
Skills-Based Hiring and AI Skills
On this page
Business
the increased opportunity promised by Skills-Based Hiring has borne out in not even 1 in 700 hires last year.
Skills-based hiring means judging candidates on demonstrated skills through work samples, structured interviews and a scorecard of outcomes, rather than relying on proxies such as degrees. The Harvard Business School and Burning Glass Institute report found that removing a degree requirement was easy to announce and much less common to translate into different hiring. Its “last year” refers to 2023.
Skills-based hiring quick reference
| Term | What it is | The number to know | The trap |
|---|---|---|---|
| Skills-based hiring | Assessment based on demonstrated abilities relevant to a role | About 3.5 percentage points more non-BA hiring in studied roles, HBS and Burning Glass Institute | Changing the posting while keeping the old assessment |
| Degree requirement removal | Dropping a degree from the stated requirements | About 3.6% of roles removed requirements, HBS and Burning Glass Institute | Treating a posting change as a hiring outcome |
| Work sample test | A task that resembles work the candidate would perform | .33 estimated validity, Schmidt, Oh and Shaffer, 2016 | Quoting the older .54 estimate as settled |
| Structured interview | Predetermined questions scored against proficiency benchmarks | .58 estimated validity, Schmidt, Oh and Shaffer, 2016 | Asking consistent questions without consistent scoring |
| Unstructured interview | A conversation whose questions and evaluation vary | .58 estimated validity, Schmidt, Oh and Shaffer, 2016 | Assuming a comfortable conversation produces comparable evidence |
| Operational validity | An estimate of how well a selection method predicts job performance | .42 for structured interviews, as reported from Sackett and colleagues, 2022 | Treating estimates from different analyses as fixed laws |
| Job description | A statement of duties and requirements | No universal number, role-specific | Listing tools without saying what success looks like |
| Scorecard | A role mission, measurable outcomes and competencies | Three to eight measurable first-year outcomes, our rule of thumb | Writing outcomes that cannot be observed |
| AI literacy | Knowledge and capability to use AI in context | Article 4 applied from 2 February 2025, AI Act Service Desk | Buying a certificate the article does not require |
| Upskilling | Building skills for a person's current role | 29 of 100 workers projected to be upskilled in role, WEF | Training people without changing their work |
| Reskilling | Building skills for a different role | 19 of 100 workers projected to be reskilled and redeployed, WEF | Assuming training alone creates an available role |
| Registered apprenticeship | A registered program combining work and related instruction | At least 2,000 on-the-job hours for a time-based term, 29 CFR 29.5 | Treating the recommended instruction hours as a requirement |
| Independent contractor, IRS | A worker classified using common-law evidence about control and the relationship | An SS-8 determination may take at least six months, IRS | Assuming a signed agreement settles status |
| Economic reality test, DOL | A separate test for employee status under the FLSA | The 2024 rule took effect on 11 March 2024, DOL | Assuming the IRS test resolves every classification question |
| Onboarding | The process of bringing a new hire into productive work | Around 12 months to peak performance potential, Gallup's experience | Confusing a first-week orientation with proficiency |
| Retention | Whether employees remain with the company | 10 percentage points higher for non-degreed workers at Leader firms, HBS and Burning Glass Institute | Assuming the finding predicts your company's result |
| Skills that complement AI | Abilities used to direct, check and apply AI output | Analytical thinking: 69% of employers, WEF | Hiring for a tool name without testing judgment |
Assessing candidates
Skills-based hiring
Definition. Skills-based hiring identifies the abilities needed for a role and collects evidence that candidates can use them. A degree may still be relevant where a role requires particular education or credentials, but it should not substitute for assessing the work itself.
The evidence. The HBS and Burning Glass Institute study examined 11,300 roles at large US firms where hiring could be observed before and after a degree requirement was removed. Across those roles, the share of hires without a BA rose by about 3.5 percentage points on average. Only about 3.6% of roles removed degree requirements during the study window, and the estimated added opportunity reached about 97,000 workers a year out of 77 million annual hires.
The researchers grouped firms that removed requirements into three types. Leaders, 37% of those firms, drove nearly all the change. About 45% were In Name Only, with no meaningful change in hiring, while about one fifth were Backsliders whose initial gains reversed.
