Predictive Hiring: What Actually Predicts Performance in Frontline Roles

What is predictive hiring?

Predictive hiring is forecasting job performance from evidence collected before the hire.

The phrase is used in two ways. In the research tradition, it means selection methods such as interviews, tests and work samples whose link with later job performance has been measured across many studies. In software marketing, it often means a model trained on data about past hires that gives each new applicant a fit percentage or places them in a sorted list.

The difference between the two is how much of the evidence you can see. Structured interviewing has decades of published studies behind it. A fit percentage from a model you cannot inspect has whatever evidence its maker chooses to publish.

This matters most in frontline hiring, where volumes are high and many applicants have little or no CV to send. A tool that promises to find the best 20 among 300 applicants is attractive. What should such a promise rest on?

What does it mean when a method predicts performance?

It means there is a measured link between the method and later job results.

Researchers express that link as a correlation, often called validity: a number between 0 (no link) and 1 (a perfect one). It describes how well a method predicts across many hires and says little about any one person: a link of 0.42 does not mean 42 percent correct decisions. In hiring assessment validity research, it means that people who did better on the method tended, on average, to do better in the job.

Two details decide whether a published number means anything for you.

What counted as performance. The large research summaries use measures of overall job performance. The researchers point out that it is often unclear what those measures covered (Sackett and colleagues, 2023). For a frontline role, the outcomes you care about may be different: turning up for the first shift, working safely, reaching the expected pick rate after training, finishing the assignment, still being there after probation. A method that predicts overall performance has not automatically been shown to predict any of these.

The spread around the average. The method with the best average in the revised research, the structured interview, comes out at 0.42. Across studies, though, the likely range runs from 0.18 to 0.66. The authors suggest reading it as ".42, plus or minus .24" (Sackett and colleagues, 2023). Treat an average as a starting point for your own site.

Which hiring methods are the best predictors of job performance?

Methods built around the specific job predict best; general signals trail.

In 2022, Paul Sackett and colleagues revisited the research summaries that HR had relied on since 1998. They found that a statistical correction had often been applied too generously, so most of the leading methods predict less well than people thought, with averages 0.10 to 0.20 lower. The order changed too. Structured interviews came out on top, and general cognitive ability, long treated as the best single predictor, fell from 0.51 to 0.31 in their revision. A later analysis of samples from this century, reported in their 2023 paper, put it at 0.23 (Sackett and colleagues, 2022 and 2023).

Sources: Sackett and colleagues (2022, 2023), also summarised in SIOP's publication TIP; experience, Van Iddekinge and colleagues (2019), a review of 81 samples.

We look at years of experience in a separate blog, and at structured against unstructured interviews in a separate article.

Why the best predictors are the hardest to use at frontline volume

The researchers draw a conclusion that matters for frontline hiring. Several methods near the top of the list only work where applicants already have job specific training, as with knowledge tests and work samples. Structured interviews are hard to run in high volume screening. Biodata needs long development and large validation samples (Sackett and colleagues, 2023).

When 300 people apply for 20 warehouse jobs, nobody has time to interview each of them in a structured way. So the process falls back on what is fast: a glance at a CV, a filter on years, a phone call that goes differently every time, or a number produced by software. Speed wins by default, and the evidence loses.

Our view: put the effort into collecting the stronger kinds of evidence from every applicant, at volume, and leave the judgement with a person. A cleverer guess about who will perform is a weaker goal than better evidence for the people who decide.

See how the AI Interviewer asks every applicant for a role the same job related questions, on their phone, whenever it suits them.

Can predictive hiring tools tell you who will perform well?

Some may, but a study of vendor claims found little published evidence.

A research team studied the public claims of 18 vendors of algorithmic assessments used to screen job applicants. Their overview listed validation information for only 5 of the 18, and they noted that vendors' websites did not make clear whether the models were validated. Eight vendors built their assessments from data about the client's own current and past employees (Raghavan and colleagues, 2020).

