The question your board will ask before you're ready
Three weeks before the board meeting, the request lands in your inbox: the PE sponsor wants to know how many jobs will AI replace by 2030, and specifically, which of your roles are on that list. You pull up the headlines. One says AI could affect the equivalent of 300 million full-time jobs globally. Another says 60 to 70 percent of work hours are technically automatable. A third says the net effect is actually positive — tens of millions more jobs created than lost. None of these numbers were built to answer your board's question, and presenting any one of them as if it applies to your organization will not survive a follow-up question.
This article does not resolve the disagreement between the big 2030 estimates, because it can't be resolved — the estimates aren't measuring the same thing. What it does is explain what each major figure actually captures, why the ranges are so wide, and what a defensible, company-specific answer requires instead of a single borrowed headline.
Why the 2030 estimates don't agree
The estimates that get quoted in board decks and news coverage come from at least four different kinds of research, each answering a different question.
Some measure technical automation potential — what a technology could theoretically do to a task, independent of whether anyone deploys it that way. McKinsey's 2023 analysis falls here: it estimates that 60 to 70 percent of work hours could theoretically be automated with current and emerging technology, up from roughly 50 percent in pre-generative-AI estimates. That is a ceiling, not a forecast of what will happen by any particular date.
Some measure exposure, meaning how much of an occupation's task content overlaps with what a technology can currently do, without claiming the task will actually be automated. Goldman Sachs's widely cited estimate that the equivalent of 300 million full-time jobs globally are exposed to automation by generative AI is this kind of figure — a measure of overlap, not a prediction of headcount loss.
Some measure net labor market churn — jobs created and destroyed simultaneously, which is the normal way labor markets behave even without AI. The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced by 2030, a 22 percent churn rate, for a net gain of 78 million jobs globally. That is not a "78 million jobs are safe" statement; it is a structural churn estimate at the level of the entire global economy.
And some measure employer sentiment — what executives say they intend to do, which is a leading indicator, not a technical assessment. PwC's Global CEO Survey found that in 2024, 25 percent of CEOs expected a headcount reduction of 5 percent or more attributable to AI, while 39 percent expected an increase of 5 percent or more. By 2025, PwC found 13 percent of CEOs reported an actual headcount reduction due to generative AI (16 percent in the most AI-exposed sectors), against 17 percent who saw an increase. These are surveys of intent and self-reported outcome, not measurements of task automatability.
Put a technical-potential number, an exposure number, a churn number, and a sentiment number side by side and of course they disagree — they were never trying to agree.
What each estimate is actually measuring
It helps to be specific about what sits underneath the headline you're quoting, because the underlying methodology tells you what the number can and can't support.
Anthropic's Economic Index, tracking real usage of its Claude models, found that the share of occupations using the tool for at least a quarter of their tasks grew from 36 percent in early 2025 to 49 percent on a pooled basis later that year — a usage-intensity measure, not an exposure or displacement estimate. A companion 2026 analysis found that 68 percent of that usage fell into tasks the model could fully complete, while only 3 percent involved tasks it could not meaningfully touch at all. That's a granular, task-level read on capability — closer to how a rigorous workforce assessment should be built, but still describing a model's technical reach, not a company's headcount decision.
The World Economic Forum's Future of Jobs Report 2025 projects 170 million jobs created and 92 million displaced worldwide by 2030 — a 22 percent churn rate, and a net gain of 78 million jobs.
The WEF's report also found that 39 percent of workers' current skills are expected to be transformed or become outdated by 2030, and that of every 100 workers, 59 will need retraining of some kind — 29 upskilled within their current role, 19 reskilled or redeployed internally, and 11 unlikely to receive adequate training at all. Separately, 63 percent of employers named skill gaps as their single biggest barrier to workforce transformation. These are the numbers that matter most for a People Analytics function, because they describe a training and redeployment problem your organization can actually act on — unlike a global exposure headline, which cannot tell you anything about your own org chart.
Brookings' 2024 analysis narrows the picture further, finding that the highest-exposure occupational categories cluster in STEM, business and finance, engineering, and law, and that roughly 12.9 million workers — about a third of those in the most exposed occupations — are highly exposed. That's sector-level granularity, a step closer to something a company can use, but it is still an occupational-category estimate, not a role-by-role assessment of your own workforce.
