Pre-Employment and Talent Assessment: The Test Was Never the Hard Part

A pre-employment test is the cheapest instrument in hiring and the most expensive thing in it to defend. The measurement science behind it is decades old, well evidenced, and — on the most recent re-examination — more modest than the sales decks suggest. The volumes hitting the funnel are not modest at all. What is scarce in 2026 is neither the test nor the candidates. It is the audit trail that lets an employer use either one.

The funnel stopped telling employers who applied

LinkedIn’s chief executive, Dan Shapero, put the problem in one sentence to the Wall Street Journal in September 2026: “It’s gotten harder and harder for employers to know who can actually do the job anymore.” The number behind it is that job seekers are sending out 30% more applications than they did before the pandemic, and that recruiters now dig through hundreds or thousands of applicants for a single posting. Job hunters were averaging 45 applications a month as of May 2025, on data shared with Fortune; if the same 30% uplift applies across the period, the pre-pandemic monthly rate works out at about 35 applications (Pheonix Research calculation: 45 ÷ 1.3 = 34.6).

The market underneath is not hot. US job openings fell to a five-month low of 7.08 million in August, and monthly hires excluding the pandemic period are nearing a ten-year low, both per Fortune‘s reporting. Candidates are not applying more because there is more to apply to. They are applying more because each application got cheaper to produce, and AI is what made it cheap.

Employers answered in kind. Roughly 99% of hiring managers say they use AI somewhere in the hiring process, and 98% report improvements in interview scheduling, resume screening and skills assessment, according to a 2025 Insight Global report cited by Fortune. LinkedIn ships an AI hiring assistant to filter incoming applications. HireVue acquired the technology behind Hireguide in March 2026 to build a voice-based AI interviewer, and said plainly why: “The rise of AI-generated applications has compounded applicant volumes, making it harder than ever to distinguish real capability from polished resumes.”

That is where assessment stops being one stage of the process and becomes the stage where a generated document stops being the deliverable. A resume is now an artefact anyone can produce. A structured work sample is not.

What the evidence actually supports

The canon here is the meta-analytic validity literature, and it was quietly rewritten. For two decades the reference summary was Schmidt and Hunter’s 1998 paper. In 2022, Paul Sackett, Charlene Zhang, Christopher Berry and Filip Lievens revisited it in the Journal of Applied Psychology and concluded that the range-restriction corrections used across the field had systematically overcorrected, inflating how well selection procedures predict job performance. Their revised estimates cut mean validity by 0.10 to 0.20 points. Structured interviews emerged as the top-ranked procedure. The authors were explicit that predictors “remain useful, but selection predictor–criterion relationships are considerably lower than previously thought.”

Three things follow for anyone buying or selling assessment.

First, the ranking survived even though the levels fell. Structured interviews, work samples and well-constructed ability tests still sit near the top of the table; the unstructured interview, which is how most hires are actually made, sits far lower. Assessment’s advantage over the status quo is real. It is simply smaller than advertised.

Second, the same paper pairs each procedure’s validity with the mean Black–White subgroup difference it produces, because the two travel together. That relationship is not new: a 2001 meta-analysis by Philip Roth and colleagues found the accepted Black–White mean difference in cognitive ability in job-applicant samples to be approximately one standard deviation, with the caveat that job complexity, applicant-versus-incumbent sampling and the exact construct all moderate the figure, frequently downward, and that Hispanic–White differences run somewhat smaller. High-validity predictors tend also to be the ones that generate the largest selection-rate gaps.

Third, and least comfortable for the industry: a validity coefficient is not a statement about the individual in front of you. A predictor that explains a meaningful slice of performance variance across a workforce will still mis-rank plenty of individuals — and it says nothing about the rejected candidate whose file cannot be reconstructed six months later.

The screen moved from GPA to skills — and from hiring to the whole lifecycle

Where assessment is used has changed as much as how much of it is used. In NACE’s Job Outlook 2026 survey of employers, collected between August and September 2025, 70% reported using skills-based hiring, up from 65% a year earlier, and 71% of those employers use the approach at least half the time. It shows up most often at interviewing (87%) and screening (65%). The substitution is explicit: in 2019, 73% of employers screened candidates on GPA. In 2026, 42% do — a fall of 42.5% (Pheonix Research calculation: (73 − 42) ÷ 73).

That shift decouples assessment volume from requisition volume. Openings can sit at a five-month low while the number of instruments run per hire keeps climbing, because every screen that used to be a grade-point threshold now has to be an instrument.

