Human skills in an AI economy: what keeps its value and how to build it

What the exposure studies measure, and what they do not

Every serious estimate of AI's effect on work starts with an exposure index: a score for how much of an occupation's task list a model could, in principle, perform. The International Labour Organization's 2025 refined index places about one in four workers worldwide in an occupation with some generative-AI exposure, 34% of employment in high-income countries against 11% in low-income ones; only 3.3% of global employment falls into the highest-exposure tier, and women's employment is roughly twice as exposed as men's (4.7% versus 2.4%).^[1] A Eloundou and colleagues estimate that about 80% of the US workforce could have at least 10% of their tasks affected by large language models and about 19% could have half or more.^[2] A An OECD summary, which we could read only as a search snippet and therefore grade C, puts around a quarter of OECD workers in exposed occupations and about 1% in highly exposed ones.^[3] C

Three limits apply to all of them. Exposure is potential, not adoption. It is scored on task descriptions, not on what the task is worth or who is accountable for it. And it says nothing about the tasks an occupation gains when others are automated. The question people actually type, "will ai replace my job", draws about 210 US searches a month and is up 91% year on year in our keyword data (DataForSEO, retrieved 2026-11-09); our role-by-role answer works through the exposure research for specific jobs. This page asks the complementary question: which human skills gain value as the exposed tasks get cheaper?

Skills that complement AI in the data

The labour-market record offers one long-run study and several recent experiments.

The long-run study is Deming's analysis of US employment from 1980 to 2012. Jobs requiring high levels of social interaction grew by nearly 12 percentage points as a share of the labour force, maths-intensive but less social jobs shrank by about 3 points, and employment and wage growth were strongest in jobs demanding both high maths and high social skill.^[4] A His explanation is that computers substitute for routine cognitive tasks but not for the flexible, interpersonal coordination that lets a team reassign work among its members. The mechanism predates generative AI, but it is the same mechanism.

The experiments show where human skill still earns its keep once AI is in the room:

• In a quasi-experiment across 5,172 customer-support agents, a generative-AI assistant raised issues resolved per hour by about 14–15% on average, over 30% for the least experienced and close to zero for the most experienced.^[5] A The tool encoded what experienced agents already knew; the experts' edge did not disappear, it became the thing being copied.
• In a field experiment with 758 consultants, GPT-4 users did better on tasks inside the model's competence and were 19 percentage points less likely to be correct on a task designed to sit just outside it.^[6] B Knowing which side of that line you are on is a human skill with a measurable payoff.
• A meta-analysis of 106 experiments found human–AI combinations under-performing the better of the two alone on decision tasks, while gaining on content-creation tasks.^[7] A Decisions remain the human's job not because models cannot make them, but because combining human and model judgement badly is worse than either.

Autor's reading of this evidence is optimistic in a specific way: AI could extend expert-level tasks to more workers and rebuild middle-skill work, but only if the people using it hold the judgement to direct it.^[8] B The World Economic Forum's employer survey expects 39% of current skills to be transformed or obsolete by 2030,^[9] B a churn figure that applies to the human skills below as much as to technical ones: they do not vanish, they are redeployed.

From this evidence, four families of skill complement AI in the data: judgement and verification, taste and framing, trust and coordination, and care. Each gets a section, with the evidence that ties it to complementarity rather than to a wish.

Judgement and verification

Judgement is the skill of deciding what is true enough and good enough to act on. Verification is its routine form: checking before trusting.

The research on automated aids is older than language models and consistent. A systematic review of 74 studies of clinical decision support found users following incorrect advice at non-trivial rates, mediated by trust, task complexity and time pressure.^[10] A An integrative review in Human Factors describes complacency and automation bias as attentional effects of trust and workload that fall with accountability and training, not as traits of careless people.^[11] A And the METR study of 16 experienced open-source developers found them 19% slower when allowed to use early-2025 AI tools while believing they were about 20% faster, a small sample with a large lesson about calibration.^[12] B

What complementarity looks like: as the production of plausible output gets cheap, the ability to tell plausible from correct gets scarce. That skill is domain knowledge plus a habit. The habit is sampling three claims from every output you rely on and tracing each to a source; the domain knowledge is what lets you see that the fourth claim, the one you did not sample, is the kind that goes wrong.

Taste and framing

Taste is knowing what good looks like before it exists. Framing is deciding what question is worth answering.