That last design deserves attention. A model trained on your current workforce learns what your current workforce looks like. The researchers point out that this skews the search towards candidates who resemble people already hired. If past performance ratings carried bias, or past hiring left some groups out, the model learns that too. It also sees outcomes only for the people you hired, never for those you turned away.

One caveat: the study reflects what vendors published in April 2019. Read it as a picture of how little evidence buyers were offered, not as a verdict on any current product.

A strong test looks different. In a field experiment, researchers Brian Jabarian and Luca Henkel randomly assigned 70,000 applicants for entry level customer service jobs in the Philippines to an interview with a human recruiter or with an AI voice agent. Human recruiters reviewed the interviews and made the hiring decisions. Applicants interviewed by AI were 12 percent more likely to receive an offer, and the gains carried through to job starts and retention, with no drop in the productivity of the people hired. The authors link the result to interviews that were more structured and consistent (Jabarian and Henkel, working paper, July 2026).

That study tested how evidence was collected, not a model that predicts who will perform, and one study in one setting does not settle the question for every role. It does show the kind of evidence worth asking for: random assignment, real applicants and outcomes measured after the hire.

How do you test a vendor's prediction claim?

Ask for the study, the outcome, the sample and the spread.

Professional guidance sets the bar. In January 2023 the Society for Industrial and Organizational Psychology (SIOP) published recommendations for AI based assessments. They say such assessments should predict future job performance or other relevant outcomes accurately, produce consistent results that reflect job related characteristics, be fair and unbiased, and have every step of development and scoring documented so it can be verified and audited (SIOP, 2023). In the Netherlands, COTAN, the test committee of the Netherlands Institute of Psychologists (NIP), assesses the quality of psychological tests on seven criteria, including reliability, construct validity and criterion validity (NJi). A tool that cannot show evidence against standards like these is asking you to take its word for it.

Use these questions in any demo, including ours. For the rest of a buying decision, see how to choose AI recruitment software.

How can you check what predicts performance in your own roles?

Compare outcomes you already record against the evidence you collected.

Published averages tell you where to start. Your own data tells you whether a method works on your sites. The researchers discuss combining published estimates with a local study of the method you use (Sackett and colleagues, 2023). You do not need a statistician for a first look.

Illustrative example. The figures below are invented to show the method. They are not data from any employer or agency.

A staffing agency fills order picker roles at one client warehouse. Over six months it places 120 people. For each placement, it noted at the time whether every requirement of the role was backed by evidence before the start (a checked certificate, a structured answer about shift times, a confirmed start date) or whether one or more rested only on the candidate's word. It also records who is still working after 13 weeks.

The gap is 16 percentage points (52 of 70 against 29 of 50). Before anyone acts on it, run four checks:

Repeat the comparison every quarter and per client site. If the gap holds, you have local evidence; if it disappears, that is useful too. Our guide to measuring quality of hire in frontline recruitment lists measures you can track.

What does this mean for frontline employers and staffing agencies?

Spend your effort on job related evidence, and keep decisions human.

What the research cannot tell you. The research summaries combine many studies from different jobs and settings, measure overall job performance, and are not specific to the Netherlands or to frontline work. Use the published numbers as direction, and your own outcomes as the test.

So, does predictive hiring work?

Yes, modestly, when the prediction rests on job related evidence.

Job specific methods predict better than years on a CV, and even the best of them improve the odds across many hires rather than foretell what one person will do. The methods that predict best are the hardest to run at volume. Collect better evidence from every applicant, ask every tool, including ours, what its claims rest on, check against your own outcomes, and let a person decide.

How Radius helps

Radius Hire is an AI hiring platform for frontline roles. It collects job related evidence from every applicant and shows it to a recruiter, who makes every hiring decision.

Radius is new, so this article explains how the product works rather than claiming results. Radius has not published validity results, and we do not claim that its output predicts job performance. It is built to give the people who decide better evidence, with a record of what each decision rested on. Read more in Why We Count Evidence and Refuse to Rank Candidates.