From Resume Upload to Shortlist: The AI Screening Pipeline

In July 2026, one HyperVerge campaign ran 24,327 resumes through HireQwik across nine open roles. Only 1,074 of those resumes turned into a completed AI interview. One posting alone — a DL/ML Research Intern role — accounted for 22,761 of the resumes and 621 of the interviews. That gap between resumes-in and interviews-out isn’t a bug. It’s the pipeline doing its job before a recruiter ever opens the dashboard.

Most vendor pitches describe AI screening as a single black box: candidates go in, a shortlist comes out. That’s not how it actually works, and pretending otherwise makes it impossible to debug when something looks wrong. HireQwik’s pipeline has six distinct stages, each with its own logic, and each one exists to answer a different question about a candidate. This post walks through all six in order — what happens, why it happens there and not earlier or later, and where to look if you want to see it for yourself.

This is the overview version. Each stage links out to a deeper post on the specific mechanism behind it, because trying to explain the resume scorer’s relevance logic and the two-evaluator speech layer and the auto-decide threshold math in one article would shortchange all three. Think of this post as the map, not the terrain — read it first if you’re new to how HireQwik actually processes a candidate, then follow the links into whichever stage you need to understand in depth.

Stage one: a sheet row becomes a trigger

Every HireQwik campaign starts with a recruiter connecting a Google Sheet to a job description. The trigger flow watches that sheet for new candidate rows and, the moment one appears, enriches it and starts the pipeline — no manual upload, no batch export at the end of the week. We’ve written about the mechanics of this in Trigger-Based Hiring: How a Sheet Drop Becomes a Screening Campaign, including what happens when the sheet connection itself breaks.

This stage matters because it sets the clock. A candidate who applies at 9 a.m. is enriched and scored before lunch, not batched into next Monday’s review. If you’re trying to work out how long the whole process actually takes, this is where the timer starts.

Stage two: relevance-aware resume scoring

Once a row is picked up, the resume gets scored against that specific job description — not a generic template. HireQwik’s scoring hierarchy checks first for a JD’s own structured rubric; if none exists, a relevance-aware scorer runs, weighing whether a candidate’s actual skills and experience match the role rather than just counting years of experience or degree level. A resume with no relevant skills and no relevant experience is capped at a low score regardless of how long it is.

This is the stage where most of that July gap opens up. Not every one of the 24,327 resumes was a plausible fit for the role it landed against, and the scorer’s job is to say so before anyone’s time — human or AI — gets spent on a call. We go deeper on how this scoring actually behaves in Why Your AI Resume Scores All Cluster at 50-60%.

Stage three: the self-schedule invite

Candidates who clear the resume stage get an invite with a real calendar attachment — an RFC 5545 .ics file that opens natively in Outlook or Gmail, with its own 15-minute reminder. There’s no HireQwik-assigned slot the candidate has to work around; they pick a time that fits their day. We cover the design reasoning behind that choice in Self-Scheduling for Candidates.

Scheduling is also where a meaningful share of candidates quietly disappear — not because the AI rejected them, but because nobody followed up on a slot that was never booked. The reschedule-limit safety valve (three attempts within 72 hours of the first booking) exists specifically to keep that stage from drifting open-ended.

Stage four: phase-0 knockout, live on the call

The interview itself doesn’t open with generic small talk. It opens with the JD’s own disqualifying questions — phase-0 knockout, built from the same structured screener-build document that drives the resume rubric. If a candidate fails one, typically inside the first one to two minutes, the call ends there with a Reject verdict tagged “knockout” so HR can see exactly why in the review queue rather than guessing.

This stage is deliberately blunt. It exists to protect the two-evaluator scoring that follows it — there’s no reason to run a full communication assessment on a candidate who’s already disqualified on a hard requirement. Knockout Questions First has the full mechanics.

Stage five: two-evaluator scoring

For candidates who clear phase-0, the rest of the 15-20-minute conversation gets scored twice — once by an LLM evaluating what the candidate actually said, and once by an audio analysis layer evaluating how they said it: pace, filler and hesitation rate, pronunciation, CEFR fluency level. Both evaluators have to agree before HireQwik recommends a Reject. The blend is deliberately asymmetric: strong speech signal can lift a borderline transcript from Reject to Hold, but it never demotes a candidate who scored well on substance.

This is the mechanism most competing tools skip entirely — most AI screening products score communication from the transcript alone, missing everything that happens between the words. The Two-Evaluator Check walks through why that gap matters and what it catches.

