The gap in enterprise AI adoption is no longer simply about whether a company uses AI. It lies in whether the company stops at a helpful conversation or moves toward workflows in which AI can perform work under control. This is a notable point in OpenAI’s analysis published on August 12, 2026: firms described as “frontier firms” generated 8.3 times as many output tokens per active user as typical firms in June, compared with a 2.6-times gap in January.
This figure is not a direct measure of productivity, revenue, or management quality. It is a platform-based indicator of depth of use and should be read within that limit. Yet, considered alongside observations about connecting agents to context, tools, and repeatable workflows, it signals an important management shift: advantage may come from the ability to organize work around AI, not merely from purchasing access to a model.
The AI gap is not closed by adding more accounts. It is closed when an effective individual practice becomes a shared workflow with clear authority, data, and review standards.
From answer assistance to work execution
OpenAI describes a move from “assistance” to “execution”: rather than merely asking AI how to do a task, users assign an agent a sequence of steps for human review. The report states that by June, Codex accounted for 64% of the combined output tokens from Codex and ChatGPT among the enterprise customers observed. This does not mean that every company should automate more. It indicates that multi-step tasks involving tools and context make up a larger share of AI activity in this sample.
The management boundary between the two states is very different. In an assistance state, an individual may use AI to summarize, suggest, or draft. In an execution state, the company must decide which data an agent may see, which systems it may call, which actions it may take, when approval is required, and who is accountable when an outcome is wrong. Value rises, but the requirements for work design and control rise with it.
Why are leading firms creating the gap?
The publication identifies a practical difference: among weekly active users, frontier firms use plugins at 21% and skills at 19%, compared with 9% and 3% respectively at typical firms. These capabilities matter because they place reusable instructions and connections to data or tools in the same workflow. In the report’s illustration, a sales agent can combine a playbook with a CRM to prepare a context-specific response for the responsible person to review.
The implication is not “install more plugins”. A skill creates value only when it encodes a tested way of working, specifies inputs, outputs, limits, and the handoff point to people. If a playbook remains ambiguous, customer data are fragmented, or approval policy is unclear, connecting an agent risks automating ambiguity. Teams should therefore treat each AI workflow as an operating-design object, not as a software feature.
Where the gap is appearing
According to the report, weekly active enterprise Codex users grew rapidly since February in legal, sales, recruiting, and marketing; the stated rates were 108 times, 41 times, 41 times, and 26 times respectively, compared with five times in engineering. These rates reflect platform data, not a representative picture of every company or country. They are nevertheless a useful signal: AI agents are reaching knowledge work that has rules, documents, approvals, and handoffs, rather than remaining solely a tool for technical teams.
Management example
A marketing team can begin with an agent that consolidates campaign feedback, checks it against the brief, and prepares a weekly report. The first version should not independently change advertising budgets or send messages externally. Only when quality and exception rules have been tested should the team consider granting more authority. This is scaling by “permission to work”, not only by “ability to answer”.
Three things managers can do now
First, choose a workflow with sufficient frequency, reasonably clear inputs, and controllable risk. Do not start with the broadest or most sensitive problem. Second, document the best version of how one employee is using AI: which data are used, which steps repeat, which quality standards are checked, who approves, and which errors must be escalated. This is the material for turning an individual habit into a team skill, SOP, or checklist.
Third, measure two sets of indicators in parallel. Efficiency measures include cycle time, throughput, rework rate, and response speed. Control measures include errors, policy violations, human-handoff rate, complaints, and cases where the agent exceeded its authority. Measuring only speed can hide quality costs. Measuring only risk does not show whether the workflow is creating additional value.
Conclusion
The management message in these data is that the AI gap is organizational. Models may become increasingly accessible, but data, authority, review processes, and the ability to learn from real use do not appear automatically. Companies should use current figures as a signal to prioritize workflow design and management capability, not as a promise of results. In the next stage, the organizations that move faster may not be those using AI most, but those that know how to turn good AI use into a reliable shared practice.
References
- OpenAI. (2026, August 12). From assistance to execution: How enterprises put AI to work. https://openai.com/index/how-enterprises-put-ai-to-work/
