Ninety-day failures.

The candidate cleared every bar. The interviews ran long because people enjoyed them, the references were warm, the offer was accepted within a day. Four months later the resignation arrived, and the exit interview produced the sentence every talent leader has heard more than once: it wasn't what I expected. Most organizations file that exit under onboarding and move on. The turnover literature files it somewhere less comfortable: early attrition is a selection outcome few hiring processes ever test for, and the sentence in the exit interview was measurable before the offer letter went out.

The evidence behind that claim is specific. Person–job fit, a construct assessable before hire, correlates .56 with job satisfaction and −.46 with the intention to quit (Kristof-Brown, Zimmerman, & Johnson, 2005), and quit intentions sit closest to the act of leaving itself and predict it best (Griffeth, Hom, & Gaertner, 2000). Realistic job previews produce consistent if modest reductions in early departures by removing the surprise (Phillips, 1998). These are ordinary measurements. Most hiring processes simply decline to take them.

A century of turnover research (Hom, Lee, Shaw, & Hausknecht, 2017) has matured into a working hierarchy of predictors, and the hierarchy has an inconvenient shape: its best predictors form inside the job, while hiring decisions are made outside it. This article is about the middle of that hierarchy, the variables that exist before the offer and feed everything downstream. An organization that measures person–job fit and previews the role truthfully buys down its early-exit rate before day one. One that does not pays the full bad-hire bill on people who could do the job but were never going to stay in it.

Early attrition is a selection outcome wearing an HR label

The first year of employment operates as its own regime, not a shorter version of the years that follow. Wanous (1992) gave its signature experience a name, entry shock: a newcomer arrives carrying expectations assembled from job postings, interviews, and recruiter enthusiasm, and the first months are where those expectations collide with the job as it actually is. Turnover later in tenure tends to follow what accumulates: stalled progression, a reorganized team, an outside offer. Turnover in the first year runs through a channel the later years do not have, the gap between what was promised and what turned out to be true.

That gap has a formal name, the met-expectations framework, and a measured consequence. Reviewing the newcomer studies meta-analytically, Wanous, Poland, Premack, and Davis (1992) found that the degree to which early experience matched prior expectations related meaningfully to satisfaction, to commitment, and to the intention to remain. The newcomer who says the job was not as described is not offering an excuse. They are citing, almost verbatim, the variable the literature identifies.

The filing error matters because it routes the fix to the wrong team. Onboarding can soften entry shock; it cannot retroactively create fit that was never assessed, and it cannot unmake expectations that a recruiting funnel spent months inflating. When the mismatch existed at offer stage, the failure belongs to selection, and as long as early attrition sits in the onboarding column, the same screen keeps producing the same exits with a clean conscience.

The stakes are the same arithmetic as any failed hire. An early exit re-runs the entire sequence of search costs, vacancy, ramp time, and manager hours; the full accounting appears in our companion analysis of what a bad hire actually costs. Early attrition runs that bill on a compressed schedule and adds a detail the averages miss: the person leaving has usually not yet repaid any of it.

What predicts leaving runs in a hierarchy, and demographics sit at the bottom

Predicting employee turnover is one of applied psychology's oldest projects, old enough that its first hundred years have been formally reviewed (Hom et al., 2017). A key synthesis is Griffeth, Hom, and Gaertner (2000), a meta-analysis in the Journal of Management that ordered the field's accumulated predictors of actual leaving and found a consistent hierarchy. At the top sit the proximal attitudes, the ones nearest the decision itself: the stated intention to quit, organizational commitment, and job satisfaction. People telegraph their exits, to themselves first.

The finding can sound circular, since people who intend to leave then leave. Its value is directional: if intentions carry the exits, then anything that shapes intentions in the first months is upstream of every resignation letter, and the practical question becomes which upstream variables can be known in advance. That is the question selection is positioned to answer.

At the bottom of the same ordering sit demographic characteristics, among the weakest correlates of leaving the meta-analysis examined. That placement deserves exactly one sentence, because it is where informal practice quietly operates: inferring "stability" from a candidate's age, family situation, or the date arithmetic on a résumé is empirically weak and legally hazardous at the same time, and nothing else in this article depends on it.

The hierarchy contains a trap for anyone who wants to act on it. Its strongest predictors are attitudes that exist only inside the job; satisfaction and commitment cannot be measured at offer stage because they have not formed yet. Selection cannot screen on the top tier directly. What it can do is measure the tier below, the variables shown in Figure 1 that feed the attitudes: person–job fit, the expectations a candidate carries in, and the temperament they bring with them.