The trap. Editing the posting and retaining the old interview can leave the same proxies in charge. The large-firm findings do not establish what will happen at a company with 30 or 100 employees.
Next action. For each open role, write one observable task that demonstrates a necessary skill and identify where candidates will have an equal opportunity to show it.
Degree requirement removal
Definition. Removing a degree requirement changes who can meet the stated conditions in a job posting. It does not by itself establish how applications will be screened or how interviewers will score answers.
The evidence. At Leader firms in the HBS and Burning Glass Institute study, non-degreed workers had retention rates 10 percentage points higher than their degree-holding colleagues. Non-degreed workers hired into roles that previously required a degree saw an average salary increase of 25%. These are findings for the studied firms and workers, not guarantees for a new hire.
The trap. A recruiter can remove “BA required” and then continue favoring familiar schools or career paths. The posting will look different while the selection decision stays much the same.
Next action. Review the screening form and remove any degree filter that is no longer a stated requirement for the role.
Work sample test
Definition. A work sample asks a candidate to perform a task resembling the job. It is especially useful when the candidate already has enough relevant knowledge to attempt the task, and when the output can be scored against an answer or observable standard.
The evidence. The estimate for work samples changed substantially between the 1998 analysis and the 2016 Schmidt, Oh and Shaffer working paper. The earlier .54 estimate rested on seven primary studies; Roth, Bobko and McFarland's 2005 reanalysis added many more studies and produced .33. They attributed the decrease to the inclusion of more service-sector samples, and the 2016 paper says .54 may still apply to manual skilled trades.
The trap. A task that depends on confidential company knowledge, unpaid hours of work or an unexplained marking scheme measures more than the skill you intended to test. The older .54 estimate also should not be presented as a universal current figure.
Next action. Use our rule of thumb: give each candidate a paid 45-minute task with 10 AI-drafted customer replies, three seeded with wrong facts, and score the errors caught and corrected. A copy of an exception queue can supply the context; the human oversight and guardrails guide explains the same seeded-check idea for live workflows.
Structured interview
Definition. A structured interview uses predetermined questions and evaluation rules tied to the role. The US Office of Personnel Management describes structured interviews as having “rules for eliciting, observing, and evaluating responses,” with answers “scored according to benchmarks of proficiency.”
The evidence. OPM describes giving applicants the same predefined lead and follow-up questions, drawing those questions from competencies identified through job analysis, and having panel members rate responses individually. The 2016 working paper estimated .58 validity for structured interviews. Sackett and colleagues' 2022 analysis ranked structured interviews first; summaries of the study report an estimate of .42.
The trap. Asking everyone the same questions helps little if “good answer” means whatever each interviewer happens to prefer. A panel discussion before individual scoring can also make it difficult to see where assessors disagree.
Next action. Apply our rule of thumb: ask every candidate the same five role-specific questions and have each panel member independently score each answer from 1 to 5 before discussion. Write anchors for every point: 1 means the answer misses the central issue, 3 means it identifies the issue and gives a plausible response, and 5 means it identifies the issue, explains a sound response and checks the result; define 2 and 4 as the observable steps between those anchors.
Unstructured interview
Definition. An unstructured interview allows the questions, follow-ups and evaluation to vary across candidates. It can reveal useful information, but the evidence collected from one person may differ substantially from the evidence collected from another.
The evidence. Schmidt, Oh and Shaffer estimated .58 for both structured and unstructured interviews in their 2016 working paper. Their newer correction for indirect range restriction changed the earlier estimates of .51 for structured interviews and .38 for unstructured interviews. That equality is an estimate under a particular correction, not proof that every informal conversation is equally predictive.
The trap. Treating an unstructured interview as useless overstates the evidence. Treating rapport as a score can be just as misleading when candidates were asked different questions.
Next action. Add one common, scored question to every candidate interview and record each answer before discussing impressions.
Operational validity
Definition. Operational validity is an estimate of how well a selection method predicts job performance. The figures below are correlations: a larger value indicates a stronger estimated relationship under the analysis used, not a percentage of hiring decisions that will succeed.
