The figure that opens almost every article on soft skills is this one: 85% of professional success depends on people skills, only 15% on technical skills. The source cited is a Carnegie Foundation study, sometimes attributed to Stanford or Harvard, dated 1918.
That study, in the form it gets cited, has never been found. Several attempts to trace the primary source stop at circular citations: articles citing articles citing 1980s training books. There are real works by Charles Riborg Mann for the Carnegie Foundation on engineering education in those years, but they don’t contain that percentage breakdown, and in any case applying them to today’s labour market would be odd.
The curious thing is that inventing it wasn’t necessary. The real data exists, it’s recent, and it says something more interesting than an 85/15 ratio.
What actually predicts performance (and the 2022 revision)
For twenty years the most cited table in personnel selection was Frank Schmidt and John Hunter’s, published in Psychological Bulletin in 1998 as a synthesis of 85 years of research. It said general cognitive ability was the strongest single predictor of job performance, with a corrected validity around 0.51, followed by work samples and structured interviews.
In 2022 Paul Sackett, Charlene Zhang, Christopher Berry and Filip Lievens published a revision of those estimates in the Journal of Applied Psychology. The technical argument is that the range-restriction corrections applied in the original studies were systematically too generous. With more conservative corrections the picture changes substantially:
| Predictor | 1998 estimate | 2022 revised estimate |
|---|---|---|
| General cognitive ability | ~0.51 | ~0.31 |
| Structured interview | ~0.51 | ~0.42 (the highest) |
| Job knowledge test | ~0.48 | ~0.40 |
| Work sample | ~0.54 | ~0.33 |
| Conscientiousness | ~0.31 | ~0.19 |
Two readings. First: the best single predictor today isn’t a cognitive test, it’s a well-run interview, meaning structured, with identical questions for everyone and a scoring rubric. Second, and more important for this article: no single predictor explains most of the variance. Even the best one leaves out most of what determines whether someone will do well in that role.
Which makes the question “do hard or soft skills matter more?” badly posed from the start.
The finding that resolves the question: it isn’t “which”, it’s “both together”
The most useful work on this point is David Deming’s, published in the Quarterly Journal of Economics in 2017. He crossed the skill requirements of US occupations with employment and wage trends between 1980 and 2012.
The most-cited result is that social-skill-intensive occupations grew by nearly 12 percentage points as a share of total employment, with relatively strong wage growth. But the result that really matters is the interaction:
- Occupations with high maths requirements and high social requirements: strongest growth in both employment and wages.
- Occupations with high maths requirements and low social requirements, including many traditional STEM professions: down by around 3 percentage points over the same period.
That is: technical skills haven’t stopped paying. They’ve stopped paying on their own. The wage premium ended up on the combination, not on either column.
Deming and Lisa Kahn confirmed the same structure in 2018 in the Journal of Labor Economics, analysing millions of real job postings: positions requiring both high cognitive and high social skills are associated with higher pay, and the joint requirement has become more frequent over time.
The hard evidence on soft skills
There’s a piece of research that settled the “but can they actually be measured?” question elegantly, by James Heckman, the Nobel laureate in economics, and Tim Kautz. Their 2012 article in Labour Economics is literally called Hard Evidence on Soft Skills.
The case they use is the American GED, the equivalency diploma obtainable by exam. GED holders have average cognitive scores comparable to regular high school graduates, because the exam measures exactly that. But their outcomes in the labour market, in continuing education and in other areas of life look far more like those of people who dropped out.
The authors’ interpretation is that the regular diploma also certifies something the exam doesn’t measure: the ability to see a long process through. Perseverance, self-regulation, reliability. Things that appear in the economics literature as personality traits and behave like capital: they’re malleable, especially at a young age, and can be invested in.
The useful distinction isn’t between “hard” and “soft” skills. It’s between skills an exam certifies and skills only a process certifies.
How to combine the two kinds of skill
- Stop choosing. If you find yourself deciding whether to study SQL or communication, Deming’s data says the right answer is the combination, not the choice.
- Find the weak axis, not the strong one. Marginal returns are almost always higher where you’re below threshold. A brilliant technician who can’t explain what they did loses more value than they’d gain by learning another framework.