The second shift is where in the employee lifecycle the instrument sits. SHRM’s State of AI in HR 2026, drawn from a survey of 1,908 HR professionals in December 2025, finds AI concentrated in recruiting (27% of organisations), HR technology (21%), learning and development (17%) and employee experience (14%). The same measurement stack that screened a candidate is being pointed at an incumbent being considered for a move — which turns “pre-employment assessment” into “talent assessment” and moves the buyer from talent acquisition to the CHRO.

The rulebook moved in both directions at once

Europe had set the sharpest deadline in the world for hiring AI: high-risk obligations for Annex III systems — the category covering employment and worker management, including recruitment, selection and evaluation of candidates — were to apply from 2 August 2026. They no longer do. The AI Omnibus, proposed on 19 November 2025 as part of the digital omnibus package, entered into force on 27 July 2026 and moved the Annex III date to 2 December 2027, with high-risk AI embedded in regulated physical products following on 2 August 2028. The obligations were not withdrawn. The clock was reset.

In the United States, the federal layer largely went quiet. The EEOC’s 2023 technical assistance on Title VII adverse impact in AI-assisted hiring has been taken down from the agency’s site with no replacement issued, while its 2022 ADA guidance on hiring tools remains posted, and comprehensive federal AI-hiring legislation still does not exist.

The cities and states kept going, with results worth reading closely. New York City’s Local Law 144 remains the only regime requiring an independent bias audit — completed within the prior year, published in summary, with candidates notified at least ten business days before the tool is used on them. Penalties run to USD 500 for a first violation and USD 500 to USD 1,500 for each violation after that, and the city counts each day of non-compliant use and each missed notice as separate violations. Then, in December 2025, the New York State Comptroller audited that enforcement and called it “ineffective”: the Department of Consumer and Worker Protection had identified a single instance of noncompliance across 32 companies reviewed over two years, after receiving two complaints, while the comptroller’s own auditors found at least 17 instances of potential noncompliance among the same companies.

Illinois runs two laws — the Artificial Intelligence Video Interview Act, in force since 2020, and Public Act 103-0804, effective 1 January 2026 — though the implementing rules for the latter were withdrawn in June 2026. Maryland, Texas and Colorado have taken narrower or later-arriving approaches, and none requires an independent audit the way New York City does.

So the enforcement vacuum is being filled somewhere else. Mobley v. Workday, filed in the Northern District of California, is proceeding on claims that an employer’s use of algorithmic applicant-screening tools produced disparate impact by race under Title VII, by disability under the ADA and by age under the ADEA, with state FEHA claims added when three new named plaintiffs joined. The case has survived repeated rounds of dismissal, including a 22 June 2026 order granting in part and denying in part Workday’s motion to dismiss the third amended complaint. What is being litigated is not only whether the tool had an impact. It is who answers for it: the vendor whose model ranked the candidates, or the employer whose requisition set the rules.

The liability sits with the buyer, and the buyer isn’t measuring

Adverse impact is a property of an employer’s applicant flow, not of a test publisher’s norm sample. A vendor can hand over a validation study; only the employer can run the arithmetic on its own funnel — and under Title VII, as under the EU’s deployer-side obligations, that is where the duty lands.

Which makes the measurement gap inside HR conspicuous. In SHRM’s 2026 survey, 56% of HR professionals said they do not formally measure the success of their AI investments at all, and only 16% use a return-on-investment metric of their own. More than half of organisations — 54% — have no AI in HR and no plans to adopt any this year, while extra-large employers are far more likely than small ones to have implemented it (60% versus 33%). The pattern is uncomfortable: adoption concentrates in recruiting, the highest-liability practice area, and the ability to prove a tool works is thinnest exactly where the exposure is greatest.

The money is already flowing to the adjacent layer that can prove something. Pheonix Research’s Global Background Check Market report sizes background screening at USD 15.80 billion in 2025, growing to USD 34.60 billion by 2033 at a 10.3% compound rate — a 2.19× expansion (Pheonix Research calculation: 34.60 ÷ 15.80 = 2.19). First Advantage, the largest listed pure-play in that market, reported second-quarter 2026 revenue of USD 448.8 million, up 14.9% year on year, at a 28.6% adjusted EBITDA margin, and raised full-year guidance to USD 1.67–1.71 billion. Its stated demand strength sits in transportation and logistics, retail and e-commerce, industrials and manufacturing, and general staffing — high-volume hourly hiring, not white-collar requisitions.