The evidence. Schmidt and Hunter published the 1998 figures; Schmidt, Oh and Shaffer revised estimates for 31 methods in a 2016 working paper. Sackett, Zhang, Berry and Lievens argued in their 2022 Journal of Applied Psychology study that earlier corrections for range restriction had overcorrected.
| Method | 1998 | 2016 | 2022 |
|---|---|---|---|
| General mental ability or cognitive ability | .51 | .65 | .31 |
| Structured interview | .51 | .58 | .42 |
| Unstructured interview | .38 | .58 | not shown here |
| Work sample | .54 | .33 | .33 |
| Job knowledge | .48 | .48 | .40 |
| Reference checks | .26 | .26 | not shown here |
| Years of job experience | .18 | .16 | not shown here |
| Years of education | .10 | .10 | not shown here |
The 2022 column shows only the estimates named in summaries of that study.

The work sample estimate fell after a larger set of studies was analyzed. Structured interview estimates also changed as researchers used and challenged different corrections.
The trap. The 2016 paper was a working paper, not peer reviewed in that form. These estimates are useful for choosing what evidence to collect, but the disputed corrections and differences between jobs make any single ranking a poor substitute for a role-specific assessment.
Next action. For your next hire, record the assessment methods you will use and the job outcome each is intended to predict.
Defining the role
Job description
Definition. A job description states duties and requirements: what the person is expected to do and what qualifications the company asks for. It is useful for communicating the role and setting boundaries around it.
The evidence. A duties list can become stale when an AI tool changes how a task is performed. An outcome such as resolving an order exception remains meaningful even if the steps used to reach it change.
The trap. A posting that lists ten tools may attract candidates who recognize the tools without showing whether they can solve the underlying problem. It can also leave interviewers with no common standard for judging success.
Next action. Mark each requirement in the current description as necessary on day one or teachable after hiring.
Scorecard
Definition. A scorecard states a role's mission, measurable outcomes and the competencies needed to achieve them. Geoff Smart and Randy Street describe the scorecard idea in Who, published in 2008.
The evidence. For an openly hypothetical customer operations coordinator at a 30-person distributor, our rule of thumb produces this scorecard:
- Mission: Resolve every order exception within one business day and tell the customer what happens next.
- Outcome by month 12: Keep the exception queue's median age under one business day.
- Ongoing outcome: Review AI-drafted replies before sending; check a weekly sample of 20 against the order system.
- Outcome by month 3: Rewrite two procedures for the AI-assisted workflow.
- Outcome by month 6: Train one colleague on the review checklist.
- Competencies: Analytical thinking, written clarity and judgment on refunds.

The description lists duties; the scorecard states the mission and outcomes. Build the interview questions and work sample from those outcomes.
The trap. A scorecard full of adjectives such as “adaptable” gives an interviewer little to score. An outcome also needs an owner and a way to observe it. The guide to roles for an AI-era company can help you redraw ownership before you write the role.
Next action. Write one mission and three to eight measurable first-year outcomes, then map each interview question and work sample to an outcome.
AI literacy
AI literacy and Article 4
Definition. AI literacy is the knowledge and capability people need to use AI in their work context, including enough understanding to recognize where an output needs checking. Article 4 of the EU AI Act addresses measures supporting AI literacy for people involved in operating and using AI systems on behalf of providers and deployers.
This is general information, not legal advice. Applicability and obligations depend on the organization, the system and how it is used.
The evidence. Article 4 has applied since 2 February 2025. The AI Act Service Desk timeline lists enforcement concerning AI literacy starting 2 August 2026.
The Digital Omnibus on AI amended Article 4 in July 2026. The original wording called for ensuring “a sufficient level of AI literacy”; the amended text asks for measures that “support the development of AI literacy.” The current Article 4 text ends its first paragraph with this sentence:
This obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual.
The trap. The article does not ask for a certificate. Nor was the obligation simply scrapped: the amended text still calls for measures, taking account of people's knowledge and the context in which systems are used.
Next action. Create a one-page record for each AI tool listing who uses it, what they use it for and what they were shown about using and checking it.
Building skills
Upskilling
Definition. Upskilling adds abilities that help someone perform their current role as its tasks change. A coordinator learning to check AI-drafted replies against an order system is an example.