How many jobs will AI replace by 2030, read correctly
So — how many jobs will AI replace by 2030? The honest answer is that no credible source claims to know, because "replace" implies an outcome (a specific job eliminated) while nearly every rigorous estimate measures something upstream of that: technical potential, task overlap, usage intensity, or stated intent. The WEF's own headline number is not a job-loss figure at all — it's a net gain, once churn is accounted for. Read literally, the most defensible summary of the current research is that a large share of work — by some estimates a majority of work hours — has AI-relevant task content, that tens of millions of jobs globally will be displaced while a larger number are created, and that the resulting churn will demand retraining at a scale most employers are not currently prepared for.
None of that tells your board which of your 340 roles are Monitor, Review, or Redeployment Candidate risk categories. That answer only exists at task level, inside your own organization's occupational mix — which is exactly the gap between macro research and a workforce plan.
For a deeper breakdown of the individual statistics behind these headlines, see our companion piece on AI job displacement statistics, and for a closer look at which occupational categories the research consistently flags as highest-exposure, see which jobs are most exposed to AI. If you want the McKinsey findings unpacked in more depth, including the shift toward healthcare, STEM, and managerial employment mix by 2030, we've summarized it in our McKinsey AI workforce report summary.
Occupation-level headlines vs. task-level reality
This is where the ONET framework earns its keep. The U.S. Department of Labor's ONET database — 1,016 occupational titles covering 923 data-level occupations and more than 55,000 real-world job titles, described by roughly 277 standardized descriptors and updated on a regular cycle with a primary annual refresh — was not built to answer "how many jobs will AI replace." It was built to describe, in granular and comparable terms, what any given occupation's work actually consists of: tasks, skills, knowledge areas, and their relative importance to the role.
That's the level at which an AI exposure assessment becomes useful rather than merely provocative. WorkforceAnalysis scores each task attached to a role against four dimensions — cognitive routine, physical routine, social/judgment, and creative — each rated 0 to 100, then weights those scores by O*NET's own task-importance ratings for that occupation. As a simplified worked example: if a role's top three tasks are weighted 40 percent, 35 percent, and 25 percent of total importance, and those tasks score 80, 65, and 30 on cognitive-routine exposure respectively, the role's weighted cognitive-routine exposure comes out to roughly 66 (0.40×80 + 0.35×65 + 0.25×30) — a single, defensible number built from the same task list the U.S. Department of Labor uses to describe that occupation nationally.
That score is an input to a Monitor, Review, or Redeployment Candidate designation — never a verdict, and never a claim that a specific role "will be automated." The distinction matters more than it sounds: a macro estimate can tell your board that engineering and finance functions are, in the aggregate, among the more exposed occupational categories. It cannot tell you whether your specific senior financial analyst role, with its particular mix of reconciliation tasks and stakeholder judgment calls, sits at 30 or 70 on that scale. Only a task-level assessment of your own role list can do that.
This site incorporates information from O*NET. Used under the CC BY 4.0 license. O*NET is a trademark of USDOL/ETA.
What this means for your next planning cycle
Gartner's 2024 research found that 86 percent of HR leaders have not implemented a strategic workforce planning capability, and that 66 percent describe their current workforce planning as limited to headcount forecasting, with difficulty demonstrating its return. That gap explains why so many teams reach for a macro headline in the first place — it's available, it's citable, and it feels like an answer. But a Goldman Sachs exposure estimate or a McKinsey automation-potential range was never built to survive a board asking "what about our engineering function specifically." For that, you need your own role list, your own O*NET mapping, and your own four-dimension scoring — read alongside, not in place of, the macro research summarized here and in how AI is changing the future of work.
Where to start
If your next step is building that company-specific picture rather than repeating someone else's global estimate, our pricing page outlines the self-serve tiers for mapping a role list to O*NET occupations and generating department-level exposure heatmaps. And if you'd rather work through the framework at your own pace first, our Workforce AI Readiness Assessment Guide walks through the same four-dimension methodology in a self-paced format.
Subscribe to our newsletter for a monthly digest of new macro research as it's published, translated into what it does and doesn't mean for a mid-market workforce plan — no borrowed headlines, no invented numbers.