MGI · Defining the skills citizens will need · Jun 2021
Distinct Elements of Talent — a field atlas

Fifty-six things
a person can
actually get better at.

McKinsey Global Institute surveyed 18,000 people across 15 countries and came back with a list. Not competencies. Not values. Fifty-six discrete, nameable, teachable behaviours — grouped into 13 skill groups across four domains — that its data links to staying employed, earning more, and being satisfied at work.

COGNITIVE 16 INTERPERS. 14 SELF-LEAD. 15 DIGITAL 11
Bar widths = share of the 56 Source: MGI 2021
56DELTAs named
13Skill groups
4Domains
18,000Adults surveyed
15Countries

What a DELTA is

and the four things it is not

The definition, precisely

DELTA stands for Distinct ELements of TAlent. MGI chose the word “element” on purpose: each one is meant to be the smallest unit of capability that can still be described as a behaviour, observed in someone, and improved on its own.

The framework deliberately mixes registers. Structured problem solving sits beside humility, which sits beside cybersecurity literacy. That mixing is the argument: in a labour market being reshaped by automation and generative AI, the durable advantage is not one type of skill but a portfolio spanning how you think, how you work with people, how you govern yourself, and how you handle machines.

  • Behavioural, not dispositional. “Asking the right questions” is a thing you do in a meeting, not a personality trait.
  • Foundational, not vocational. They transfer across roles and sectors, which is exactly why they survive job disruption.
  • Independently movable. You can be excellent at synthesising messages and terrible at coping with uncertainty.

What it is not — read this before you use it

  • Not a hiring rubric. MGI published a research taxonomy, not a validated selection instrument. Scoring candidates out of 56 will produce noise and legal exposure. Pick the three DELTAs a specific role actually fails on.
  • Not “soft skills.” The word soft implies unmeasurable. Half this list — programming literacy, data analysis, work-plan development — is hard-scoreable. The other half is scoreable too, just with different instruments.
  • Not objectively measured in the original study. Proficiency came from an online self-assessment. Those numbers are perceptions of skill, not demonstrations of it. Treat every headline figure as correlational and self-reported.
  • Not a curriculum, and not a ranking. The four domains are categories, not stages. Nobody “completes” cognitive and moves on to digital.
Fifty-six is not a syllabus. It is a diagnostic vocabulary — a way to say what is missing without saying “they’re just not very good.”

The full map

All 56, by group — with one worked deep dive per domain
COG · 16

Cognitive

How you break a problem apart, sequence work, put an idea into someone else’s head, and change your mind. The domain most exposed to generative AI — and, paradoxically, the one where the human premium is moving up the stack, from producing analysis to judging it.

Critical thinking 4

Structured problem solvingLogical reasoningUnderstanding biasesSeeking relevant information

Planning & ways of working 3

Work-plan developmentTime management & prioritisationAgile thinking

Communication 4

Storytelling & public speakingAsking the right questionsSynthesising messagesActive listening

Mental flexibility 5

Creativity & imaginationTranslating knowledge to different contextsAdopting a different perspectiveAdaptabilityAbility to learn
Deep dive — Structured problem solvingCognitive / Critical thinking
What it actually is

Taking a vague, data-poor question and turning it into a set of components that don’t overlap and don’t leave gaps; forming a falsifiable hypothesis; then sequencing the analysis by value of information rather than by what’s easy to pull.

What it looks like in the room

They draw the tree before they open the spreadsheet. They can tell you, unprompted, what evidence would prove them wrong. When you hand them a contradicting fact at minute 30, they redraw a branch instead of defending the answer.

What it is not

It is not intelligence, speed, or being well-read — plenty of very clever people solve unstructured and get lucky. It is not slide craft. It is not “having the answer”: someone who arrives with the conclusion and reverse-engineers support is demonstrating the opposite skill.

How to test it

A 45–60 minute live case on a problem they’ve never seen, with deliberately missing data. Score the process on five anchored dimensions: clarified the question / decomposition is non-overlapping / prioritised by impact × effort / stated a falsifiable hypothesis / updated on new evidence. Two independent raters, then check they agree. McKinsey’s own Solve assessment does a gamified version of exactly this.