The evidence. The World Economic Forum's Future of Jobs Report 2025 surveyed over 1,000 employers representing more than 14 million workers across 22 industry clusters and 55 economies. Employers expected 39% of existing skill sets to be transformed or become outdated over 2025 to 2030, down from 44% in the 2023 edition and 57% in 2020. These are survey expectations, not measurements of completed change.
Of a representative 100 workers, the report projected 29 would be upskilled in their current roles, 19 reskilled and redeployed, 11 unlikely to receive needed training, and 41 needing no significant training. Among employers, 85% planned to prioritize upskilling, 70% expected to hire people with new skills, and 40% planned to reduce staff whose skills became less relevant.
The trap. Tool training without a changed task leaves people with a lesson and no agreed place to use it. Training also needs to include what the person should check and when to seek a decision from someone else.
Next action. Use our rule of thumb: select one employee who already knows the context needed to check the output and train that person on the tool gap. The AI change management guide helps plan the rollout of changed tasks to existing staff.
Reskilling
Definition. Reskilling develops abilities for a different role. It may involve moving someone from work that is declining into a role with a different set of outcomes.
The evidence. In the WEF survey, employers expected 19 out of 100 workers to be reskilled and redeployed by 2030. Half of employers planned to move staff from declining to growing roles. Both figures describe expectations or plans, not confirmed moves.
The trap. A course does not create a vacancy or give someone practice in a new role. A reskilling plan needs an actual destination and someone who can assess progress against its requirements.
Next action. Name one available or planned role, attach its scorecard and identify the first task the person can practice.
Hiring or contracting for a skill gap
Definition. Hiring brings a capability into a role you expect to maintain. Contracting can obtain a bounded piece of work from an independent provider, subject to the applicable classification rules.
The evidence. The WEF survey found that 70% of employers expected to hire staff with new skills. Our rule of thumb is to hire when the scorecard needs a capability nobody has and the expected ramp fits the business; Gallup's estimate of around 12 months to peak performance potential is a useful planning prompt, not a deadline for every role.
If an outside operator runs an AI layer, the employee's scorecard may emphasize reviewing outputs and handling exceptions. Iron Goo's operations service follows an audit, build and operate model for chatbots answering repeat inbound questions and an agentic operations layer for internal operations work with scoped credentials and audit logging.
The trap. A short-term need does not automatically make a worker an independent contractor. A permanent need for judgment may also remain after a tool is deployed.
Next action. Use our rule of thumb: write the missing capability, the expected duration and who controls how the work is done before choosing to hire or contract.
Registered apprenticeship and on-the-job training
Definition. A US registered apprenticeship combines on-the-job learning with related instruction. Under 29 CFR 29.5, its term can be time-based, competency-based or hybrid.
The evidence. A time-based term requires at least 2,000 hours of on-the-job learning. The regulation recommends a minimum of 144 hours of related technical instruction for each year of apprenticeship; it does not make that number a requirement.
The Stanford Digital Economy Lab's August 2026 revision, using ADP payroll data through June 2026, found employment among 22- to 25-year-olds in highly exposed occupations about 19% below the level implied by keeping pace with similarly aged workers in less-exposed occupations. That is a gap against a comparison trend, not a 19% fall. The authors describe the patterns as descriptive, not causal, and say the adjustment appears to come mainly through reduced hiring.
The trap. Calling informal shadowing a registered apprenticeship does not make it one. If entry-level hiring into exposed work shrinks, our inference is that a company's own on-the-job ladder may become its pipeline for judgment-heavy roles.
Next action. Choose one recurring task and document the supervised steps a new person must demonstrate; the guide to processes and SOPs that survive automation can help turn those steps into written procedures.
Contractors and employees
Independent contractor under IRS common law
Definition. For US federal tax classification, the IRS considers evidence in three categories: behavioral control, financial control and the type of relationship. The facts of the working relationship matter.
This is general information, not legal advice. Federal and state rules can apply differently to the same working relationship.
The evidence. The IRS classification guidance states: “The keys are to look at the entire relationship and consider the extent of the right to direct and control the worker.” Form SS-8 requests an IRS determination, which the IRS says may take at least six months.
The trap. A contractor agreement or invoice does not settle the classification on its own. The actual right to direct and control the work remains part of the assessment.
