Set Resume Auto-Decide Bands Before Voice Invites Fire

A recruiter on one of our early campus pilots asked how to set resume auto-decide AI interview bands, then parked the fanout line at 40% and left for the night. By morning the booking calendar was packed with people she would never have called by hand. The queue looked productive. The shortlist quality did not. That night is why this calibration guide exists. Placement of the two thresholds is the contract between paper signal and microphone time.

This is a how-to for the resume gate only. It is not a second essay on post-call accept/reject bands. Those live after the conversation and use a different label set. We will contrast the two once, then stay on the pre-call side: Score-Decision ranges, trigger thresholds, relevance-aware scoring, the Needs Review filter, calibration steps, and the two mis-sets that cause most of the pain. If you want the full post-call middle-band logic, start with auto-decide AI screening accept/reject bands.

Resume Score-Decision bands are not post-call verdicts

HireQwik shows four resume-match bands on the pre-call queue:

• Strong Go: 90-100%
• Go: 75-89%
• Maybe: 60-74%
• Reject: 0-59%

Those percentages are the resume pass against the job description. Nobody has spoken yet. The labels answer one question: does this paper look close enough to invite into a voice screen?

After the call, the product uses a different four: Strong Go, Go, On Hold, No Go. Those come from the structured conversation, communication, and role fit. We kept “Strong Go” and “Go” on both sides because the intent is the same (this person looks ready to move), but the evidence is not. Treating a 92% resume match like a post-interview Strong Go is a category error the system never claimed to support.

If you want the full post-call middle-band logic (auto-reject floor, auto-advance ceiling, human review in between), that sibling already covers it. This article stops at the invite.

Trigger auto-decide: floor, ceiling, and the invite path

Resume-side auto-decide shipped as a per-JD opt-in on the trigger path. Recruiters set two match-score thresholds on the job:

1. Auto-reject floor. Below this score, the candidate does not get a voice invite. They land as Reject on the resume side.
2. Auto-fanout ceiling (invite bar). At or above this score, the trigger path emails a self-schedule link without HR clicking Invite per row.

Everyone between those two lines sits in Needs Review. That middle band is where a human still decides whether the resume deserves a voice slot.

The invite path itself is short once the sheet is connected: new row lands, resume scores against the JD, auto-decide applies, and cleared candidates get the booking email. The full handoff from shortlist into autonomous voice is the sibling pillar piece on resume shortlist to voice screening. The mechanics of the email after a clear score sit in auto invite after resume score. Here the only dial that matters is where you place the two thresholds relative to the four Score-Decision bands.

A practical mapping most campus roles start from:

That table is a starting point, not law. A high-volume support campus drive and a 40-seat specialist drive should not share the same floor. Per-JD thresholds exist so they do not have to.

Relevance-aware scoring: why bands need a real spread first

Bands only work if scores spread. If every resume lands between 50% and 60%, no floor or ceiling can save you. That compression is the relevance-blind failure mode: the engine rewards total years and degree level without asking whether any of it fits this role. Non-skill factors drag almost everyone into the same middle strip, so Reject never fires cleanly and Go never separates from Maybe.

We fixed that with relevance-aware resume scoring: irrelevant experience contributes close to nothing, and a profile with no relevant skills and no relevant experience caps low instead of floating to the middle. The hierarchy is JD rubric first, relevance-aware scorer next, relevance-gated keyword matcher last. The deeper root-cause write-up is why AI resume scores cluster at 50-60%.

Before you argue about 72% versus 75%, open the last campaign’s score distribution. If the whole field sits in a ten-point band, fix scoring (or the one-row rubric that forces 33/67/100) before you touch auto-decide. Thresholds cannot invent signal the scorer never produced.

Employment selection tools are supposed to distinguish candidates on a job-relevant basis before they include or exclude anyone. That is the same bar the U.S. EEOC guidance on employment tests has held for years. A compressed 50-60% cloud fails that bar long before legal asks a question.

Needs Review Score/Decision filter: triage the middle without scrolling

Once auto-decide is on, most of the recruiter’s morning lives in Needs Review: the Maybe pile, plus anything you chose not to auto-fanout. Scrolling a sorted score column and drawing the line by eye is how teams “set bands” before we shipped a real filter.

The Score/Decision filter on Needs Review groups the pre-call queue by those four resume bands (Strong Go 90-100%, Go 75-89%, Maybe 60-74%, Reject 0-59%). It sits next to Role and Source, scopes bulk select to the filtered list, and clears when you leave the tab. It is not the post-call Verdict filter. Same toolbar family, different stage.

Use it like this after a sheet drop overnight:

1. Open Needs Review for the JD.
2. Select Strong Go + Go. Confirm the count matches what auto-fanout should already have invited (or invite the ones you held back).
3. Switch to Maybe alone. That is the judgment queue.
4. Spot-check Reject only when you distrust the floor (odd low scores on known-good profiles).

Campus teams that leave every “shortlisted” resume sitting until someone remembers to click Invite are living in the campus shortlist voice gap. Bands plus the filter close that gap without turning Maybe into an unsupervised invite blast.

