The AI Interview Odds Nobody Publishes

Why Nobody Publishes a Pass Rate

No employer publishes a pass rate for the AI interview stage that a candidate could read as their own odds. The figure would be a per-role, per-quarter number that moves with the size of the applicant pool and the number of seats open, so it would be stale by the time anyone read it. It was never a public number.

So the gap gets filled by whoever is willing to type a number, which is how the resume myth spread. The claim that applicant tracking systems automatically reject three quarters of resumes before a person sees them circulates with no sample, no denominator, no named employer and no year attached. It is still repeated in career advice. A pass rate for the AI interview round is the same shape of claim, specific enough to sound measured and attached to nothing.

Employers often cannot answer the question either.

The share who clear a screening stage lives inside an applicant tracking system, under a stage name that team invented, and it moves whenever the role changes or the pool grows. A recruiter may be able to tell you last month's figure for one requisition. Ask about the round in general and you are asking about an average across roles, tools and quarters that the company may never have calculated.

Treat any percentage you find for this stage the way you would treat a price with no seller attached. Ask where it was measured, across how many candidates, at which employer, and in what year. If the page cannot say, the number is decoration.

Two kinds of number survive that test, and neither is about your employer.

What the Traceable Numbers Cover

The applicant tracking numbers describe conventional hiring pipelines rather than AI interviews. Ashby, an applicant tracking vendor, reports across more than 54 million applications and 93,000 jobs that recruiter screens pass roughly 35% of the candidates who reach them ¹. Later stages convert far higher in the same data, at 95% after the onsite and 81% at offer ¹.

Every part of that carries weight. These are stage-to-stage rates among candidates who already reached the stage, so they do not multiply into an end-to-end chance from application, and the heaviest filtering happens earlier still, at application review, which is not one of these numbers ¹. Stage names are configured by each customer inside the software, so a recruiter screen at one company is not the same object as a recruiter screen at another.

And the 95% is not evidence that onsites are easy ¹. It means the decision has effectively been made before the onsite ends.

The sample is Ashby's own customers, which skew toward venture-backed technology companies.

A second figure from the same vendor gets quoted as if it were an applicant's odds. Across more than 109 million applications, the share that resulted in an interview fell from about 7-8% in 2021 to between 3.6% and 4.7% by 2026, depending on the role type ². That is a ratio of two organization-level totals, applications in one column and interviews in the other. It is not the probability that your application becomes an interview.

The same report names automated and fraudulent submissions as one driver of the rising volume, without saying how much of it they account for ².

These stage rates have an employer's companion piece, what conversion rate each stage of a hiring funnel should have, written for the person setting the cut line rather than the person facing it.

What One Randomised Experiment Found

One randomised field experiment reports outcomes after an AI interview rather than a pass rate for the stage itself. It assigned 37,000 applicants for a junior developer role to either a resume-screening pipeline or a pipeline that added an AI-run structured video interview first ³. Both tracks then met the same final human interview, with the interviewers blind to how each candidate had been shortlisted ³.

In that final round, 54% of the candidates shortlisted through the AI track passed, against 34% of those shortlisted from resumes alone ³. The gap is an average treatment effect of 20 percentage points, with a standard error of 12 ³. A standard error that size on an effect that size is wide, which makes the finding directional rather than precise, and it should never be quoted as a flat twenty points.

One employer, one country, one job family, and a preprint rather than peer-reviewed work.

A later version of the same paper reports two experiments, 4,323 applicants across four engineering roles and 34,493 for one junior frontend role, and gives the effect across both runs as 17.5 to 20 points, with standard errors of 8.5 and 11.8 ⁴. Two of that paper's four authors are affiliated with the recruitment platform that ran the experiments and sells the AI interview ⁴. That is a commercial interest in the result, and it belongs beside the number rather than in a footnote.

What both papers measure is the final round, not a pass rate for the interview stage itself.

One finding inside that work is more useful to a candidate than either figure. The system tended to advance younger applicants with less experience and fewer advanced credentials ³. In this experiment, the AI stage advanced people with thinner resumes on paper.

That is a description of who advanced, not an endorsement of the design.

Where the Stage Is Actually Lost

In the one experiment that measured it, the stage was lost far more often by not being finished than by being failed. Of 23,000 candidates invited to the AI interview there, 24% completed it ⁵. The later version of the same paper reports a junior frontend run with 34,493 applicants and puts completion among those invited at 25.07% ⁴. Three quarters of the invited pool never produced an answer to judge.

Non-completion is not the same as refusal. It mixes the candidate who objects on principle with the one who lost the email, hit a dead link, had no working camera, ran out of time before the deadline, or took another job that week. The study does not separate them, and neither should anyone quoting the figure ⁵. What the number does show is where the attrition sits.

In that pipeline, most invited candidates never finished, whatever the reason.

The candidates who declined were slightly older and more experienced, and their later labour-market outcomes, read from LinkedIn profiles, did not differ statistically from those of the people who finished ⁵. That is an absence of a detected difference in one sample, not proof that declining is free. Inside that pipeline, not finishing meant not advancing, so the decision worth making is a deliberate one rather than a default.

HireVue's candidate FAQ describes its interviews and assessments as first steps that, for the vast majority of jobs, help recruiters make the first cut, after which the top group moves on to person-to-person interviews ⁶. That is the vendor's account of its intended design, not a promise about how any single employer has configured it.

It is still the kind of page worth reading before the interview rather than after the rejection.

If you are weighing whether to skip the format entirely, turning down an AI interview and still being considered sets out what that costs. Whether a human ever watches the recording is the other question worth settling first, and it is a question about one employer's process rather than about the vendor's design.

A recording nobody finished has no pass rate.

See how it works >