If it’s weak — how to shift it
  1. Change the artefact, not the attitude. For twelve weeks, no analysis begins without a one-page issue tree signed off by a peer. Skill follows the mandatory artefact.
  2. Force the falsifier. Every recommendation carries a sentence beginning “This is wrong if…”. It is the cheapest debiasing intervention that exists.
  3. Review the tree, not the answer. If managers only ever critique conclusions, people optimise for confident conclusions.
  4. Re-test cold at week 0 and week 12 on a fresh case with different raters. If the rater score hasn’t moved, the problem was never training — it was that nobody has time to think.
INT · 14

Interpersonal

Not charm. The mechanics of moving other people and other systems — building the trust that makes information flow toward you, and structuring groups so the quietest person’s knowledge reaches the decision.

Mobilising systems 4

Role modellingWin-win negotiationsCrafting an inspiring visionOrganisational awareness

Developing relationships 4

EmpathyInspiring trustHumilitySociability

Teamwork effectiveness 6

Fostering inclusivenessMotivating different personalitiesResolving conflictsCollaborationCoachingEmpowering
Deep dive — Fostering inclusivenessInterpersonal / Teamwork effectiveness
What it actually is

Changing the conditions of a group so that the least-powerful member’s information reaches the decision. It is an engineering skill applied to conversations: who speaks, in what order, with what protection.

What it looks like in the room

Air time is noticed and redistributed. The most junior person speaks first on a contested question. Dissent is requested by name. Credit is re-attributed live — “to be clear, that was Priya’s point.” Interruptions get a soft, immediate correction.

What it is not

Not being liked. Not demographic representation on an org chart — that’s an input, not this behaviour. Not conflict avoidance: genuinely inclusive teams argue more, because disagreement has stopped being expensive. And absolutely not a training-completion percentage.

How to test it

Never ask the person. Use (a) 360 feedback with behaviour-anchored items answered by subordinates and peers, weighted to subordinates; (b) meeting telemetry — distribution of speaking time, interruption counts, who tables agenda items; (c) Edmondson’s seven-item psychological-safety scale at team level, tracked quarterly. The sharpest signal: does the variance between team members’ safety scores shrink under this leader?

If it’s weak — how to shift it
  1. Use structure, not exhortation. Round-robin openings; silent-read-then-write-before-discussion; a rotating chair. Protocols move behaviour in weeks; “be more inclusive” moves nothing.
  2. Log dissent. Every significant decision records who disagreed and why. Naming dissent legitimises it.
  3. Re-measure the team at 90 days, not the individual. This DELTA lives in the group’s numbers.
  4. If it still doesn’t move, stop training. When safety scores are flat under a well-intentioned manager, the constraint is usually authority, tenure precarity or workload — a system fault wearing a skill costume.
SLF · 15

Self-leadership

The domain MGI found most strongly tied to job satisfaction and to staying employed — and the one formal education is worst at building. How you regulate yourself, start things nobody asked for, and finish them when the ground moves.

Self-awareness & self-management 6

Understanding own emotions & triggersSelf-control & regulationUnderstanding own strengthsIntegritySelf-motivation & wellnessSelf-confidence

Entrepreneurship 4

Courage & risk-takingDriving change & innovationEnergy, passion & optimismBreaking orthodoxies

Goals achievement 5

Ownership & decisivenessAchievement orientationGrit & persistenceCoping with uncertaintySelf-development
Deep dive — Coping with uncertaintySelf-leadership / Goals achievement
What it actually is

Holding decision quality steady when the outcome distribution is genuinely unknown. Acting on incomplete information without either freezing or over-committing to the first plausible story.

What it looks like in the room

They make reversible decisions fast and irreversible ones slowly, and they can tell you which is which. They express confidence as a range, not a verdict. They buy cheap optionality instead of demanding certainty. They still sleep.