The writing experiment is the clearest evidence that these move up in value. ChatGPT cut time on mid-level professional writing tasks by about 40%, raised rated quality by about 18% and, crucially, compressed the gap between weaker and stronger writers.^[13] A When average quality rises and converges, the differentiator stops being execution and becomes the choice of what to make and the recognition of when a draft is merely competent. The meta-analysis makes the same point from the other side: content creation is where human–AI teams gained, which is where a human with taste directing a fast producer beats either alone.^[7]

Framing matters earlier in the chain. A brief that asks for the wrong thing is executed perfectly and uselessly; most of the "AI produced rubbish" stories in any team are framing failures with a technical excuse. The skill is learned by writing the problem statement before the prompt, and by reviewing a month of outputs you discarded to see how many were answers to badly posed questions.

Trust, coordination and care

Deming's twelve points of growth were in jobs where people coordinate flexibly with other people.^[4] Three conditions make that skill scarcer, not cheaper, over the next decade.

First, coordination load is rising. Microsoft's 2025 Work Trend Index, a vendor survey of 31,000 workers, counts an interruption every two minutes, about 275 a day.^[14] A Somebody has to decide what deserves a response, and nobody wants an agent making that call about their relationships.

Second, distance is permanent. Eurostat records 22.6% of EU employees usually or sometimes working from home in 2024, with France at 33.9%, and the US time-use survey records 33% of employed people doing some work at home on days worked, 50% among degree-holders.^{[15] [16]} A Trust built across screens is a skill, and hybrid teams need more of it per unit of output, not less.

Third, care is the least exposed tier. The ILO index places the highest exposure in clerical work and the lowest in occupations where the task is the relationship: teaching, nursing, counselling, management of people.^[1] Care is not soft; it is the work of noticing what someone needs before they say it, and giving feedback a person can hear. Nothing in the controlled literature suggests a model substitutes for being remembered as the person who handled a hard conversation well.

How to build each deliberately

These skills are built the way any skill is built: by a defined practice, a measure and a review. The table is ours; the evidence is cited where a practice derives from it.

Skill	Weekly practice	Measure	Evidence anchor
Verification	Sample and trace three claims in every AI output you rely on; log the error rate	Errors found per 10 outputs, trending down for tasks you keep, up for tasks you drop	Automation bias falls with accountability[10] [11]
Frontier judgement	Keep a two-column log of tasks where AI helped and where you corrected it	Share of delegated tasks that pass without correction	Jagged frontier[6]
Taste	Review the outputs you discarded this month; name why each fell short	A written quality rubric that gets more specific over time	Quality compression in the writing experiment[13]
Framing	Write the problem statement before any brief; revisit it when the output disappoints	Share of disappointing outputs traced to the question rather than the tool	Decision-task losses in combinations[7]
Coordination	Run one meeting a week with no AI in the room, with a written decision and owner at the end	Decisions that stuck a month later	Social-skill premium[4]
Care	One piece of specific, in-person feedback a week; one check-in with no agenda	Whether people bring you problems early	Low exposure of relational work[1]

Pick two, not six. Verification and framing are the pair with the fastest payoff for most knowledge workers, because they change the quality of everything else you delegate.

Our forecast, kept separate from the measurements

Everything above is measured or experimentally observed, with its grade. What follows is a model output. In our Future Demand Radar, the category "future-of-work skills and AI literacy" scores 45.8 out of 100 in the baseline scenario (p10–p90 band 41.4–50.2), the "monitor" class; the conservative scenario gives 42.5 ("monitor") and the accelerated scenario 48.2 ("validate in 12 months"). Our baseline scenario expects US search demand for the category to reach about 2.2 times today's level by 2030 (band 1.7–2.8) and 2.8 times by 2035. "Skills for the future" itself draws about 320 US searches a month, growing 30% a year over three years (retrieved 2026-11-09). The three scenarios for knowledge work to 2035 describe what each path means for a career; under scenario C, where capable agents absorb much routine knowledge work, individual productivity reduces to judgement, orchestration and recovery, which is to say to the skills on this page. Forecasts are model outputs with stated uncertainty, not facts.

What to do next

The practices above need protected time, and the OS Builder is where that time gets scheduled: it turns thirteen answers about your work into a daily rhythm with deep-work blocks, a weekly review and an AI-delegation section that already separates what to delegate from what to keep. Build the system, then put two of the practices from the table into the weekly review. For the technical side of the same question, which AI skills to learn and in what order, read AI skills for 2030; the two pages are designed as a pair.

Frequently asked questions

Which human skills become more valuable as AI improves?

The evidence points to four families: judgement and verification (telling correct from plausible), taste and framing (knowing what good looks like and what question is worth asking), trust and coordination (the social skills whose share of US employment grew nearly 12 points from 1980 to 2012^[4] A), and care, the relational work that exposure indices place in the lowest tier.^[1]

Does AI exposure mean my job disappears?

No. Exposure measures how much of an occupation's task list a model could in principle perform. The ILO places 25% of global employment in exposed occupations but only 3.3% in the highest tier, and frames the effect as augmentation for most and transformation for few.^[1] A Our page on whether AI replaces your job walks through the research by role.

