The AI adoption gap is no longer about who has a chatbot license. It is about who can delegate complete work. Leading teams give agents approved context, tools, persistence, and review paths. Other teams still ask isolated questions and carry every step themselves. Business leaders can close that gap with role-based AI training, shared workflows, and controls built around actions.
The practical move is to find employees already completing dependable work with AI. Turn their methods into shared playbooks. Connect only approved context and tools. Define what agents may read, draft, write, send, publish, or transact. Keep people accountable for consequential actions. Expand one verified workflow at a time beyond engineering. Measure accepted work, cycle time, rework, and business results. Feed failures and successes back into training so every team advances.
What OpenAI's August 12 Enterprise Signals report says
OpenAI's August 12 Enterprise Signals update defines frontier firms as the top 10% of enterprise customers by monthly output tokens per active user. Typical firms sit between the 45th and 55th percentiles. By June, frontier firms generated 8.3 times more output tokens per active user. The gap was 2.6 times in January.
The composition of use changed too. Codex produced 64% of combined ChatGPT and Codex enterprise output tokens in June. OpenAI treats Codex tokens as a measure of agentic use. Frontier firms also used advanced capabilities more often. Among weekly active users, 21% used plugins and 19% used skills. Typical firms reached 9% and 3%, respectively.
Those measures do not prove business value. OpenAI explicitly calls tokens an imperfect proxy. Long output can be wasteful, while one short answer can be decisive. The report still reveals an important pattern. The widening gap is associated with deeper delegation, reusable instructions, connected tools, and longer work.
Assistance and execution are different operating models
An assistant answers a question, drafts a passage, or suggests a next step. The employee still finds each source, moves data, applies policy, updates systems, and closes the loop. An agent receives a defined outcome. It can gather context, use permitted tools, complete multiple steps, test its work, and return evidence for review.
OpenAI's research on how agents are transforming work describes that shift as a new unit of knowledge work. Chat interactions are often short and self-contained. Agents can operate for minutes or hours while using tools and iterating toward a result. In May, 70.2% of sampled individual Codex users made at least one request estimated above one hour of human work.
That estimate is directional, not exact. OpenAI used a model to estimate human task time from a 0.1% sample. Leaders should not turn it into an hours-saved claim. The useful lesson is that agent use increasingly covers longer tasks. Those tasks need clearer outcomes, access boundaries, acceptance tests, and accountable reviewers.
The transition has also moved beyond developers. Since February, weekly active enterprise Codex users grew 108 times in legal. Sales and recruiting each grew 41 times. Marketing grew 26 times, compared with five times in engineering. A custom AI agent is therefore an operating-model decision for many functions, not only an engineering tool.
Current coverage shows a skills and governance gap
Deloitte's 2026 State of AI in the Enterprise reports that the skills gap is the largest integration barrier. It also finds only 34% of surveyed organizations are truly reimagining the business. Another 37% use AI at a surface level with little process change. Only one in five has a mature governance model for autonomous agents.
These findings explain why license access cannot close the gap. Agentic work needs role fluency and workflow redesign. It also needs governance that matches increased action. An assistant can produce a weak draft. A connected agent can update a record, send a message, publish content, or start a transaction. Authority raises both value and consequence.
RingCentral shows experimentation becoming infrastructure
OpenAI's August 12 RingCentral case study provides a concrete operating pattern. RingCentral's Office of the CEO sponsored an AI-Native Challenge. Participants received ChatGPT Work and Codex, then built complete projects. Thousands of employees took part, including nontechnical staff and executives. Nearly every participant produced a working repository.
The case did not present autonomy as a substitute for accountability. RingCentral kept people responsible for requirements, business context, architecture, testing, and verification. That is the useful pattern for leaders. Broad experimentation can discover frontier users, but reviewed outputs determine what deserves to become a shared method.
RingCentral's PMO then moved from experimentation into recurring operations. ChatGPT Work connected context across Jira, Google Sheets, CRM systems, and other sources. Workflows supported status tracking, reporting, release governance, and knowledge transfer. Meetings began with blockers, owners, and actions already surfaced. The metric was operational readiness, not chat activity.
OpenAI's current ChatGPT Work examples add another useful detail. RingCentral reports scaling an early-access program from six pilot customers to about 80. Shopify describes research across 3,500 non-R&D employees to find high and low adopters. The goal is to spread successful patterns and raise the floor. That is an adoption system, not a prompt contest.
A realistic mid-sized business scenario
Consider a 140-person commercial services company. One sales operations manager already uses AI to prepare weekly pipeline reviews. She combines CRM opportunities, meeting notes, proposal history, and renewal risks. Her report flags stalled deals, missing owners, and the next action. Other managers still paste pipeline totals into chat and rewrite the answer by hand.
Leadership should not simply tell everyone to copy her prompts. The company first documents the recurring decision: which opportunities need intervention this week. It checks the sources and quality rules. It turns her method into a shared playbook. Then it connects approved CRM fields and proposal examples with read-only access.
The agent may draft account briefs and proposed actions. A sales leader approves external follow-up. CRM writes remain limited to selected fields and produce an action log. The company measures accepted briefs, preparation time, corrections, pipeline actions, and conversion outcomes. After a controlled pilot, the method becomes part of CRM and sales AI training for every manager.
The same architecture can later support contract intake, candidate coordination, or campaign reporting. Each function receives separate sources, permissions, acceptance tests, and owners. Shared principles scale. Unreviewed authority does not.