The reason verification keeps growing while hiring does not is that the candidate has become the uncertain variable. When Checkr launched a fraud-detection suite in October 2026, it reported that in a beta analysis of customer resumes, 36% of inbound resumes triggered flags for further review, and that flags fired on 50% of software and AI engineering resumes against 23–27% in sales, marketing and operations. The company cited FBI warnings that North Korean IT workers are using stolen identities to obtain employment at US organisations, and Department of Justice prosecutions covering stolen American identities used to secure jobs at more than 100 US companies, in schemes estimated to generate about USD 800 million a year. An assessment administered to someone who is not the person on the application measures nothing at all.

What Pheonix Research covers here

Pheonix Research does not publish a market called “pre-employment assessment” or “talent assessment”, and any figure presented as one deserves the same interrogation a component figure does: licences only, or the whole HR technology stack; hiring-stage assessment only, or development and internal mobility folded in; seats, tests or hires as the unit. What the graph does carry is the structure this market sits inside — the Information Technology (IT) & Software industry where the platforms are built, the finance & business services segment that houses hiring-adjacent services, and the Global Background Check Market report covering the verification layer that assessment is increasingly bundled with.

What to watch

Three things will decide whether assessment becomes a priced, defensible category or stays a procurement line item nobody can justify in a deposition.

The first is whether the audit becomes a specification. Aerospace parts travel with a qualification file and buyers accept nothing without one. If enterprise procurement starts demanding the same from an assessment vendor — validation evidence and adverse-impact analysis run on the buyer’s own applicant flow, not a norm sample — the vendor that can produce it wins the deal, and the vendor that can only produce a dashboard does not.

The second is whether the identity layer swallows the assessment layer. Continuous verification across application, interview and pre-hire turns “who scored highest” into “who actually sat the test”. The screening companies are already selling that, and they have the volume data to show why it is needed.

The third is which regulator is left standing. With the EU’s high-risk deadline pushed to December 2027, Washington’s guidance withdrawn and the only municipal audit regime with real penalties found ineffective by its own state comptroller, the binding constraint on hiring assessment will not be a regulator. It will be the plaintiff’s bar and the procurement committee. The test was never the hard part. Proving you were entitled to use it is.

Frequently Asked Questions

What is pre-employment assessment?

Pre-employment assessment is the use of standardised instruments — ability and aptitude tests, job knowledge tests, work samples, structured interviews and personality inventories — to predict how a candidate will perform in a role before they are hired. It is distinct from background screening, which verifies facts rather than predicting performance, and from talent assessment used after hire for development, promotion and internal mobility.

How well do pre-employment tests actually predict job performance?

Less well than the industry claims, but better than the alternative. Sackett, Zhang, Berry and Lievens revisited the field's meta-analytic record in the Journal of Applied Psychology in 2022 and found that corrections for range restriction had systematically overcorrected, cutting mean validity estimates by 0.10 to 0.20 points. Structured interviews emerged as the top-ranked procedure. The authors' conclusion was that selection procedures remain useful, but predictor–criterion relationships are considerably lower than previously thought — and validity is a workforce-level statistic, not a verdict on an individual candidate.

Is AI used in hiring regulated?

Patchily. New York City's Local Law 144 requires an independent bias audit, publication of a summary and ten business days' candidate notice, with penalties of USD 500 for a first violation and USD 500–1,500 for each subsequent one — but a December 2025 New York State Comptroller audit found the city's enforcement ineffective. Illinois runs the AI Video Interview Act alongside Public Act 103-0804, effective 1 January 2026; Maryland, Texas and Colorado have narrower or later-arriving rules. There is no comprehensive US federal AI-hiring statute, and the EEOC's 2023 Title VII guidance on AI hiring tools has been withdrawn with no replacement. In the EU, the AI Omnibus that entered into force on 27 July 2026 moved high-risk Annex III obligations, including recruitment and selection systems, to 2 December 2027.

Why is employer spending on assessment and screening rising while hiring is flat?

Because the work per hire has gone up. Job seekers are sending about 30% more applications than before the pandemic, and roughly 99% of hiring managers report using AI somewhere in the hiring process. At the same time, employers have shifted screening criteria: 70% now use skills-based hiring, while GPA screening fell from 73% of employers in 2019 to 42% in 2026. More instruments are run per hire even when fewer people are hired.