Why do experts gain less from AI than beginners?

Because the tool encodes what experts already do. In a study of 5,172 support agents, novices gained over 30% in issues resolved per hour and the most experienced close to zero.^[5] A The experts' advantage did not vanish; it moved to selecting tasks, checking outputs and handling the cases the tool gets wrong.

Is 'soft skills' just a consolation prize?

The wage data says otherwise. Deming found employment and wage growth strongest in jobs requiring both high maths and high social skill, and the social-skill share of employment rising while routine cognitive work fell.^[4] A Coordination is what lets a team reassign work flexibly, and that is exactly the task automation handles worst.

How do I practise judgement without a course?

Sample three claims from every AI output you rely on, trace them to a source and log the error rate; keep a two-column list of tasks where AI helped and where you corrected it. Research on automation bias shows checking habits and accountability reduce the errors people make with automated aids.^{[10] [11]} A

Sources and evidence

1. Generative AI and Jobs: A Refined Global Index of Occupational Exposure (ILO Working Paper 140) — International Labour Organization (2025), report A www.ilo.org/publications/generative-ai-and-jobs-refined-global-index-o (retrieved 2026-11-09)
2. GPTs are GPTs: Labor market impact potential of LLMs — Eloundou, Tyna; Manning, Sam; Mishkin, Pamela; Rock, Daniel (2024), observational A doi.org/10.1126/science.adj0998 (retrieved 2026-11-09)
3. Who will be the workers most affected by AI? (OECD AI Papers No. 26; figure read from a search snippet only) — OECD (2024), report C www.oecd.org/content/dam/oecd/en/publications/reports/2024/10/who-will (retrieved 2026-11-09)
4. The Growing Importance of Social Skills in the Labor Market — Deming, David J. (2017), observational A doi.org/10.1093/qje/qjx022 (retrieved 2026-11-09)
5. Generative AI at Work — Brynjolfsson, Erik; Li, Danielle; Raymond, Lindsey (2025), observational A doi.org/10.1093/qje/qjae044 (retrieved 2026-11-09)
6. Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality — Dell'Acqua, Fabrizio; McFowland III, Edward; Mollick, Ethan R.; Lifshitz-Assaf, Hila; Kellogg, Katherine; Rajendran, Saran; Krayer, Lisa; Candelon, François (2023), RCT B doi.org/10.2139/ssrn.4573321 (retrieved 2026-11-09)
7. When combinations of humans and AI are useful: A systematic review and meta-analysis — Vaccaro, Michelle; Almaatouq, Abdullah; Malone, Thomas (2024), meta-analysis A doi.org/10.1038/s41562-024-02024-1 (retrieved 2026-11-09)
8. Applying AI to Rebuild Middle Class Jobs (NBER Working Paper 32140) — Autor, David (2024), report B doi.org/10.3386/w32140 (retrieved 2026-11-09)
9. Jobs market at a crossroads: which are the fastest growing and declining jobs? (summarising WEF Future of Jobs 2025) — Euronews (2025), report B www.euronews.com/business/2025/02/01/jobs-market-at-a-crossroads-which (retrieved 2026-11-09)
10. Automation bias: a systematic review of frequency, effect mediators, and mitigators — Goddard, Kate; Roudsari, Abdul; Wyatt, Jeremy C. (2012), systematic review A doi.org/10.1136/amiajnl-2011-000089 (retrieved 2026-11-09)
11. Complacency and Bias in Human Use of Automation: An Attentional Integration — Parasuraman, Raja; Manzey, Dietrich H. (2010), systematic review A doi.org/10.1177/0018720810376055 (retrieved 2026-11-09)
12. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity — Becker, Joel; Rush, Nate; Barnes, Elizabeth; Rein, David (METR) (2025), RCT B doi.org/10.48550/arXiv.2507.09089 (retrieved 2026-11-09)
13. Experimental evidence on the productivity effects of generative artificial intelligence — Noy, Shakked; Zhang, Whitney (2023), RCT A doi.org/10.1126/science.adh2586 (retrieved 2026-11-09)
14. 2025 Work Trend Index Annual Report: The year the Frontier Firm is born — Microsoft WorkLab (2025), report B www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-fro (retrieved 2026-11-09)
15. Employed persons working from home by professional status (lfsa_ehomp) — Eurostat (2024), dataset A ec.europa.eu/eurostat/databrowser/view/lfsa_ehomp/default/table?lang=e (retrieved 2026-11-09)
16. American Time Use Survey – 2024 Results (USDL-25-1060) — US Bureau of Labor Statistics (2025), dataset A www.bls.gov/news.release/archives/atus_06262025.htm (retrieved 2026-11-09)