Human-resources workers are among the ChatGPT users most likely to ask for help with tasks outside their stated profession, but OpenAI’s new 69% figure is narrower—and more consequential—than HR Dive’s weekly roundup makes it sound. The number applies only after OpenAI removes generic work such as drafting, summarizing and scheduling, leaving occupation-specific messages; it does not mean that 69% of all HR prompts fall outside HR.
HR Dive highlighted the finding from OpenAI’s July 27 Work at the Frontier report alongside a ManpowerGroup Talent Solutions survey claiming only 3% of organizations have leaders “highly prepared” for AI-enabled work. Taken together, the evidence points to a practical problem for IT and HR: workers are already using AI as an informal bridge to finance, marketing, legal and technical work, while management’s rules for approving, reviewing and owning that work remain thin.
The danger is not that an HR generalist asks ChatGPT to draft a flyer or explain a spreadsheet formula. It is that the same tool makes unfamiliar work easy to attempt, which can erase the normal handoff to the person accountable for the result. In HR, where prompts can touch pay, hiring, employee relations, benefits and policy interpretation, that is a governance issue before it is a productivity story.
OpenAI analyzed more than 800,000 work-related messages from U.S. users whose self-reported role information came from ChatGPT Business accounts. The messages themselves, however, came from those users’ individual ChatGPT accounts, not their ChatGPT Business or Enterprise conversations. OpenAI says it used automated classification without employees reading the underlying messages.
That distinction is missing from the headline takeaway. The report is not a measurement of governed enterprise AI use inside a company tenant, and it is not a survey of the U.S. HR workforce. It is a descriptive sample of individual-account activity associated with people who had also provided Business-account role information. OpenAI explicitly says the estimates should not be generalized to Enterprise users or to the entire workforce.
The headline percentage also has a denominator problem that readers should keep in view. Across the full sample, 61.5% of messages were classified as generic: activities shared broadly among occupations, including emails, writing, summarization and scheduling. Of all work-related messages, 16.8% were classified as cross-occupation work. Once generic tasks are removed, the cross-occupation share rises to 43.5% overall and 69% for HR.
In other words, the HR result says that among HR prompts OpenAI considered specific enough to map to a traditional occupational task, more than two-thirds resembled work normally associated with another occupation. It does not establish that HR has suddenly stopped doing HR work, nor that the model performed the work correctly, nor that its output was used in a real business decision.
OpenAI is unusually direct about those limitations. The study does not observe whether the answer was adopted, whether it saved time, whether the user could have completed the task without AI, or whether a qualified specialist reviewed it. It also assigns each message one primary work activity, even though real requests may involve several tasks and roles.
Those caveats do not make the finding irrelevant. They define the finding: ChatGPT is being used at the edge of job boundaries, where the user is attempting work that may previously have gone to another function.
For HR, OpenAI’s table assigns 18% of non-generic messages to engineering-like tasks, 16% to finance, 23% to marketing, 10% to legal, 9% to customer experience and only 10% to HR itself. Those row shares are a classification of task resemblance, not a claim that HR teams are replacing IT, finance or counsel. But they show why a simple policy saying “AI may be used for drafting” no longer describes what people are doing.
An HR practitioner may use ChatGPT to troubleshoot a spreadsheet, interpret a technical system issue, calculate compensation scenarios, prepare communications material or understand a regulation. Each may begin as a sensible request for assistance. The operational question is what happens after the draft or explanation arrives: Does the output move into a payroll model? Does it shape a benefits answer? Does an HR business partner make a policy call based on a model’s summary? Does a manager act on a technical recommendation that bypasses the service desk or security team?
The model can reduce the friction that once forced a handoff. It cannot transfer responsibility for the decision. That is the line organizations need to make visible.
For Windows administrators and IT leaders, this reinforces the need to treat AI access as an application-governance issue rather than an employee training perk. A company cannot meaningfully govern sensitive HR use if it only licenses an approved assistant while employees continue putting work into personal accounts, browser sessions or unmanaged AI tools. OpenAI’s research itself captures individual-account behavior linked to Business users, a reminder that the boundary between sanctioned and unsanctioned usage is often porous.
The immediate controls should be mundane but specific: approved tools, clear data classifications, identity-backed access, retention rules, auditability, and an escalation path when an AI-assisted task crosses into payroll, legal interpretation, security troubleshooting or an employment decision. “Use good judgment” is not a control when the point of the tool is to let people act outside their established expertise.
The 3% finding is striking, but its evidentiary base is small. The research drew on 80 C-suite, CHRO and senior talent-acquisition leaders in the United States and United Kingdom across healthcare, life sciences, manufacturing and technology. ManpowerGroup commissioned the report, which Everest Group developed. A percentage of 3% in a sample of 80 equates to roughly two or three respondents, depending on rounding.
