OpenAI says enterprise AI adoption is moving beyond question-answering toward delegated, multi-step work, but its own data shows that progress is highly uneven. The company’s new research describes a widening gap between “frontier” firms and typical adopters, with the most active organizations using agents, company context and connected tools more extensively.

The findings come from two OpenAI studies: Enterprise Signals, which examines agentic AI across the company’s enterprise customer base, and the working paper How Organizations Use AI: Evidence from ChatGPT. OpenAI says the studies draw on more than 10 million messages and examine adoption across firms, functions and levels of seniority.

Codex points to a shift toward delegated work

The clearest signal is the reported growth of Codex, OpenAI’s coding and software-work agent. As of June, Codex generated 64% of combined Codex and ChatGPT output tokens among enterprise customers, according to OpenAI. The company argues that the measure indicates a move toward longer, more substantive tasks because agentic workflows typically produce more output than short conversational requests.

That interpretation matters for enterprise buyers. An assistant may help an employee reason through a presentation or research task; an agent can gather information, use connected systems, create files and prepare a result for human review. OpenAI positions ChatGPT Work and Codex as examples of this model, although the data does not establish how much of the reported output was accepted without substantial human editing or how much work was completed successfully.

The use of Codex is also spreading beyond engineering. 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, OpenAI reports. Engineering grew five times over the same period. These are large relative growth figures, but OpenAI does not provide the underlying user counts in the supplied material, so they should not be read as evidence that nontechnical teams now have larger deployments than engineering teams.

OpenAI cites Virgin Atlantic as an example of cross-functional use. The airline says engineering teams have used Codex to refactor legacy code in 30 minutes rather than two weeks, while product teams used ChatGPT Work for competitive research. Those statements are examples supplied by OpenAI and are not independently verified in the company’s research summary.

The frontier gap is widening

OpenAI ranks enterprise customers each month by output tokens per active user. It defines frontier firms as those in the top 10% and typical firms as those around the 45th to 55th percentiles. In June, frontier firms generated 8.3 times as many output tokens per active user as typical firms, compared with a 2.6-times difference in January.

The company describes output tokens as a proxy for depth of use, not a direct measure of business value. More generated text or code can reflect longer workflows, but it can also result from inefficient prompting, repeated revisions or tasks that require substantial review. The widening ratio therefore indicates a difference in usage intensity, while leaving open questions about productivity, quality and return on investment.

The same pattern appears in the use of advanced capabilities. OpenAI says 21% of weekly active users at frontier firms used Plugins, compared with 9% at typical firms. Skills usage was 19% at frontier firms versus 3% at typical firms. Plugins can connect an agent to company data and actions, while skills provide reusable instructions, according to OpenAI.

OpenAI also reports that 95% of its own employees use Plugins weekly. Because OpenAI is both the product vendor and the source of this internal comparison, the figure is best treated as a vendor-reported adoption signal rather than a market benchmark.

Evidence is useful but controlled by the vendor

The research offers a view that conventional surveys often miss: actual usage patterns in workplace conversations and product telemetry rather than employees’ descriptions of their behavior. OpenAI says its data shows early-career employees using ChatGPT more frequently than executives. Six months after adoption, early-career workers sent 13 more messages per week than executives, the company reports.

OpenAI also says that, among the U.S. public companies studied in its working paper, enterprise adopters had more assets, employees and R&D investment than non-adopters. That is an association, not proof that AI adoption produced stronger financial performance. Larger companies may simply have more resources to buy, deploy and govern AI systems.

The central claims remain vendor-reported. The supplied evidence does not include an independent methodology review, customer-wide cost data, error rates, task completion rates or a comparison with non-OpenAI systems. Buyers should therefore use the reports to form hypotheses about deployment patterns, not as a neutral measurement of the enterprise AI market.

What the findings mean for AI builders and enterprises

For product teams, the research points to a shift in the unit of design. The valuable product may not be a general-purpose chatbot, but a workflow that combines instructions, permissions, business data and actions. A sales Plugin, for example, could combine a company playbook with CRM access and past proposals before producing a response for review.

That architecture introduces operational requirements that are less important for simple assistance. Teams need permission boundaries, audit logs, data access controls, approval steps and ways to recover when an agent uses the wrong context or takes an unsuitable action. Human review remains particularly important when outputs affect customers, legal work, financial decisions or production systems.

The adoption gap also suggests that buying access to an AI model is not the same as scaling usage. OpenAI’s own conclusion is that companies need continuous employee learning, shared workflows, data infrastructure and governance. For enterprise leaders, a practical starting point is to identify effective individual workflows, measure where they save time or improve quality, and convert them into repeatable processes rather than relying on informal prompting expertise.

For competitors and builders, the expansion of AI agents into legal, sales, recruiting and marketing raises the stakes around integrations. The systems that can safely retrieve current business context and complete bounded actions may become more useful than systems that only generate polished text. But the research does not show whether OpenAI’s products outperform alternatives on reliability, total cost or security.

What to watch next

The next signals will be more detailed evidence on task completion, review rates and measurable business outcomes. OpenAI’s reports may also clarify how the 8.3-times output gap varies by industry and function, and whether high token use translates into lower operating costs or faster delivery.

Enterprise buyers should watch for broader availability of agent permissions, workflow templates and audit controls in ChatGPT Work, Codex and Plugins. They should also look for independent customer evaluations that report unsuccessful runs, human intervention and deployment costs, not only usage growth.

A further test will be whether adoption continues outside engineering once agents are connected to sensitive enterprise systems. Growth in legal, sales, recruiting and marketing is notable, but sustained usage will depend on data quality, compliance review and whether agents can complete work reliably enough to earn trust.

Creati.ai perspective

OpenAI’s announcement is less a claim that enterprises have solved agentic AI than a measurement of how unevenly they are approaching it. The reported frontier gap suggests that the competitive advantage may come from organizational capability—workflow design, data access and governance—as much as from model access.

The important question for builders is therefore not simply how many employees use an AI tool. It is whether a system can complete a defined piece of work, show what it did, stay within its permissions and produce an output that people can verify. OpenAI’s data makes that transition visible, while its vendor-controlled methodology means the market still needs independent evidence on quality and economics.