions and access to company context and tools, making longer delegated tasks possible.

That pattern helps explain why output volume can diverge even when firms have access to the same underlying models. An agent carrying out a multistep assignment can generate extensive output while finding information or revising files; an employee using chat for a concise answer may generate little. The observed association between intensive firms and advanced capabilities does not identify which feature caused the gap. It also leaves room for differences in task mix, permissions, employee habits and the extent to which useful individual practices spread within a firm.

What the data can establish about performance

The public-company comparison finds stronger financial characteristics among adopters and shows high-intensity adopters toward the upper end of revenue and market value per employee. Its financial sample is limited to matched U.S. public companies, and its usage measures differ from the headline ratio. The analysis does not establish that ChatGPT Enterprise raised revenue, valuation or productivity: better-resourced firms may have been more likely to adopt and use it intensively in the first place.

The account data also omit work done through other AI systems, personal accounts and some internally built tools. Classified messages describe likely tasks, but they do not show whether a finished output passed review, reduced rework or changed a business result. For a rollout, the useful measurement sequence is to distinguish eligible staff from active users by role, examine repeat use and task mix, then assess accepted work, quality and time against an appropriate baseline. That framework treats the 8.3× ratio as a reason to investigate how work is performed, not as a multiplier for business value.