In brief: a new OpenAI report shows not a gap in access to models, but a gap in the depth of how they are used. In June 2026, companies in the top 10% by output tokens per active user generated 8.3x more than typical companies; in January, the gap was 2.6x. Tokens are only a proxy, not ROI. The practical takeaway: scale not subscriptions and prompts, but proven workflows with context, tools, permissions, tests, and an owner.
This article is intended for executives, process owners, CIOs/CDOs, and implementation teams. It analyzes OpenAI data from August 12, 2026, and offers a method for replicating use cases. It does not measure the Russian market, does not prove a causal link between usage intensity and financial results, and does not promise to close the gap by exactly 8.3x.
Contents
- What OpenAI actually measured
- Five signals for leaders
- Why model access does not become a process
- Four levels of AI implementation maturity
- CONTOUR: how to replicate a workflow
- Reusable AI process card
- What metrics to track
- Implementation plan for one process
- Where companies break scaling
- Frequently asked questions
- How AI Dawn helps build a controlled AI workflow
- Conclusion
What OpenAI actually measured
On August 12, 2026, OpenAI published Enterprise Signals and the overview From assistance to execution. The researchers compared the depth of OpenAI product usage among enterprise customers and separately examined AI adoption across functions, roles, and tasks.
In Enterprise Signals, “leading” companies are the top 10% of customers by output tokens per active user per month. “Typical” companies are those between the 45th and 55th percentiles. In June, the former generated 8.3x more tokens per active user than the latter; in January, the ratio was 2.6x.
OpenAI explicitly calls tokens an imperfect measure. A long answer may be useless, and a short one may be valuable. So the metric should be read as a proxy for how deeply work is delegated to AI, not as productivity, profit, or quality.
The accompanying working paper How Organizations Use AI links anonymized ChatGPT Enterprise usage to roles and tasks. The workforce sample includes 1,764 organizations and 17,446,551 messages; the analysis captures the state six months after an organization is onboarded. Content was classified automatically; researchers did not read individual messages by hand.
Five signals for leaders
| Signal | Measured result | Practical meaning | What the result does not prove |
|---|---|---|---|
| The depth gap is widening | 8.3x in June versus 2.6x in January | the same license is not enough; organizational practice differs | that tokens created 8.3x more value |
| Work is becoming agentic | Codex accounted for 64% of total output tokens from Codex and ChatGPT among enterprise customers | employees are delegating longer, multi-step tasks | that 64% of processes are automated |
| Tools and skills are used unevenly | plugins: 21% of active users vs. 9%; skills: 19% vs. 3% | packaged instructions and system access create the advantage | that a plugin by itself boosts productivity |
| Growth has moved beyond engineering | since February, active Codex users have grown 108x in legal, 41x in sales and recruiting, 26x in marketing, and 5x in engineering | processes should be sought across all functions, not just in IT | the absolute number of users or the effect by function |
| More junior employees use it more intensively | early-career employees sent roughly 8–9 more messages per week than the average active user inside their company | the best use cases may emerge from the bottom up | that younger employees are more productive or should own control |
The main signal is that access is diffusing faster than the company’s ability to turn an individual habit into a durable way of working.
Why model access does not become a process
In implementation discussions, four states are often conflated:
- Access — an employee has been issued a license.
- Usage — the employee asks questions or creates materials.
- Repeatable workflow — the task has approved inputs, sources, instructions, permissions, validation, and an owner.
- Business outcome — compared with the baseline, the timeline, cost, quality, risk, or throughput changed.
The report measures the first two states well and shows movement toward the third. It does not measure the fourth. The working document separately warns: classified messages do not reveal downstream work products, performance, or changes in organizational procedures.
Code became an early test bed for a reason. The repository stores context, the task is often spelled out explicitly, and tests provide an acceptance criterion. In knowledge work, rules are often scattered across email, CRM, spreadsheets, and expert memory; a "good result" is not defined, and the cost of an error often becomes visible only after it is sent to the client.
That means the bottleneck is not only model quality. It becomes the work specification: what the agent reads, what it does, what confirms the output, when it asks for approval, and who is responsible for the outcome.
