AI and Jobs in 2026: What the Data Actually Says

The major studies on AI and jobs look like they are shouting past each other, but they are not actually in conflict. They measure different units, task exposure, job automation, and net job change, over very different time horizons, from today out to 2030, and as far as 2060. Once you line up the unit being counted and the timeline it applies to, the contradictions dissolve and a coherent, actionable picture appears. Large shares of work are exposed at the task level, the way that exposure plays out splits between augmentation and automation rather than wholesale elimination, and the net macro outlook is positive while the disruption underneath it is real, uneven, and concentrated.

A worker or an HR leader does not need another headline percentage. What is useful is understanding why the numbers differ, what the macro picture means, and how to translate any of it into a reading for one specific occupation. That is the work this piece does: define the three measurement lenses, walk each major study at its true horizon, retire the 2013 figure that still haunts consumer calculators, and then move from the aggregate to a single role.

Exposure, automation, augmentation: why does the word change the number?

Most of the apparent disagreement between studies is really a disagreement about which of three different things is being counted. Get the words straight and the numbers stop fighting.

Task-level exposure describes a task that an AI tool could touch or speed up. This is the broadest and biggest number, and it says nothing about whether a job disappears. When OpenAI reports that 80 percent of workers have tasks exposed, it is measuring this and only this.

Automation means AI performs a task with little human involvement. Augmentation means AI assists a person who stays in the loop. The same exposed task can resolve to either outcome depending on how the work is organized, what the tool can reliably do, and what the employer decides. Exposure is the upstream condition; automation and augmentation are the two downstream paths.

There is a second axis that drives the disagreement just as hard: technical potential versus realized adoption. Technical potential is what is theoretically possible with today’s tools. Realized adoption is what organizations have actually deployed at scale. The two can sit a decade or more apart, held back by cost, integration work, regulation, and the slow grind of organizational change. A study that measures technical potential and a study that measures deployed reality will produce very different numbers about the same technology, and both can be correct.

What does the WEF Future of Jobs Report 2025 actually project by 2030?

The World Economic Forum answers the net job change question, and it answers it to 2030. By that horizon the WEF projects 170 million new jobs created and 92 million displaced, a net increase of 78 million jobs (World Economic Forum, Future of Jobs Report 2025, 2025). The headline is growth, but the displacement underneath it is large at the same time.

Those two figures together represent roughly 22 percent structural churn of the formal-jobs base the WEF studied by 2030 (WEF, 2025). Growth and disruption are not opposites here; they are happening simultaneously, in different roles and different places.

On the skills side, the WEF expects 39 percent of workers’ core skills to change or become outdated between 2025 and 2030, and 59 of every 100 workers to need reskilling or upskilling by 2030 (WEF, 2025). Of those 100 workers, 11 are unlikely to receive the training they need, a group the WEF puts at over 120 million workers at medium-term risk of redundancy (WEF, 2025). That is where the disruption concentrates: not in a collapse of total employment, but in specific people whose skills shift faster than their support does.

Read it this way and the WEF stops sounding contradictory. It says total employment grows and specific roles face serious disruption, both to 2030. Those are two answers to two different questions, not two sides of an argument.

How much work could AI automate, and on what timeline?

McKinsey answers a different question again: not how many jobs change, but how much work activity could technically be automated, and over what horizon. The Global Institute estimates that generative AI could add the equivalent of 2.6 trillion to 4.4 trillion US dollars in value annually across the 63 use cases it analyzed (McKinsey Global Institute, The economic potential of generative AI, 2023). That frames AI primarily as a productivity and value engine.

On automation, McKinsey finds that current generative AI and related technologies could technically automate work activities that absorb 60 to 70 percent of employees’ time today (McKinsey, 2023). The word that carries the weight is technically. This is potential, not deployment.

The horizon is where McKinsey and the WEF stop looking like rivals. McKinsey’s adoption scenarios place the automation of half of today’s work activities somewhere between 2030 and 2060, with a midpoint of 2045 (McKinsey, 2023). That is a far later horizon than the WEF’s 2030, which is precisely why both can be right at once. The same study showing a 60 to 70 percent technical figure and a 2045 midpoint is showing you the realized-adoption gap directly: what is possible now leads what is in place by years or decades.

What did the OpenAI, Anthropic, and Microsoft task-exposure studies find?

These three measure the present, at the task level, and their findings are large but specific. OpenAI researchers estimate that about 80 percent of the US workforce could have at least 10 percent of their work tasks affected by large language models, and about 19 percent could see at least 50 percent of tasks affected (Eloundou, Manning, Mishkin and Rock, GPTs are GPTs, 2023). The same study finds that about 15 percent of all worker tasks could be done significantly faster at equal quality with LLM access alone, rising to 47 to 56 percent once LLM-powered software is included (OpenAI, 2023). This is exposure, not job loss.

Anthropic’s Economic Index adds something the others cannot: what people actually do with the tools. Its first report found that AI use on Claude.ai leaned toward augmentation at 57 percent versus 43 percent automation, mapping real conversations to O*NET tasks (Anthropic Economic Index, first report, 2025). By the September 2025 report the consumer split had tipped the other way: for the first time, automation usage on Claude.ai surpassed augmentation, settling at roughly even with automation slightly ahead, near 51 percent automation and 49 percent augmentation, while enterprise API use ran around 77 percent automation (Anthropic Economic Index, September 2025 report, 2025). The mix is not fixed: it shifts by context and over time, which is exactly why a single automation-versus-augmentation number for “AI” is misleading.