Next action. Before signing, write down who controls how the work is done, including who chooses the methods and directs the worker.
DOL economic reality test
Definition. The US Department of Labor uses a separate economic reality test for employee status under the Fair Labor Standards Act. It addresses a different legal question from the IRS tax classification.
The evidence. The DOL misclassification page, read 30 October 2026, still pointed to the 2024 rule, which took effect on 11 March 2024. The DOL proposed replacing it in February 2026, so the applicable rule should be checked when a decision is made. State tests can be stricter.
The trap. A favorable answer under one test does not automatically resolve the others. The rule's status has also been in flux, making an old checklist unreliable.
Next action. Check the current DOL rule and the relevant state test before classifying the worker.
Keeping people
Onboarding
Definition. Onboarding is the process through which a new employee learns the work, the people involved and the standards for performing the role. Orientation is one part of that process.
The evidence. In an August 2021 Gallup article, only 12% of US employees said their company did a good job of onboarding. Gallup's experience suggested new employees typically take around 12 months to reach peak performance potential.
The trap. A completed first-week checklist does not show that someone can handle an exception or recognize a bad AI output. Twelve months is a planning reference from Gallup's experience, not a fixed probation period.
Next action. Use our rule of thumb: create a 90-day plan tied to scorecard outcomes, with a named reviewer and an observable task at each 30-day checkpoint.
Retention
Definition. Retention describes whether employees remain with the company over a specified period. It is useful alongside evidence about performance and the conditions that let people do the job.
The evidence. In the HBS and Burning Glass Institute study, non-degreed workers at Leader firms had retention rates 10 percentage points higher than degree-holding colleagues. That result describes the studied large US firms; it does not predict a particular company's retention rate or establish a single cause.
The trap. A hiring change can widen access while the job itself still gives a new employee unclear expectations or no route to progress. Retention figures also need a defined period and comparable groups to be useful.
Next action. Record the share of new hires still employed at a defined checkpoint, and review it alongside their progress on the role scorecard.
Skills that complement AI
Judgment and core skills
Definition. Judgment is the ability to decide what an output means, whether it is acceptable and what action should follow. In an AI-assisted role, it includes checking an answer against the relevant business facts and recognizing when a person must decide.
The evidence. In the WEF's 2025 employer survey, the ten most frequently named core skills were:
| Core skill | Share of employers |
|---|---|
| Analytical thinking | 69% |
| Resilience, flexibility and agility | 67% |
| Leadership and social influence | 61% |
| Creative thinking | 57% |
| Motivation and self-awareness | 52% |
| Technological literacy | 51% |
| Empathy and active listening | 50% |
| Curiosity and lifelong learning | 50% |
| Talent management | 47% |
| Service orientation and customer service | 47% |
AI and big data ranked 11th at 45%; programming was named by 17%. These figures show what surveyed employers called core skills, not a measured ranking of what causes better performance.
Humlum and Vestergaard's March 2026 revision of Still Waters, Rapid Currents studied 25,000 workers across 7,000 workplaces and 11 occupations in Denmark. Two years after ChatGPT, the authors ruled out earnings effects above 2%; adopters reported saving about 3% of work hours, and 85% reallocated saved time. Reported take-up was about 41% with no employer initiative, 76% with encouragement alone and 93% with all initiatives; those comparisons are descriptive and bundle selection.
Census CES 26-25 is an unreviewed working paper, not official statistics. It reports that 2.0% of AI-using firms reported AI-related employment decreases, or 2.4% employment-weighted, while increases were 2.3%. It also reports that 52% of AI-using firms saw no task effect, and about 66% of firms reporting a task effect only augmented work.
The trap. A candidate's familiarity with a model or prompt does not establish that they can catch an incorrect refund, identify a missing order or explain a decision to a customer. The wider question of which decisions remain with the owner appears in what running a business means in the AI era.
Next action. Add one realistic incorrect AI output to the role assessment and score whether the candidate identifies the error, checks the source and chooses a sound next action.
Put your next steps in order
Build on what you have just read with a free ebook on semantic SEO and AI visibility. Get worked examples, decision rules, checklists and a do-this-week action in every chapter, plus a six-week plan for your site.