- Make the soft part verifiable. “Good communication skills” on a CV is worth nothing. “I ran the internal training on tool X for 40 colleagues” is worth something, because it’s checkable. I set out a method for writing this stuff in how to write a CV.
- Prepare for the structured interview, because it’s the most predictive filter. Concrete examples, situation-action-result, numbers where they exist. I’ve collected the most frequent questions in job interview questions.
- Don’t treat soft skills as innate traits. In Heckman’s literature they’re malleable. Which means they can be trained, but also that they take time and correction, not a weekend course.
- Look at where demand is heading. I’ve gathered the data in most in-demand skills and most in-demand soft skills.
My take: “soft” makes something essential sound optional
I think the term “soft skills” does more damage than it prevents. “Soft” suggests optional, and in many organisations it gets treated that way: the communication course is the first to go when the budget tightens. I much prefer the formulation economists use, social and behavioural skills, because it doesn’t carry an implicit judgement.
I also suspect that part of the corporate enthusiasm for soft skills comes from the fact that they’re cheaper to declare than to develop. Nobody can easily verify whether a company “values empathy”. Anyone can verify whether it has updated its technology stack.
And a note on AI, since it’s the variable that makes this article different from one written five years ago: if a growing share of basic technical work becomes assisted, the non-automatable part is exactly the coordination between people that Deming was measuring. But that’s a forecast, not a finding, and should be read as such.
Frequently asked questions
So is the 85% false? It isn’t false in the sense of refuted: it has no verifiable source. Nobody has managed to produce the 1918 study in the form it’s cited. A number with no source isn’t data.
If I were starting from zero, which would I begin with? The one that gets you past the first filter in your field, which is almost always the technical one. But with the awareness that the wage premium, in the data, is on the combination.
Can soft skills be measured in an interview? Imperfectly. The 2022 revision says the structured interview is the best single predictor, and much of what it measures is behavioural. The unstructured, chatty interview predicts far less.
Does conscientiousness still matter? Yes, but in the revised estimate its weight is more modest than was claimed (~0.19). It remains one of the most robustly performance-correlated personality traits.
Why isn’t the 2022 revision better known? Because the 1998 table entered textbooks and HR presentations across half the world, and circulating figures have far more inertia than methodological updates.
Sources
- Schmidt, F. L., & Hunter, J. E. (1998). The Validity and Utility of Selection Methods in Personnel Psychology: Practical and Theoretical Implications of 85 Years of Research Findings. Psychological Bulletin, 124(2), 262-274. The original table.
- Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting Meta-Analytic Estimates of Validity in Personnel Selection: Addressing Systematic Overcorrection for Restriction of Range. Journal of Applied Psychology, 107(11), 2040-2068. The revision of the estimates.
- Deming, D. J. (2017). The Growing Importance of Social Skills in the Labor Market. Quarterly Journal of Economics, 132(4), 1593-1640. The data on the interaction between maths and social skills, 1980-2012.
- Deming, D. J., & Kahn, L. B. (2018). Skill Requirements Across Firms and Labor Markets: Evidence from Job Postings for Professionals. Journal of Labor Economics, 36(S1), S337-S369. The analysis of real job postings.
- Heckman, J. J., & Kautz, T. (2012). Hard Evidence on Soft Skills. Labour Economics, 19(4), 451-464. The GED case and the malleability argument.
- Heckman, J. J., Stixrud, J., & Urzua, S. (2006). The Effects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior. Journal of Labor Economics, 24(3), 411-482.
- Autor, D. H. (2015). Why Are There Still So Many Jobs? The History and Future of Workplace Automation. Journal of Economic Perspectives, 29(3), 3-30. The framework on complementarity between automation and human skills.
- Barrick, M. R., & Mount, M. K. (1991). The Big Five Personality Dimensions and Job Performance: A Meta-Analysis. Personnel Psychology, 44(1), 1-26. The foundation of the conscientiousness literature.
- Mann, C. R. (1918). A Study of Engineering Education. Carnegie Foundation for the Advancement of Teaching, Bulletin Number Eleven. The Carnegie document that actually exists from that year, and which does not contain the 85/15 breakdown.
This article was written following the process described in how we write. If you find a figure reported incorrectly, write to me: I’ll correct it and flag it at the bottom of the article.