What it is not

Not confidence, not optimism, not appetite for risk — a gambler is not coping with uncertainty, they are ignoring it. Not stoic silence. Not the absence of anxiety: this DELTA is functioning while anxious, which is why calm-seeming people often score poorly on it under a real shock.

How to test it

Self-report is close to worthless here — everyone rates themselves calm. Use calibration: have them log 20 forecasts with confidence intervals over a quarter and score the calibration curve (Brier score). Add a simulation with a mid-exercise information shock and observe whether the plan is revised or defended. Add retrospective decision audits: pull ten past calls and ask what was knowable at the time.

If it’s weak — how to shift it
  1. Build the forecasting log. Nothing improves tolerance of uncertainty like discovering, with evidence, that your 90% confidence is really 60%.
  2. Separate “decide” from “commit.” Most paralysis is people treating a two-way door as a one-way door.
  3. Give uncertainty an end date. Explicit tripwires — “if X hasn’t happened by 14 March, we switch” — convert open-ended dread into a bounded wait.
  4. Check the substrate first. Tolerance for ambiguity collapses under sleep debt and role ambiguity. Fix who-decides-what before you book the resilience workshop.
  5. Measure the Brier score, not self-rated confidence. The second one goes up when the first one doesn’t.
DIG · 11

Digital

The smallest domain in the 2021 taxonomy and the one that has aged fastest. Written before ChatGPT, it still holds up — because it separates using tools from understanding systems, and the second half is precisely what generative AI has made scarce.

Digital fluency & citizenship 4

Digital literacyDigital learningDigital collaborationDigital ethics

Software use & development 3

Programming literacyData analysis & statisticsComputational & algorithmic thinking

Understanding digital systems 4

Data literacySmart systemsCybersecurity literacyTech translation & enablement
Deep dive — Data literacyDigital / Understanding digital systems
What it actually is

Arguing with a number. Knowing what a figure was made of, which population it represents, how it was collected, and — the hard part — what it structurally cannot support.

What it looks like in the room

They ask for the denominator before the trend line. They separate correlation, causation and selection out loud. They spot survivorship bias in a customer-satisfaction chart. They are willing to say “this dataset cannot answer that question” and stop.

What it is not

Not dashboard usage. Not tool proficiency — Excel, SQL and Tableau live one group over, in software use. Not a statistics qualification. And emphatically not fluency with an AI assistant: a person who accepts a confidently wrong model output has demonstrated a data-literacy failure, not a prompting one.

How to test it

Short applied assessment, 30 minutes. Hand over three real, flawed charts and a claim attached to each; score errors found against false alarms so you don’t reward reflexive scepticism. Then ask them to reproduce one company KPI end-to-end from the source system. Finally, give them a plausible, wrong gen-AI answer with a fabricated citation and see whether it survives.

If it’s weak — how to shift it
  1. Make provenance mandatory. Every number in every deck carries source, definition and date. Literacy is a by-product of having to write that line.
  2. Teach denominators before tools. Most organisations buy the analytics licence and skip the sampling lesson, then wonder why the dashboards are ignored.
  3. Run a monthly chart clinic. Twenty minutes, one real internal chart, everyone hunts the flaw. Cheap, social, and it changes what gets published.
  4. Retest with a fresh flawed chart at 90 days — never the same one, or you’re measuring memory.

What the data actually said

MGI 2021 · 18,000 respondents · 15 countries
Outcome 01

Linked to being employed

  • Adaptability
  • Coping with uncertainty
  • Synthesising messages
  • Achievement orientation
  • Work-plan development

Notably: three of the five sit in self-leadership, not in cognitive or digital.

Outcome 02

Linked to higher income

  • Work-plan development
  • Asking the right questions
  • Self-confidence
  • Coping with uncertainty

MGI also reported digital DELTAs mattering more for income at higher education levels — the digital premium is not evenly distributed.

Outcome 03

Linked to job satisfaction

  • Self-motivation & wellness
  • Coping with uncertainty
  • Self-confidence
  • Understanding own emotions & triggers

Satisfaction tracked almost entirely with self-leadership. None of the top drivers were technical.