That does not invalidate the survey’s direction of travel. It does mean the figure should not be presented as a precise estimate of leadership readiness across all employers. ManpowerGroup’s public release does not provide a full response distribution, a sampling-error range, or a reproducible definition separating “highly prepared” from the lower readiness categories.
Still, its central observation matches the OpenAI evidence: deployment has moved faster than operating discipline. ManpowerGroup says 34% of respondents saw their strongest productivity gains in AI-augmented roles, versus 8% in fully automated roles. That is a useful counterweight to a familiar automation narrative. The near-term management job is not simply deciding which work to remove; it is deciding which human review, skills, approvals and systems must remain attached to AI-assisted work.
For HR, the immediate assignment is to define role boundaries that can bend without becoming unaccountable. A recruiter can use an approved tool to prepare an interview guide, for example, but should not turn a generated assessment into an automated hiring decision. A people-operations team can use AI to find themes in properly handled feedback, but should not upload identifiable employee-relations material into a service lacking contractual and technical safeguards.
That principle travels cleanly to AI rollout. Front-line managers, HR staff and service-desk workers are often the first to identify when a generated answer is wrong, a policy is too vague, or an approved tool is too cumbersome to compete with a public chatbot. If their reports disappear into a survey platform or an annual governance meeting, shadow use will continue and policies will lose credibility.
IT teams already understand this pattern from endpoint management and security reporting. The most useful signal is rarely a polished adoption dashboard; it is the recurring workaround that tells you the official process is failing. Organizations rolling out AI need an equivalent feedback loop: a short path for workers to flag unsafe prompts, bad outputs, missing access, tool gaps and tasks that need specialist review—and a way to show that the report changed something.
OpenAI’s data does not prove that AI is rewriting HR job descriptions. It does show that workers are testing broader task bundles before employers formally redesign roles. The concrete consequence is that AI policy can no longer be written around departments alone. It has to govern the task, the data involved, the tool being used, and the person who remains responsible when the answer crosses from a helpful draft into an action.
The danger is not that an HR generalist asks ChatGPT to draft a flyer or explain a spreadsheet formula. It is that the same tool makes unfamiliar work easy to attempt, which can erase the normal handoff to the person accountable for the result. In HR, where prompts can touch pay, hiring, employee relations, benefits and policy interpretation, that is a governance issue before it is a productivity story.
What OpenAI’s 69% HR number actually measures
OpenAI analyzed more than 800,000 work-related messages from U.S. users whose self-reported role information came from ChatGPT Business accounts. The messages themselves, however, came from those users’ individual ChatGPT accounts, not their ChatGPT Business or Enterprise conversations. OpenAI says it used automated classification without employees reading the underlying messages.That distinction is missing from the headline takeaway. The report is not a measurement of governed enterprise AI use inside a company tenant, and it is not a survey of the U.S. HR workforce. It is a descriptive sample of individual-account activity associated with people who had also provided Business-account role information. OpenAI explicitly says the estimates should not be generalized to Enterprise users or to the entire workforce.
The headline percentage also has a denominator problem that readers should keep in view. Across the full sample, 61.5% of messages were classified as generic: activities shared broadly among occupations, including emails, writing, summarization and scheduling. Of all work-related messages, 16.8% were classified as cross-occupation work. Once generic tasks are removed, the cross-occupation share rises to 43.5% overall and 69% for HR.
In other words, the HR result says that among HR prompts OpenAI considered specific enough to map to a traditional occupational task, more than two-thirds resembled work normally associated with another occupation. It does not establish that HR has suddenly stopped doing HR work, nor that the model performed the work correctly, nor that its output was used in a real business decision.
OpenAI is unusually direct about those limitations. The study does not observe whether the answer was adopted, whether it saved time, whether the user could have completed the task without AI, or whether a qualified specialist reviewed it. It also assigns each message one primary work activity, even though real requests may involve several tasks and roles.
Those caveats do not make the finding irrelevant. They define the finding: ChatGPT is being used at the edge of job boundaries, where the user is attempting work that may previously have gone to another function.
HR’s borrowed tasks include IT, finance and marketing work
OpenAI’s underlying task map shows the types of work moving between functions. Among the eight occupation groups in its study—customer experience, design, engineering, finance, human resources, legal, marketing and sales—financial calculations and troubleshooting computer applications or systems ranked among the three most common outside tasks in every applicable non-specialist group.For HR, OpenAI’s table assigns 18% of non-generic messages to engineering-like tasks, 16% to finance, 23% to marketing, 10% to legal, 9% to customer experience and only 10% to HR itself. Those row shares are a classification of task resemblance, not a claim that HR teams are replacing IT, finance or counsel. But they show why a simple policy saying “AI may be used for drafting” no longer describes what people are doing.