Four levels of AI adoption maturity
| Level | Observed behavior | Primary risk | Transition condition |
|---|---|---|---|
| 1. Access | licenses issued, basic training in place | activity is mistaken for value | identify a recurring real task |
| 2. Personal practice | employees save their own prompts and techniques | knowledge does not transfer, data is entered inconsistently | lock down the input, output, and quality criterion |
| 3. Shared workflow | there is a skill/instruction, a data source, permissions, and tests | the workflow expands without an owner or version control | run a pilot on a frozen dataset and assign an owner |
| 4. Managed operation | the scenario is built into the system, measured, updated, and rolled back | automation keeps running after rule changes | continuous quality monitoring and regression testing |
This model is an original synthesis by AI Dawn, not an OpenAI scale. Its goal is not to assign a company a score, but to identify the next testable transition.
CONTOUR: how to scale a working scenario
We propose the framework CONTOUR: Context → Operation → Access rules → Test → Logging → Replication.
Context
Document the approved sources: documents, spreadsheets, CRM fields, correspondence, glossary, and the effective date. Separate company facts from the model's general knowledge. If one source conflicts with another, define priority or an escalation path.
Operation
Describe the task as a transformation: event and input → actions → verifiable output. "Help sales" is not an operation. "Before the meeting, compile a customer brief from the CRM and the three most recent meeting notes using a specified template" is.
Access rules
Name the allowed data, tools, and actions. Separate reading, drafting, suggesting changes, writing after approval, and autonomous action. Secrets, personal data, and legally significant operations require separate rules.
Test
Freeze a set of real anonymized cases, the rubric, and stop rules before launch. Test not the persuasiveness of the text, but the facts, completeness, allowed sources, format, critical errors, and escalation path.
Logging
Assign a process owner and keep versions of the instructions, sources, model, output, edits, and reviewer decision. The log is needed not for total surveillance, but to explain a failure and safely update the workflow.
Replication
After the pilot, package the working practice into a shared skill or SOP, train on concrete examples, and grant access by role. Move the workflow to a neighboring team only after checking differences in data, risk, and acceptance criteria.
CONTOUR turns an individual user's talent into an organizational asset, while keeping accountability with the process owner.
Reusable AI process card
For each candidate, fill out one card:
| Field | What to record | Example |
|---|---|---|
| Event | what triggers the work | new opportunity in CRM |
| Input | required data and format | customer card, 3 notes, price list |
| Sources | approved systems and date | CRM and knowledge base, updated daily |
| Output | specific artifact | meeting brief in the approved template |
| Acceptance criterion | verifiable conditions | all figures with references, 5 required sections |
| Prohibited error | what stops the workflow | fabricated fact or someone else's data |
| Permissions | read/draft/write | read CRM, draft without sending |
| Escalation | when a person is needed | source conflict or a blank required field |
| Owner | who is responsible for the outcome | head of sales operations |
| Version | instructions, model, tests | skill 1.2, test set 2026-08-13 |
This card is intentionally shorter than the technical requirements. Its purpose is to check at a glance whether there is actually a process that can be scaled.
Which metrics to measure
| Metric | Formula | What it shows | What it does not show |
|---|---|---|---|
| Weekly active coverage | active users / target group | breadth of adoption | quality or value |
| Workflow reuse | runs of the shared workflow / all relevant cases | repeatability | the correctness of each result |
| Acceptance rate | accepted results / reviewed results | the usefulness of the output | hidden errors outside the rubric |
| Critical error rate | critical errors / reviewed results | risk by the defined taxonomy | the completeness of the taxonomy itself |
| Review time | median review time in minutes | human workload | queue wait time |
| Cycle time | median from event to acceptance | end-to-end speed | causality without a control comparison |
| Cost per accepted unit | model, infrastructure, review, and rework / accepted units | the economics of the workflow | future revenue |
| Replication rate | teams that passed the gate / candidate teams | portability of the practice | the effect in a new team without retesting |
Output tokens are useful as a diagnostic trace: a longer agent task usually produces more output. But this is not a success KPI. The KPI should be tied to accepted units of work and risk.
Implementation plan for one process
The four-week cycle below is an Estimated starting format, not a promised timeline for every company.
- Week 1 — discovery. Gather 10–15 real individual scenarios from active users, including early-career employees. Choose one common process with an observable output and an owner.
- Week 1 — baseline. Measure current cycle time, rework, critical errors, cost, and queue volume.
- Week 2 — card and permissions. Fill out the CONTOUR, connect only the required sources, and keep external actions turned off.
- Week 2 — test set. Prepare 30–50 anonymized cases; the volume is an
Estimatedstarting point and should grow with variability and risk. - Week 3 — shadow mode. AI produces the result in parallel with the current process, but does not send or change anything.