The finding people quote most, and the caveat they skip.

MGI’s headline conclusion was that higher educational attainment predicted higher proficiency in only a subset of the 56 — strongly for cognitive and digital DELTAs, far more weakly for self-leadership and interpersonal ones. The implication drawn was blunt: school systems are not currently building the skills most associated with employment and satisfaction, and curricula need to teach DELTAs explicitly rather than hope they arrive as a side-effect of subject learning.

But: proficiency was self-assessed in an online survey, the associations are correlational, and self-report is systematically distorted — least-skilled respondents over-rate themselves, and cultural response styles differ across the 15 countries. Use these findings to choose where to look. Do not use them as evidence that a named person lacks a named skill.

How to test a DELTA

Four instruments, ranked by fidelity

Every instrument below is legitimate. The failure mode is using a cheap one for a high-stakes decision. Bars indicate fidelity — how closely the measurement resembles the actual behaviour in the actual job.

Instrument 01

Self-report inventory

Good for: cohort-level trends, opening a development conversation, cheap coverage of all 56.
Fails at: anything with consequences. Social desirability, unfamiliar-language effects, and the fact that low performers over-rate themselves hardest.
Rule: never for selection, promotion or pay. This is the instrument MGI itself used — which is why its numbers are a map, not a verdict.

Instrument 02

Situational judgement & gamified assessment

Good for: volume screening on cognitive and interpersonal DELTAs; scales to thousands; harder to fake than a questionnaire.
Fails at: distinguishing “knows the right answer” from “does the right thing at 6pm on a Friday.”
Rule: validate the scoring key against actual performance in your organisation before you trust it.

Instrument 03

Work sample & live simulation

Good for: almost everything cognitive and digital. A real case, a real messy dataset, a real hostile stakeholder, scored on a behaviourally anchored scale by two independent raters.
Fails at: traits that only appear over months — grit, integrity, coping with sustained uncertainty.
Rule: write the anchors before you see anyone, and check inter-rater agreement. Unanchored “gut feel” panels measure similarity to the panel.

Instrument 04

Multi-source behavioural observation over time

Good for: the interpersonal and self-leadership DELTAs that no exercise can simulate. 360s with behavioural items, team-level climate scales, decision audits, meeting telemetry.
Fails at: anything where raters have an incentive. Politics contaminate 360s fast.
Rule: weight subordinate and peer input above the manager’s, and read the trend rather than the level.

Five rules that decide whether any of this is real

  • Define the behaviour before you build the scale. “Collaboration: 3/5” is not data. “Shares work in progress before it is finished: observed 4 of 5 sprints” is.
  • Two raters, then check they agree. If they don’t, your anchors are vague — fix the instrument, not the person.
  • Measure change, not level. Absolute DELTA scores are hopelessly confounded by role, tenure and confidence. Deltas in DELTAs are the signal.
  • Re-test with a different instrument than the one that flagged the gap. Otherwise you are measuring practice effects.
  • Assume the field has moved. Sackett and colleagues’ 2022 re-analysis reordered the validity rankings that had been quoted from Schmidt & Hunter for two decades — structured interviews and job-knowledge tests came out at the top once range restriction was handled properly. Check your assessment vendor knows this.

If something is wrong, how to shift it

A diagnostic sequence — run in order

Most organisations skip straight to step 04 and buy a course. That is why most skill programmes produce completion certificates and no behaviour change. The order matters more than the interventions.

Separate skill, will and system

Before naming a DELTA gap, ask: can they do it elsewhere? Would they, if it were rewarded? Does the structure permit it? An analyst who never asks the right questions in a meeting where the partner has already announced the answer has a system problem. Training will make it worse — you’ll have taught a behaviour the environment punishes.

Test: find one context where the person already displays the DELTA. If you find one, stop; it’s not a skill gap.

Pick three, not fifty-six

Choose the DELTAs that are load-bearing for the specific work this specific team fails at, and that MGI’s data links to the outcome you care about. Broad-spectrum capability programmes dilute to nothing. Three named behaviours, one quarter.

