Removing a degree requirement does not create skills-based hiring. Employers also need a way to define the work, recognize evidence from different paths, assess it consistently, and explain why the result predicts performance.

AI can help translate job information, retrieve candidates with unfamiliar titles, organize work evidence, and support structured assessment. It should not invent proficiency from keywords or replace validation with a similarity score.

Begin with the work to be done

List the recurring tasks, decisions, tools, knowledge, working conditions, and consequences of failure. Interview people who perform and supervise the job. Review actual work products and incidents. Separate what a person must know on day one from what can be learned after hiring.

Convert that analysis into a small set of observable requirements. “Can diagnose a production data pipeline failure” is more useful than “five years of data engineering.” “Can explain a complex policy decision to a customer” is more useful than “excellent communication.”

The US Office of Personnel Management’s competency policy resources describe competencies as measurable patterns of knowledge, skill, ability, behavior, and other characteristics. OPM’s framework applies directly to federal workforce practice; private employers can use the definition as a design reference, not as certification.

Remove credentials only where they are unnecessary

Some credentials are legally required or represent a defensible body of training. Others are convenient proxies that exclude people who learned through work, military service, community college, apprenticeships, certificates, or self-directed practice.

Review every degree, school, years-of-experience, and prior-title filter. Ask what job evidence it stands for. If the team can collect that evidence more directly, replace the proxy. If it cannot explain the requirement, remove it.

The US Department of Labor released a Skills-First Hiring Starter Kit for employers in November 2024. The announcement reflects federal support for reducing unnecessary degree barriers; it does not show that every skills-first program produces better outcomes.

Opportunity@Work describes workers who are skilled through alternative routes rather than a bachelor’s degree. Its research and advocacy provide evidence about this talent population, while employers still need to test their own job requirements and selection process.

Build an evidence ladder

Accept multiple forms of evidence, then rank them by relevance and reliability for the job:

  1. A work sample that closely resembles a job task.
  2. A structured interview response scored against behavioral anchors.
  3. A portfolio or prior work product with clear authorship and context.
  4. A verified record of performing comparable work.
  5. Training, credential, or course completion tied to the skill.
  6. Self-report or a resume keyword that requires confirmation.

The ladder avoids two extremes. Credentials do not receive automatic priority, and unverified claims do not become skills merely because an AI model extracted them.

For each skill, record the evidence source, recency, context, level of independence, and uncertainty. A person who used a tool once should not be scored like someone who owned a production system. A candidate should be able to correct a mistaken inference.

Use AI for translation and organization

Job and resume language varies across industries, countries, and employers. AI can propose related titles, map descriptions to a controlled skill vocabulary, and surface evidence that a literal keyword search would miss.

Keep the mapping visible. Show the original text beside the normalized skill and explain why the system made the connection. Let a recruiter accept, reject, or change it. Do not let the model add a skill that has no source evidence.

Treat confidence as a routing signal, not a proficiency score. Low-confidence mappings can go to review. High confidence still does not prove the person can perform the work.

AI can also draft work-sample variants, but a subject-matter expert must review difficulty, answerability, leakage, accessibility, and scoring. Record the version each candidate received and verify that variants measure the same construct.

Validate the assessment, not the marketing phrase

OPM’s hiring assessment guidance lists considerations including validity, reliability, adverse impact, applicant reactions, cost, and administration. These are useful questions for any employer choosing among work samples, interviews, tests, and experience records.

The federal Uniform Guidelines on Employee Selection Procedures provide principles for validation and recordkeeping in the adverse-impact context. They do not make an assessment valid merely because it is called skills-based or uses AI.

Evaluate the exact procedure in its intended role and population. Test scoring consistency and criterion relevance. Examine false negatives, not only average accuracy. Monitor results after deployment and repeat the analysis when the job, assessment, model, prompt, threshold, or applicant pool changes.

Avoid inferred personality, emotion, honesty, or culture fit. These labels are difficult to tie to observable work and can encode disability, language, culture, or evaluator preference.

Design for candidates without familiar signals

Skills-first hiring should make evidence legible across different paths. Give candidates examples of acceptable work products. Explain how an assessment is scored. Offer time and technology accommodations. Provide a route to clarify a credential or experience the system did not recognize.

Do not require an elaborate unpaid project when a shorter sample can measure the same skill. Compensate substantial work and avoid using candidate submissions as free production output. Protect confidential material from prior employers by allowing simulated or redacted evidence.

Use structured questions to establish context: What part did the candidate personally own? What constraints existed? What alternatives were considered? What changed as a result? Those answers often reveal skill level better than a keyword count.

Measure whether the funnel changed

Before removing a credential filter, record who enters and advances through the current process. After the change, track applicant mix, retrieval, assessment invitation, completion, stage progression, offers, acceptance, early job outcomes, and retention where appropriate.

Preserve denominators. More applications do not mean more qualified access. A higher pass rate does not prove the assessment predicts work. Better representation at the top of the funnel can disappear if later interview criteria remain unstructured.

Compare job-related outcomes without treating performance ratings as perfect ground truth. Ratings may contain manager effects and unequal opportunity. Use multiple measures and investigate discrepancies.

The NIST AI Risk Management Framework can organize ownership, measurement, monitoring, and response for AI components. It is voluntary and does not validate an employment procedure.

A staged implementation

Choose one role with sufficient hiring volume, managers willing to define evidence, and a clear business need. Remove unjustified credential filters. Build the evidence ladder and assessment rubric. Train reviewers and double-score a sample.

Run AI skill mapping in observation mode. Compare extracted skills with source records and expert review. Measure unsupported additions, missed evidence, and differences across candidate paths. Correct the taxonomy and prompts before the output influences advancement.

Then allow AI to organize evidence while a person makes the stage decision. Show the source, mapping, rubric, and reviewer action in one record. Establish override and candidate-correction routes.

Only automate a decision after the organization has validation evidence, monitored outcomes, appropriate legal review, and a tested rollback. Version every material component so historical decisions remain explainable.

Buyer questions for skills platforms

Ask the provider:

  • Which skill taxonomy is used, who maintains it, and how are changes versioned?
  • Does every inferred skill link to candidate-supplied or verified evidence?
  • How does the system distinguish exposure from independent proficiency?
  • Can candidates correct mappings and supply alternative evidence?
  • What validation applies to this role, assessment, population, and threshold?
  • How are accommodations handled?
  • Can the employer export the rubric, evidence, decisions, and version history?
  • What triggers retesting, and how are affected candidates reconsidered after an error?

A useful skills system broadens how evidence can enter without lowering the standard for what counts as evidence. It replaces weak proxies with closer measures of the work and makes uncertainty easier to inspect.

Sources and limits

This article draws on public resources from the US Department of Labor, OPM, Opportunity@Work, NIST, and the federal Uniform Guidelines. It is not legal advice. These sources do not certify an AI product or guarantee that a specific skills-first program will improve hiring outcomes.