Read as a supply chain rather than a ranking, Figure 1 turns from discouraging to useful. The tiers are not competing predictors; the middle one feeds the top one, which means an organization that cannot select for satisfaction can still select for the conditions under which satisfaction tends to form. The best-quantified of those conditions is fit.

Fit predicts the attitudes that predict leaving

The reference estimates come from Kristof-Brown, Zimmerman, and Johnson (2005), a meta-analysis in Personnel Psychology of the consequences of fit at work, careful to keep the varieties of fit distinct. Person–organization fit is alignment with a company's values; person–group and person–supervisor fit describe the immediate humans. Person–job fit is narrower and harder-edged: the match between a person and the work itself, with two operational faces. Demands–abilities fit asks whether the person can do what the role requires; needs–supplies fit asks whether the role supplies what the person needs (the pace, autonomy, variety, and recognition that make the work sustainable for that particular individual).

The distinctions are not academic bookkeeping. The meta-analysis finds that the forms of fit carry distinct outcome patterns, and interviews that chase "culture fit" are, at best, probing the person–organization form: whether the candidate shares the company's values and style. For the early-exit question, the coefficients that matter belong to the person–job form. A candidate can admire the company, like the team, and still be mismatched with the daily work itself, and the mismatch with the work is the one that surfaces in the first months, when the work is nearly all there is.

For person–job fit the corrected correlations, estimates adjusted for measurement artifacts, are .56 with job satisfaction, −.46 with the intention to quit, and .20 with overall job performance. The shape of that trio is this article's argument in miniature. Fit is a modest performance signal, and the .20 should be said plainly: measuring fit does not replace measuring ability. It is a strong attitude signal, and attitudes are where exits are manufactured.

Figure 2 draws the trio to scale, and the −.46 deserves the longest look. It says that a variable measurable before the offer runs strongly against the specific attitude that most directly precedes leaving. A screen built entirely on ability captures the .20 and is blind to the channel that governs staying. That is how an organization ends up with the hire from the opening scene: fully capable, accurately assessed on capability, and gone in month four.

Fit is measurable pre-offer because both of its faces decompose into assessable parts. The demands side is a profiling exercise: specify what the role actually requires (sustained unsupervised focus, tolerance for interruption, constant client contact, tight procedural discipline) and measure the candidate's standing on each. The supplies side is symmetric: characterize what the role offers (its pace, its autonomy, its ceiling, its social texture) and compare it with what the candidate's profile indicates they need. Neither step requires the candidate to have held the job. That is what makes person–job fit a selection construct rather than a post-exit explanation: it is a distance between two profiles, and both can be in hand before the offer.

Previews trade surprise for retention, at almost no cost

If fit closes the gap between person and job, the realistic job preview closes the gap between expectation and job. The intervention is almost embarrassingly simple: before the candidate accepts, show them the role as it is, unglamorous parts included, rather than the role as the recruiting funnel has rendered it. Phillips (1998), meta-analyzing realistic job preview studies across organizational outcomes in the Academy of Management Journal, found that previews produced consistent, modest reductions in turnover and improvements in met expectations.

The finding turns on both of its adjectives, consistent and modest. The honest framing is that a preview trims early turnover rather than transforming it, and anyone selling previews as a retention cure has overrun the evidence. But the cost side is what makes the lever unusual: a realistic preview requires no new instrument, no vendor, and no budget line. It is a scheduling decision and a script.

The mechanism runs through the same variable as entry shock. A preview cannot make a hard job easy; what it changes is the expectation the newcomer carries into month one, so that reality arrives as confirmation rather than betrayal. Part of the effect plausibly operates even earlier, when a candidate hears the preview and withdraws: a mismatch surfaced during recruitment rather than on payroll. And previews compound with measurement, because met expectations feed the same attitudes that fit feeds (Wanous et al., 1992); the two levers belong in the same pre-offer stage, not in different departments.

One failure mode is worth naming, because it is common. A preview that is really an extended pitch, the hard parts sanded off, changes nothing, since the variable it must move is the accuracy of expectations, not their warmth. The discipline that produces a working preview is the same discipline that produces a fit assessment: a truthful profile of the role's demands. Organizations that build the profile once can spend it twice.

Temperament tilts the decision, and not always through attitudes

Personality earns its place in the retention conversation, provided the claim is sized correctly. Zimmerman (2008), building a meta-analytic path model in Personnel Psychology, found that emotional stability and conscientiousness predicted turnover decisions with correlations in the .1–.2 range. Alone, that is a modest signal, nothing an organization should decide on. Measured alongside fit and ability, it adds an increment of prediction inside a battery the organization is already running.