The seven-decision adoption checklist
Use this matrix before expanding any individual workflow. Every row needs a named owner and retained evidence.
| Decision | Start | Control | Evidence |
|---|---|---|---|
| Frontier users | Find people already completing valuable, repeatable work with AI | Review quality, data handling, and judgment before copying their method | Finished work, time saved, exception rate, and named workflow owner |
| Shared playbooks | Turn a proven personal workflow into reusable instructions and examples | Version the playbook and assign an owner for every material change | Workflow definition, approved examples, test cases, and change log |
| Context and tools | Connect only the systems and records needed for the defined outcome | Use approved sources, least privilege, data boundaries, and access reviews | Source inventory, permission map, connection owner, and access logs |
| Agent actions | Separate read, draft, write, send, publish, and transact permissions | Require human review for high-impact, external, or hard-to-reverse actions | Approval record, action log, exception queue, and rollback procedure |
| Function expansion | Pilot one bounded workflow in sales, legal, recruiting, or marketing | Give each function its own data, policy, quality, and review criteria | Function owner, baseline, acceptance test, and expansion decision |
| Completed work | Measure dependable outcomes instead of chats, seats, prompts, or tokens | Count work only when it meets quality, timeliness, and policy standards | Accepted deliverables, cycle time, rework, exceptions, and business result |
| Training loop | Pair frontier users with teams still limited to basic prompting | Teach the workflow, test it in role context, and refresh it from failures | Practice completion, adoption by cohort, quality gains, and updated playbook |
Identify frontier users by finished work
Find the people already using AI to complete valuable recurring tasks. Ask managers where cycle time fell without creating more rework. Review actual deliverables and decisions. Look across job level and function. OpenAI found early-career employees used AI more, so seniority should not decide who teaches the method.
Do not publish a leaderboard based on prompts, tokens, or time in the tool. That rewards activity and may encourage sensitive or low-value use. A frontier user should demonstrate accepted work, sound judgment, compliant data handling, and a method another person can repeat.
Turn individual workflows into shared playbooks
Capture the outcome, trigger, required inputs, sequence, exceptions, acceptance criteria, and reviewer. Include approved examples and known failure cases. Package reusable instructions as a skill, template, or standard operating procedure. Give the playbook an owner, version, and review date.
A playbook should reduce dependence on one person's memory. It should not freeze experimentation. Teams can propose improvements, but changes should pass the same tests as the current version. Our guide to agentic workflows in operations applies the same owner-and-gate discipline to production agents.
Connect approved context and tools
Agents become useful when they can reach the right work context. They also become risky when their reach is vague. Start with a data and AI readiness audit. List each source, record owner, sensitivity, retention rule, and permitted use. Remove stale and conflicting sources before connecting them.
Then grant the smallest tool set needed for the outcome. A sales brief may need read access to selected CRM records and approved proposals. It does not need permission to export the whole CRM. Test revoked access, missing records, conflicting instructions, and sensitive content before launch.
Define permissions and human review for actions
Write an action matrix for read, draft, write, send, publish, and transact. Each level needs a business owner, technical control, log, exception path, and rollback plan. Require human approval when an action affects customers, candidates, contracts, money, regulated records, or public claims.
Review requirements can shrink only after evidence supports the change. A low-risk internal draft may graduate to automatic generation with sampled review. A contract commitment or customer send may always need approval. Our AI governance and training guide provides the broader policy layer.
Expand beyond engineering one workflow at a time
Engineering moved first because code has rich context and testable outputs. Other functions need equivalent verification. Sales can test account briefs against CRM facts. Legal can test clause extraction against an approved rubric. Recruiting can test candidate packets for completeness and bias controls. Marketing can test campaign packages against source facts and brand rules.
Start where the decision repeats, sources are available, and quality can be evaluated. Use ChatGPT and Codex training to teach each function how delegation, context, and review differ from basic prompting. Do not force every team into one generic agent workflow.
Measure dependable work completed
Chat volume shows exposure, not adoption value. Count a unit only when the work meets quality, timeliness, security, and policy standards. Sales might count approved account briefs that trigger a documented action. Legal might count reviewed intake packets. Recruiting might count complete interview plans. Marketing might count campaign analyses accepted without factual correction.
Pair volume with cycle time, reviewer corrections, exception rate, rework, incidents, and the business result. Track cost per accepted unit. Compare against the old process. The strongest scorecard shows whether dependable capacity increased, not whether employees talked to AI more often.
Create training loops that raise the floor
Pair frontier users with the teams still using basic prompts. Run role-based labs on real, sanitized work. Let learners practice scoping an outcome, choosing approved sources, setting permissions, checking evidence, and handling exceptions. Assess the result instead of attendance alone.
Feed production failures into the next training session and playbook revision. Share strong examples in an internal library. Hold short showcases where owners explain what changed and what remains human. Recognize reviewers and workflow maintainers, not only builders. This loop converts individual advantage into organizational capability.
A 90-day path from assistance to execution
During the first month, inventory current use and select one recurring workflow. Identify the frontier user, owner, sources, baseline, and acceptance test. During the second month, build the shared playbook and run it with limited access. Review every output and log every exception.
During the third month, expand to a small cohort. Compare completed work, quality, cost, and cycle time with the baseline. Train lagging users on the same workflow. Expand tools or action authority only when the evidence clears the control gate.
The widening AI adoption gap is not inevitable. It grows when successful methods remain personal and ungoverned. Leaders can close it by converting frontier behavior into shared systems. Give agents approved context, bounded authority, human accountability, and measurable work. Then use training to make the next reliable workflow easier than the first.