Test: can every person name the three from memory in week six? If not, you picked too many.

Change the artefact and the meeting, not the mindset

Behaviour follows structure faster than it follows belief. Mandatory issue tree. Written-first meetings. A “this is wrong if…” line. A decision log. These are boring, cheap, and they work in weeks — where mindset workshops produce a warm afternoon and a flat curve at 90 days.

Test: is the new behaviour required by a template or a ritual, or merely encouraged? Only the first survives a busy quarter.

Then, and only then, practise deliberately

Real stakes, small scope, feedback inside 24 hours against the same anchored criteria you tested with. Reps at the edge of current ability — not a five-hour module. Ericsson’s condition holds: practice without immediate, specific feedback produces experience, not improvement.

Test: how many hours between the attempt and the feedback? Over 48 and you have a course, not practice.

Re-measure with a different instrument

If a self-report survey flagged the gap, close it out with a work sample or a team-level scale. Evaluate at Kirkpatrick level 3 and 4 — behaviour and results — and treat satisfaction scores as attendance data. Training hours delivered is not a metric; retire it publicly.

Test: can you point to one changed outcome — a decision, a cycle time, a retention number — not one changed feeling?

If it still hasn’t moved, change the job

After two honest cycles with no shift, the humane and accurate conclusion is fit, not deficiency. Self-leadership and interpersonal DELTAs are the slowest to move and the most context-dependent. Redesigning the role around someone’s strong DELTAs is a legitimate answer — and usually a cheaper one than a third programme.

Test: which three of the 56 does this person already score high on, and what job is built out of those?

Most skill gaps are system gaps wearing a skill costume. Check the costume before you buy the course.

What has changed since 2021

Where the taxonomy still holds — and where to cross-check it

The gen-AI reweighting

MGI’s 2023 work on generative AI and the US workforce projects a sharp reweighting of hours by 2030: demand for technological skills rising steeply, demand for social and emotional skills rising materially, and demand for basic cognitive skills — routine data entry, basic literacy, elementary processing — falling.

Read against the DELTA map, that promotes the judgement-side cognitive DELTAs (understanding biases, asking the right questions, synthesising messages) over the production-side ones, and it moves data literacy and digital ethics from nice-to-have to load-bearing. The 2021 list didn’t need rewriting. It needed reordering.

The independent cross-check

The World Economic Forum’s Future of Jobs Report 2025, drawn from employers rather than individuals, lands remarkably close: employers expect roughly 39% of workers’ core skills to change by 2030; analytical thinking remains the most-cited core skill; and the fastest-growing set pairs AI and big data with creative thinking, resilience, flexibility and agility, and curiosity and lifelong learning.

Two independently-built frameworks converging on the same handful of behaviours is the strongest evidence available that this list is not a consulting artefact. Where they disagree — WEF is far more emphatic about AI-specific skill demand — the more recent source should win.

Sources

Every figure above, traceable

A · Primary source for the 56 DELTAs

R01 McKinsey Global Institute — Defining the skills citizens will need in the future world of work (25 June 2021). Dondi, Klier, Panier, Schubert. mckinsey.com/industries/public-sector/our-insights/defining-the-skills-citizens-will-need-in-the-future-world-of-work Supports: the 56 DELTAs, 13 skill groups, 4 domains, 18,000 respondents in 15 countries, and all employment / income / satisfaction associations.
R02 McKinsey Global Institute — Skill shift: Automation and the future of the workforce (May 2018). mckinsey.com/featured-insights/future-of-work/skill-shift-automation-and-the-future-of-the-workforce Supports: the pre-DELTA modelling of rising demand for higher cognitive, social/emotional and technological skills.

B · The gen-AI reweighting

R03 McKinsey Global Institute — Generative AI and the future of work in America (July 2023). mckinsey.com/mgi/our-research/generative-ai-and-the-future-of-work-in-america Supports: projected 2030 shifts in hours by skill category.
R04 McKinsey Digital / MGI — The economic potential of generative AI: the next productivity frontier (June 2023). mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier Supports: which work activities are most exposed, by occupation.