The path finding matters more than the coefficients. Emotional stability's link to leaving was partly direct rather than fully routed through job attitudes: some of the propensity to leave travels with the person, not with the job's measured satisfaction. For early attrition specifically that is worth pausing on, because a direct path is exactly the kind that engagement surveys, which watch the attitude route, will not see coming.

Domain scores are averages, though, and turnover is a mechanism-level question. Which strand of "emotional stability" carries the risk plausibly differs by role: resilience to rejection matters in one seat, composure under interruption in another. The case for measuring personality at the facet level rather than the domain average is made in full in our companion analysis of facet-level personality measurement, and it applies here with the same force: the sharper the resolution, the more precisely a temperament signal can be read against a particular role's frictions.

What compounding mismatch looks like across the first year

Correlations describe cohorts, and cohorts are where modest-sounding numbers become visible. Figure 3 plots 12 months of retention for two illustrative cohorts, one hired with strong person–job fit and one without. The values are constructed for exposition, not observed: the high-fit cohort retains 97% of its members at month one, 95% at month three, 92% at month six, and 89% at month 12, while the low-fit cohort passes the same posts at 93%, 85%, 76%, and 68%. What the construction preserves is direction, the pattern the −.46 implies: lower fit, stronger quit intentions, more exits.

Two features of the picture generalize beyond its invented values. The curves separate fastest in the first three months, exactly where entry shock operates and before most managers have formed a view of the newcomer. And the gap never closes: a cohort that bleeds early does not drift back, because the members it lost are gone and their replacements restart clocks of their own. New hire turnover is not a rough patch a cohort grows out of; it is a loss that was priced in at selection.

The practical instruction hiding in Figure 3 is that an organization should possess this figure about itself, with real data where the illustration sits. Retention curves by hiring cohort, linked to pre-hire assessment scores, convert early attrition from an anecdote into an instrument: when a curve bends early, the organization can see which measured variables the leavers shared, and the screen that passed them can be adjusted rather than defended.

Assembled, the sections of this article form the timeline in Figure 4. Fit and expectations are set, measured or not, before the offer. Entry shock converts unmeasured mismatch into disappointment; disappointment forms the quit intention; the intention, the most proximal predictor on the record, becomes the exit that HR logs in month four. The chain runs left to right. The leverage runs right to left, and it terminates in the one segment where measurement is standardized, comparable across candidates, and cheap: before the offer.

Run new hire retention as a selection KPI

For organizations willing to treat the first year as an output of the hiring process rather than a run of luck, the evidence converts into four operating rules.

• Measure fit constructs before the offer. Profile the role's actual demands and supplies, assess candidates on the matching constructs, and score the distance, identically for every candidate. This replaces the inference habit the turnover literature ranks weakest: reading "stability" off demographics or the shape of a résumé's job dates. A sample scored fit readout shows what this looks like in practice.
• Preview truthfully, late in the funnel. Give final-stage candidates the unvarnished version of the role while withdrawing is still free. The evidence promises modest, reliable gains (Phillips, 1998); the cost is one conversation and the nerve to describe the job accurately.
• Instrument the cohort. Track retention curves by hiring cohort, joined to pre-hire scores, so that exits trace to causes. Without that join, every early exit is an anecdote, every explanation is plausible, and the screen never learns.
• Treat regretted 90-day exits as a selection KPI. Count them where hiring decisions are governed, not only where offboarding is processed. A regretted early exit is direct evidence that the screen passed someone the role could not hold.

The fourth rule earns its qualifier. "Regretted" is doing the sorting: an early exit the organization would accept again, a misconduct termination or a mutual quick goodbye, is a different event from losing a capable hire to a mismatch nobody measured. The KPI counts the second kind, because that is the kind a better screen would have prevented, and it belongs on the same review cadence as time-to-fill, in front of the people who own the screen.

The budget logic is the standard utility argument: per-hire gains multiplied across hiring volume compound into sums that dwarf assessment costs, and the framework for that arithmetic is set out in our companion analysis of the return on selection investment. Retention enters that calculation twice, once as replacement cost avoided and once as longer tenure over which every other per-hire gain accrues. Fit, which correlates with satisfaction and performance at once, moves both entries.

This is administrative work, not research work. High-volume employers run the natural experiment every quarter whether they look at it or not; enterprise-scale deployments differ mainly in whether anyone joins the pre-offer measurements to the post-offer curve. That join is the whole method: measure the middle tier of Figure 1 before the offer, then let the retention curve grade the screen. Early attrition stops being a mystery at roughly the moment it starts being a metric.