- Week 3 — gate. The owner compares the result with the baseline using preapproved metrics and stop rules.
- Week 4 — packaging. Document the skill, sources, permissions, tests, change log, and brief team training.
- Week 4 — limited rollout. Connect one role or a neighboring team and repeat the gate on its data.
The first goal is not to “get everyone engaged,” but to prove that one shared workflow preserves quality when it is handed off from the author to other employees.
Where companies break scaling
They buy seats instead of changing the work
The share of active users grows, but the decisions, sources, and handoff stay the same. Fix: choose one process outcome and work backward from it to the data, permissions, and review.
They copy the prompt without context
The wording is transferred, but the CRM fields, vocabulary, template, and exceptions stay in the author's head. Fix: package not just the prompt text, but the entire CONTOUR.
They make tokens the KPI
The team starts producing more output instead of accepted work. Fix: measure cost per accepted unit, critical errors, and review time.
They make the experienced user the owner of risk
An employee found a strong use case, but should not single-handedly determine access to data or approve external actions. Fix: separate the roles of scenario scout, process owner, IT/security, and reviewer.
They scale before regression testing
The workflow works on ten convenient examples and breaks when the team or source changes. Fix: the new group gets a new sample, a separate gate, and the ability to roll back.
Research limitations
| Limitation | Why it matters |
|---|---|
| The data applies to OpenAI customers | the use of other models, APIs, internal systems, and personal accounts is not fully covered |
| Output tokens are a proxy for depth | they do not measure quality, productivity, profit, or causal effect |
| Job data is incomplete | it is not possible to calculate the share of all employees in each role who adopted AI |
| Tasks are classified automatically | taxonomy and message classification errors are possible |
| The working paper analyzes messages, not the final product | downstream outcomes and process changes were not measured |
| The financial comparison is limited to related U.S. public companies | it is descriptive, not causal, and does not generalize to Russian companies |
Important detail: in the working paper, early adopter companies are larger and more R&D- and SG&A-intensive than non-adopters. So you cannot conclude that simply adopting ChatGPT created their advantages; more likely, underlying organizational capabilities help companies both adopt and embed the technology more deeply.
Frequently asked questions
What does the 8.3x gap mean?
In June 2026, the top 10% of OpenAI enterprise customers by output tokens per active user generated 8.3 times more of them than companies between the 45th and 55th percentiles. This is a proxy for depth of use, not a ROI gap.
Should AI be given to all employees?
Broad, safe access helps surface use cases, but access alone does not create a process. It needs to be combined with specific tasks, approved sources, permissions, tests, and training.
Why do companies need AI skills and plugins?
A skill packages a repeatable instruction, while a plugin can add access to data, tools, and actions. Together, they reduce dependence on an individual prompt, but they require versioning, permissions, and validation.
Who should look for deployment use cases?
Employees who already use AI intensively can scout for use cases. The process owner defines the outcome and risk; IT and security define access; a subject matter expert handles acceptance. One enthusiast should not fill all roles.
How can you tell a use case is scalable?
Another employee using the same card gets a result that passes frozen criteria; critical errors do not exceed the stop rule; permissions are limited; the owner, version, and rollback are defined; and total cost is better than the baseline on the target metric.
How AI Dawn helps build a controlled AI environment
AI Dawn can help turn disconnected personal workflows into one testable process:
- Audit the process and document the baseline, data, constraints, owner, and acceptance criteria.
- Prepare a corporate knowledge base or RAG environment, a version-controlled skill, and a test set.
- Build an AI agent or agentic RPA with minimal permissions, approvals, an activity log, and escalation.
- Integrate the gate-passed environment with corporate systems, train the team, and support updates.
The safest first step is to choose one process, its current baseline, data sources, constraints, and acceptance criteria. This lets you test value before giving an agent write access or external action rights and does not assume predetermined timelines, savings, or ROI.
Conclusion
A new OpenAI report shows that the enterprise gap exists even when companies have access to the same models. Leading companies delegate work more deeply to agents and are more likely to use packaged capabilities such as context, tools, plugins, and skills. But 8.3x remains a measure of intensity, not proven profit or productivity advantage.
The practical answer for leaders is not to maximize tokens. Find one personal workflow that solves a recurring task, build a CONTOUR for it, compare it with the baseline, and scale it only after the gate. That is how an experiment becomes a reproducible company asset, not a trick that disappears with the tab or the employee.