An HR practitioner may use ChatGPT to troubleshoot a spreadsheet, interpret a technical system issue, calculate compensation scenarios, prepare communications material or understand a regulation. Each may begin as a sensible request for assistance. The operational question is what happens after the draft or explanation arrives: Does the output move into a payroll model? Does it shape a benefits answer? Does an HR business partner make a policy call based on a model’s summary? Does a manager act on a technical recommendation that bypasses the service desk or security team?
The model can reduce the friction that once forced a handoff. It cannot transfer responsibility for the decision. That is the line organizations need to make visible.
For Windows administrators and IT leaders, this reinforces the need to treat AI access as an application-governance issue rather than an employee training perk. A company cannot meaningfully govern sensitive HR use if it only licenses an approved assistant while employees continue putting work into personal accounts, browser sessions or unmanaged AI tools. OpenAI’s research itself captures individual-account behavior linked to Business users, a reminder that the boundary between sanctioned and unsanctioned usage is often porous.
The immediate controls should be mundane but specific: approved tools, clear data classifications, identity-backed access, retention rules, auditability, and an escalation path when an AI-assisted task crosses into payroll, legal interpretation, security troubleshooting or an employment decision. “Use good judgment” is not a control when the point of the tool is to let people act outside their established expertise.
The 3% leadership-readiness claim is a warning, not a workforce census
ManpowerGroup Talent Solutions released its New Talent Equation: Activating Workforce Confidence at Scale report on July 22 and said only 3% of organizations surveyed considered their leaders highly prepared to manage AI-enabled ways of working. The same release says 78% reported employee concern about AI’s effect on jobs, 63% saw resistance after deployment, and 86% put AI-oriented upskilling and reskilling among priorities for the next 12 to 18 months.The 3% finding is striking, but its evidentiary base is small. The research drew on 80 C-suite, CHRO and senior talent-acquisition leaders in the United States and United Kingdom across healthcare, life sciences, manufacturing and technology. ManpowerGroup commissioned the report, which Everest Group developed. A percentage of 3% in a sample of 80 equates to roughly two or three respondents, depending on rounding.
That does not invalidate the survey’s direction of travel. It does mean the figure should not be presented as a precise estimate of leadership readiness across all employers. ManpowerGroup’s public release does not provide a full response distribution, a sampling-error range, or a reproducible definition separating “highly prepared” from the lower readiness categories.
Still, its central observation matches the OpenAI evidence: deployment has moved faster than operating discipline. ManpowerGroup says 34% of respondents saw their strongest productivity gains in AI-augmented roles, versus 8% in fully automated roles. That is a useful counterweight to a familiar automation narrative. The near-term management job is not simply deciding which work to remove; it is deciding which human review, skills, approvals and systems must remain attached to AI-assisted work.
For HR, the immediate assignment is to define role boundaries that can bend without becoming unaccountable. A recruiter can use an approved tool to prepare an interview guide, for example, but should not turn a generated assessment into an automated hiring decision. A people-operations team can use AI to find themes in properly handled feedback, but should not upload identifiable employee-relations material into a service lacking contractual and technical safeguards.
Front-line feedback must result in visible action
HR Dive’s “quote of the week” came from Gerald Danniel, chief strategy officer and executive vice president at Onvo, who described a workplace where front-line employees can speak directly with owners and leaders rather than “jump through hoops.” The accompanying HR Dive reporting stresses the part companies frequently miss: collecting feedback is not enough. Employees need to see what happened to it.That principle travels cleanly to AI rollout. Front-line managers, HR staff and service-desk workers are often the first to identify when a generated answer is wrong, a policy is too vague, or an approved tool is too cumbersome to compete with a public chatbot. If their reports disappear into a survey platform or an annual governance meeting, shadow use will continue and policies will lose credibility.
IT teams already understand this pattern from endpoint management and security reporting. The most useful signal is rarely a polished adoption dashboard; it is the recurring workaround that tells you the official process is failing. Organizations rolling out AI need an equivalent feedback loop: a short path for workers to flag unsafe prompts, bad outputs, missing access, tool gaps and tasks that need specialist review—and a way to show that the report changed something.
OpenAI’s data does not prove that AI is rewriting HR job descriptions. It does show that workers are testing broader task bundles before employers formally redesign roles. The concrete consequence is that AI policy can no longer be written around departments alone. It has to govern the task, the data involved, the tool being used, and the person who remains responsible when the answer crosses from a helpful draft into an action.
References
- Primary source: HR Dive
Published: 2026-08-03T11:13:00+00:00
Loading…
www.hrdive.com - Related coverage: axios.com
Loading…
www.axios.com - Related coverage: openai.com
Loading…
openai.com - Related coverage: openai.com
ChatGPT usage and adoption patterns at work | OpenAI
A data-driven report on how workers across industries use ChatGPT—covering adoption trends, top tasks, departmental patterns, and the future of AI at work.openai.com - Related coverage: cdn.openai.com
- Related coverage: itpro.com
Loading…
www.itpro.com