C · Human capital, skills-based organisations

R05 MGI — Human capital at work: The value of experience (June 2022). mckinsey.com/mgi/our-research/human-capital-at-work-the-value-of-experience Supports: how much lifetime skill is built on the job rather than in school.
R06 MGI — Performance through people: Transforming human capital into competitive advantage (Feb 2023). mckinsey.com/mgi/our-research/performance-through-people-transforming-human-capital-into-competitive-advantage Supports: the firm-level economics of investing in capability.
R07 McKinsey People & Organizational Performance — Taking a skills-based approach to building the future workforce (Nov 2022). mckinsey.com/capabilities/people-and-organizational-performance/our-insights/taking-a-skills-based-approach-to-building-the-future-workforce Supports: operationalising a skill taxonomy inside an organisation.
R08 Deloitte Insights — The skills-based organization: A new operating model for work and the workforce. www2.deloitte.com/us/en/insights/topics/talent/organizational-skill-based-hiring.html Independent counterpart to R07 — useful for disconfirming evidence.

D · Independent cross-checks on the skill list

R09 World Economic Forum — Future of Jobs Report 2025 (Jan 2025). weforum.org/publications/the-future-of-jobs-report-2025/ Supports: ~39% of core skills changing by 2030; analytical thinking as top core skill; fastest-growing skill list.
R10 OECD — Future of Education and Skills 2030 / Learning Compass. oecd.org/education/2030-project/ Supports: the education-system argument, with a competency framework built independently of MGI.
R11 European Commission JRC — DigComp 2.2: The Digital Competence Framework for Citizens. joint-research-centre.ec.europa.eu/digcomp_en Use this rather than the 11 digital DELTAs if you need assessable, levelled digital criteria.

E · Measurement science — for the testing section

R12 Sackett, Zhang, Berry & Lievens (2022) — Revisiting meta-analytic estimates of validity in personnel selection, Journal of Applied Psychology. doi.org/10.1037/apl0000994 The current best estimate of which selection instruments actually predict performance.
R13 Schmidt & Hunter (1998) — The validity and utility of selection methods in personnel psychology, Psychological Bulletin. doi.org/10.1037/0033-2909.124.2.262 The classic that R12 corrects. Read both, in that order.
R14 Edmondson (1999) — Psychological safety and learning behavior in work teams, Administrative Science Quarterly. doi.org/10.2307/2666999 Source of the 7-item team scale used in the inclusiveness deep dive.
R15 Ericsson, Krampe & Tesch-Römer (1993) — The role of deliberate practice in the acquisition of expert performance, Psychological Review. doi.org/10.1037/0033-295X.100.3.363 Supports: feedback latency as the binding constraint on skill change.
R16 Yeager et al. (2019) — A national experiment reveals where a growth mindset improves achievement, Nature. doi.org/10.1038/s41586-019-1466-y Supports: mindset interventions work only where the surrounding norms permit — the “system gap” argument.
R17 Mellers et al. (2014) — Psychological strategies for winning a geopolitical forecasting tournament, Psychological Science. doi.org/10.1177/0956797614524255 Supports: calibration training as the testable route into “coping with uncertainty.” Practice publicly at gjopen.com.
R18 Kirkpatrick Partners — The Kirkpatrick Model. kirkpatrickpartners.com/the-kirkpatrick-model/ Supports: evaluating capability programmes at behaviour and results level, not satisfaction.
R19 Duckworth — Grit Scale. angeladuckworth.com/grit-scale/ A worked example of a self-report instrument, with the author’s own published cautions about using it for selection.
R20 McKinsey Careers — Solve (the firm’s gamified problem-solving assessment). mckinsey.com/careers/interviewing Supports: what a scaled, process-scored test of structured problem solving looks like in practice.

Where a claim above is stated without a reference — the deep-dive test protocols and the shift sequence, in particular — it is applied practice built on these sources, not a finding published by MGI. It is marked as such throughout so you can separate the evidence from the interpretation.