Stage six: auto-decide, then the dashboard

The interview produces a verdict — Strong Go, Go, On Hold, or No Go — and that’s where auto-decide comes in. Recruiters can set per-JD thresholds so HireQwik auto-rejects candidates who land below one score and auto-fans-out candidates who clear another, leaving HR to review only the middle band that genuinely needs a human judgment call. Everything else, verdicts and all, lands in the dashboard at app.hireqwik.in/dashboard/hr, where the Score/Decision filter on the Needs Review tab lets a recruiter narrow the queue by band instead of scrolling a flat list.

Inside the HireQwik Dashboard covers what that review screen actually looks like. And once a candidate clears it with a Strong Go, the handoff to the next hiring step is its own set of decisions — we cover that separately in What Happens After a Strong Go.

What happens to each of the four verdicts

A finished interview doesn’t just get a pass or fail — it lands in one of four bands, and auto-decide treats each one differently. A Strong Go above the recruiter’s upper threshold can auto-fan-out straight to the next hiring step without waiting on a human click. A No Go below the lower threshold auto-rejects the same way, and nothing about how it got there is hidden afterward — the full record stays there for anyone who wants to check the reasoning later. Go and On Hold sit in the middle band recruiters actually opted into reviewing — the Needs Review queue on the dashboard — because the two evaluators landed too close to call it confidently either direction.

That middle band is deliberately small relative to the whole funnel. Of the 1,074 completed interviews in that July HyperVerge campaign, the goal of stages two through five isn’t to push every candidate into a human queue — it’s to resolve as many as the rubric can resolve on its own, so the recruiter’s attention goes to the genuinely ambiguous cases instead of re-litigating decisions the pipeline already made confidently.

Running all six stages at campus-drive scale

None of this matters if it only works for a trickle of applicants. The pipeline was built around a specific stress test: a single evening where 3,000 candidates were screened in about two hours, start to finish, resume to verdict. That’s only possible because stages three through five run concurrently, not one candidate at a time — by default, up to 20 candidates can be on a live call at once within any given quarter-hour window, so a 3,000-candidate evening is really a long series of 20-wide batches, each one taking roughly as long as a single call.

Compare that to a manual phone screen: 10-15 minutes per candidate, one recruiter at a time, at a market rate of roughly ₹85-150 per call for outsourced screening. Run the same 3,000-candidate volume through that model and you’re looking at weeks of recruiter hours, not a single evening — which is where the pilot’s 89% reduction in HR time per candidate against a manual phone screen actually comes from. It isn’t that any one stage above is dramatically smarter than a human recruiter; it’s that all six stages run in parallel across thousands of candidates at once, and a human reviewer only has to look at the ones the pipeline couldn’t confidently resolve on its own.

Why the gap between stages is the point

SHRM’s State of AI in HR 2026 report found that a meaningful share of organizations using automation in hiring said their tools had overlooked or screened out qualified applicants — a real risk with any automated filter, ours included. The reason HireQwik’s pipeline is six visible stages instead of one opaque score is so a recruiter can tell which stage did the filtering and why, instead of trusting a single number.

Most published recruiting-funnel research backs up why that matters: across more than ten million applications tracked in one 2025 benchmarking study, only about 3% of applicants ever reached an interview, and under 1% were hired. If your funnel loses candidates at that rate anyway, the honest question isn’t whether to filter — it’s whether you can see where the filtering happens. That’s what stage-by-stage visibility buys you, and it’s the reason we built the pipeline this way instead of shipping a single “match score” and calling it done.

Which stage rejects the most candidates?

In practice, relevance-aware resume scoring at stage two and phase-0 knockout at stage four do most of the work — one before a candidate is ever invited to a call, the other inside the first two minutes of it. Two-evaluator scoring at stage five rejects far fewer people outright, because its asymmetric blend is built to rescue borderline candidates rather than filter them out.

Does a candidate ever skip a stage?

No — every candidate who clears the resume score goes through the same six stages in the same order, whether the role is a campus internship or a lateral senior hire. What changes JD to JD is the rubric each stage checks against, not whether the stage runs at all.

If you’re setting up your first campaign and want the practical version of all six stages, Setting Up Your First AI Screening Campaign in HireQwik is the step-by-step companion to this overview. If a specific candidate looks stuck somewhere in the middle of it, Why “It’s in the Pipeline” Isn’t Good Enough is the debugging guide.