Calibration steps: set the bands on one JD before you scale

Do not invent thresholds from a vendor default and push them across every open role. Calibrate on one live JD with a finished score distribution.

Step 1: Confirm the distribution is usable. Count distinct scores. Look at the lowest score. If everything is 100% or nothing falls below ~33%, stop and fix the rubric. If everything is 52-61%, stop and fix relevance. Only then set bands.

Step 2: Pick the auto-reject floor. Find the score where, on a human skim, you would never invite that resume. For most fresher India campus roles with a healthy scorer, that sits near the Reject/Maybe border (around 55-60%), not at 40%. Place the floor at or just below that line so clear mismatches never touch the invite path.

Step 3: Pick the auto-fanout bar. Find the score where you would invite without a second look. That is usually the bottom of Go (75%) or the bottom of Strong Go (90%) if seats are scarce and false invites are expensive. Start conservative: prefer a slightly smaller auto-invite set over a flooded calendar.

Step 4: Decide what Maybe means on this JD. Default: Needs Review, human invite. Some volume roles auto-fanout from mid-Maybe upward after a week of clean spot-checks. Do not start there on day one.

Step 5: Run a shadow check. On the first 50-100 scored resumes, list who auto-fanout would invite and who you would invite by hand. Aim for high overlap. Where they disagree, adjust the bar, not the whole philosophy. Write the disagreements down: resume keyword inflation, missing projects the JD cared about, or a seniority band capping someone you still wanted to hear. Patterns beat gut feel.

Step 6: Re-check after the first voice wave. Resume Go who become post-call No Go are a rubric or JD problem. Resume Maybe who become post-call Strong Go are a reminder that the voice layer exists for a reason. Adjust the invite bar after one full invite cycle, not after a single anecdotal resume.

Step 7: Document the JD’s band contract. One line is enough: “Fanout at 75+, Needs Review 60-74, Reject below 60.”

Volume hiring teams already feel the pressure: more applications than any recruiter can skim, which is why SHRM’s coverage of AI in talent acquisition keeps returning to screening automation. Bands keep that automation a filter instead of a loudspeaker.

One more calibration habit that saves arguments later: keep a dated screenshot or export of the score histogram the day you set the floors. When a hiring manager asks why Go starts at 75% instead of 65%, you point at the distribution you calibrated against, not at a memory of a Slack thread. If the next sheet drop produces a different shape, recalibrate. Do not keep the old floors out of pride.

Common mis-sets: invite flood vs dead shortlist

Two mistakes account for almost every angry Slack message we get about resume auto-decide.

Floor too low (invite flood). Setting auto-fanout at 40%, or auto-reject so low that Reject is empty, means the trigger path treats half the pile as ready for voice. Overnight you get an invite queue that looks successful and a slot calendar that is not. Voice ends up doing resume work the gate already failed. If that is already happening, read too many voice invites after resume screen for recovery. Raise the fanout bar into Go, put Maybe back under human review, and stop the drive only if wrong invites are still joining.

Ceiling too high (dead shortlist). Setting auto-fanout at 95% when almost nobody scores above 90% means autonomous invites never fire. Needs Review fills with Go and Strong Go waiting on a click automation was supposed to remove. Lower the bar into the real Go band, or bulk-invite Strong Go + Go once via the Score/Decision filter, then leave auto-fanout on for the next sheet drop.

New thresholds apply forward only. Do not expect yesterday’s 68% Maybe to rewrite itself when you raise the bar this morning. Hard cuts on invites already sent are revoke territory, not a band tweak.

What “good” looks like after one week

On a healthy fresher JD with relevance-aware scores and calibrated bands, you should see:

• A visible Reject share that never received invites
• Auto-fanout covering Strong Go and most of Go without daily babysitting
• A Needs Review Maybe pile small enough that one recruiter can clear it in a sitting
• Voice-slot usage that tracks intended shortlist size, not raw applicant volume

If Reject is empty, the floor is wrong or the scorer is compressed. If Needs Review is empty and invites are huge, the fanout bar is too low. If Needs Review is huge and invites are tiny, the bar is too high or auto-decide is off.

After the first week, also compare voice outcomes by resume band. If Strong Go paper matches are landing No Go after the call more often than Maybe, the JD or rubric is over-crediting keywords. If Maybe is producing Strong Go after the call, your invite bar may be too shy for that role. Use one full invite cycle of evidence before you move the floors again. Write the comparison into the role notes so the next recruiter inherits the lesson.

Resume bands are not a personality setting. They are the contract between paper signal and voice capacity. Set them with a real distribution in front of you, keep Maybe human until the data says otherwise, and never confuse a pre-call Strong Go with a post-call one. The invite should feel boring: the right resumes get a link, the wrong ones do not, and the middle waits for a person who still has time to think.

Want to walk the Score-Decision filter and per-JD thresholds on one of your live roles? Book a demo and we will set the first floor and fanout bar together, not from a slide.