Microsoft Research approached it from yet another angle, computing an AI applicability score per occupation from real Bing Copilot conversations. It found the highest applicability for knowledge work, including computer and mathematical roles, office and administrative support, and sales, and the lowest for healthcare support and physical-labor roles (Microsoft Research, Working with AI, 2025). Crucially, Microsoft attached an explicit warning to its own number.

Where does the 47 percent figure come from, and why is it outdated?

The single number that still dominates consumer “will AI take my job” tools predates large language models entirely. Carl Benedikt Frey and Michael Osborne estimated that about 47 percent of total US employment was at risk of computerisation, using a model over 702 detailed occupations (Frey and Osborne, The Future of Employment, Oxford Martin School, 2013). It was a serious academic exercise for its moment. Its moment was thirteen years ago.

More than a decade later, many calculators still recycle it, attaching a 2013 probability to a 2026 worker. The problem is not only the age of the model but the result it predicted. The forecast job loss did not materialize: an Information Technology and Innovation Foundation analysis notes that US employment grew rather than collapsed in the years that followed, which undercuts the 47 percent as a basis for present-day per-occupation risk (Information Technology and Innovation Foundation, 2022).

Using a single dated probability gives a specific worker a false reading. It was never built to be queried per role today, and it predates the technology people are actually worried about.

Is exposure the same as your job being eliminated?

No. Exposure means an AI tool could touch some of your tasks. Elimination means the role goes away. The studies that measure exposure are explicit that they are not measuring job loss, and reading one as the other is the most common mistake in this whole conversation.

Put the threads together and a consistent near-term pattern emerges. OpenAI’s 80 percent exposed at the task level, Anthropic’s first report showing augmentation leading automation on Claude.ai at 57 versus 43 percent, and Microsoft’s explicit caveat all point the same way: task reshaping, not wholesale replacement, dominates the near term (OpenAI, 2023; Anthropic Economic Index, first report, 2025; Microsoft Research, 2025). The mix is moving, not static. By Anthropic’s September 2025 report, automation on Claude.ai had edged past augmentation for the first time, while enterprise API use leaned far more heavily toward automation (Anthropic Economic Index, September 2025 report, 2025). The operative point is not which side leads in a given quarter; it is that exposure resolves into augmentation and automation in different proportions by context and over time, so it cannot be read as a single verdict.

That said, the real risk should not be waved away either. The WEF’s 11 of every 100 workers unlikely to be reskilled, and the concentration of high exposure in specific knowledge-work occupations, mean the disruption is genuine, uneven, and concentrated even when the net is positive (WEF, 2025). The honest position holds both halves at once.

A high task-exposure score is best read as a planning signal. It tells you which parts of a role to redesign, augment, or retrain around. It does not tell you the role is doomed.

Will AI create or destroy more jobs on balance?

On the best current macro evidence, the net is positive. The WEF’s net increase of 78 million jobs by 2030 is the clearest top-line number available, and it sits alongside McKinsey’s framing of generative AI as a productivity and value engine rather than a pure displacement force (WEF Future of Jobs Report 2025; McKinsey Global Institute, 2023).

Net positive does not mean evenly distributed. The same WEF report shows 92 million jobs displaced and 39 percent of core skills changing by 2030, so individual occupations, regions, and demographics can face sharp disruption inside a growing aggregate (WEF, 2025). The national average does not arrive evenly in any one career.

This is why the macro question and the personal question must be kept apart. The macro question asks whether total employment grows. The personal question asks what happens to your specific role. A national net number, however well sourced, simply cannot answer the personal one. That gap is the missing piece, and it points to what is actually needed: a per-occupation measurement.

How do you go from a macro headline to one specific occupation?

The studies hand you horizons and aggregates. What a person or an employer can act on is a single role, assessed against all of them at once. That is the pivot from reconciliation to action.

This is where JobRoute operates, as a measurement engine rather than a competing headline statistic. It reconciles the scattered literature into one reproducible, version-locked, source-traceable exposure score for any of 1,016 ONET occupations, and it separates task exposure from job elimination the way Microsoft Research itself urges. The data graph is named in the open: ONET 30.2, ESCO, Lightcast Open Skills, the Anthropic Economic Index, the WEF Future of Jobs Report, and BLS, locked to specific versions so a score is reproducible and citable rather than a black box.

In practice that means you can check your own role against the reconciled sources, read how JobRoute measures exposure and how the score works, and Take the AI Ready Score for free. The same engine serves three audiences from one data graph: individuals get the free AI Ready Score, enterprises get workforce exposure mapping for enterprises, and government gets WIOA-fundable workforce program measurement.

What does a reconciled per-occupation score add that a headline cannot?

The differentiator is narrow and defensible. Consumer calculators recycle one dated 2013 number. The big studies publish episodic reports at incompatible horizons. Enterprise tools are proprietary black boxes. The value is in reconciling all of them and doing it reproducibly, per occupation, rather than publishing yet another scare statistic.

A reconciled score also keeps the augmentation-versus-automation distinction intact rather than collapsing it into a single frightening figure, mirroring exactly what Anthropic’s shifting usage data and Microsoft’s caveat show. That distinction is the difference between knowing a role will be redesigned and fearing it will vanish.

Exposure is the start of a plan, not the end of a career. The right next step is to measure a specific role, not to read one more headline.

You can take the free AI Ready Score for any role today, and read the full formula, sources, and version-locking in the methodology.