# The Applicant-Flood Triage Playbook: Raw Export to a Corroborated Screen List

*Turn one role's raw applicant export into a ranked list of 10 to 25 names, each showing which claim was confirmed, by what source, and which fraud flags were cleared.*

- Canonical URL: https://www.refolk.ai/guides/inbound-applicant-triage-playbook
- Pillar: Recruiting and sourcing
- Format: Playbook
- Published: 2026-11-05
- Last reviewed: 2026-11-05
- Reading time: 20 min
- Keywords: how to screen hundreds of job applications for one role, how to spot fake job applicants, AI generated resumes how to screen, verify applicant work history before phone screen, high volume applicant triage process

## Key takeaways

- At 30 to 60 seconds per resume, reading 500 applications takes 4.2 to 8.3 hours, so dedupe, cluster detection and knockouts have to shrink the pile before anyone reads it.
- Technical roles in the Americas draw 610 applications per hire against 229 for business roles (Ashby), so set stage budgets by role type, not by one average.
- The strongest official fraud signal is structural: the same phone number, email or resume text under different names, which you can check without judging anyone's prose.
- Never use an AI-writing detector as a knockout: seven detectors flagged 61.22% of human-written TOEFL essays as AI-generated, which turns the detector into a proxy for national origin.
- Record the raw counts behind every four-fifths ratio, because federal guidance treats a selection rate under 80% of the highest group's as evidence of adverse impact but warns that small samples may not be significant.
- In Refolk's index, only 84% of top US Technical Recruiter results and 80% of UK results held that exact title, so a title mismatch is a reason to check function and employer, not proof a claim is false.

This playbook is for the in-house recruiter, sourcer or founder who posted one role, got 300 to 600 applications in a week, and needs a phone-screen list by Friday. It sets out a timed procedure in order. You collapse duplicates and coordinated submission clusters, apply two or three documented knockouts with a recorded adverse-impact check, and corroborate each survivor's key claims against public evidence. What you hand the hiring manager is 10 to 25 names, and every row shows what was confirmed, by what source, and which flags were cleared.

It covers one requisition at a time. It does not replace identity verification, which belongs later in the process.

## How many applications should one role expect?

A few hundred per hire is now normal, and remote technical roles draw far more. Ashby reports the average recruiter processing 291 applications per hire, against roughly 100 a few years earlier, and the average has since held above 300.

That average hides the spread you need for planning. Role type moves the number more than anything else:

| Region | Technical hire | Business hire | Technical vs business |
|---|---|---|---|
| Americas | 610 | 229 | 2.66x |
| EMEA | 254 | 170 | 1.49x |

*Inbound applications per hire. Ashby figures as reported second-hand; ratios derived.*

Remote work adds another multiplier. In Ashby's startup data, remote jobs at companies with 100 to 300 employees drew 426 applications per hire against 281 for in-office jobs, about 1.52x. A remote technical role in the Americas is the heaviest case this playbook covers.

Published benchmarks also disagree with each other. Greenhouse gives 244 applications per job in one benchmark and 95 per job in another. Treat any published average as a rough starting point and use your own ATS history for the same role family as the real baseline.

One number is missing everywhere: the share of applicants who meet most of a posting's stated criteria is not established publicly. Count it yourself for each requisition. The knockout log later in this guide records it. After three or four roles you will have a local figure that beats any published one.

Two more numbers set the stakes:

- **The funnel has tightened.** Ashby finds candidates are about 50% less likely to reach the interview stage than five years ago, and engineering hires average 17.9 interviews each. So each qualified person you wrongly reject is a bigger share of the viable pool.
- **AI use is common and fraud is a minority.** In a Gartner survey of 3,000 candidates, 39% said they used AI during the application process and 6% admitted to taking part in interview fraud. Gartner's "one in four profiles fake" figure is a forecast, not a measured rate.

## How long does triage take for a 500-application pile?

Reading every resume properly takes a working day. At 30 to 60 seconds each, 500 applications take 4.2 to 8.3 hours of reading alone. The time budget only holds if steps that need no reading shrink the pile first.

| Method | Seconds per resume | Hours for 500 |
|---|---|---|
| Ladders skim | 7.4 | 1.0 |
| ResumeGo low | 30 | 4.2 |
| ResumeGo high | 60 | 8.3 |

*Hours derived from the per-resume figures.*

The 7.4-second figure comes from Ladders' eye-tracking work. It measured an initial skim, not a full review, and nobody can judge fit or spot a made-up work history in 7.4 seconds. The ResumeGo range is closer to an honest read: 47% of respondents in that survey said they spend 30 seconds to a minute.

So the order of operations is also the time plan. Export, dedupe, cluster detection and knockouts all work on fields like email, phone, location and work authorisation, not on prose. They run fast and shrink the pile before the per-resume clock starts. At the ResumeGo rate, every 100 resumes that reach the reading stage cost 50 to 100 minutes.

#### Shrink first, read second

1. **Freeze** - Export the pile once, record the count
2. **Collapse** - Merge true duplicates, hold contact-sharing clusters
3. **Knock out** - Two or three yes/no requirements, with an adverse-impact check
4. **Read** - 30 to 60 seconds per survivor, rank a top 40 to 60
5. **Corroborate** - Check one or two load-bearing claims against public evidence
6. **Hand off** - 10 to 25 auditable rows to the hiring manager

*Every stage before reading works on structured fields, so the expensive per-resume time is spent only on survivors.*

Nobody has published time per application under structured triage, or how many applicants survive each middle stage. The stage budgets below are my estimates, pieced together from the sources in this guide. They are not documented benchmarks. On your first requisition, log your actual time and count at each stage and use them to set the budget for the second.

## The triage procedure, step by step

The procedure has eight stages, run in order, and each ends in a recorded "done" state so the hiring manager can audit it later. Do not reorder them. Reading before clustering wastes the reading budget, and corroborating before knockouts wastes the research budget.

#### One requisition, raw export to screen list

1. **Export and freeze the pile** - Export applications with contact fields, source and timestamps into one file nobody edits, and record the count. About 15 minutes (estimate).
2. **Exact dedupe, review before merge** - Match on email, phone and LinkedIn URL, then check name, work history and location for each pair before merging. 30 to 60 minutes (estimate).
3. **Detect submission clusters** - Group records where different names share a phone, an email or near-identical resume text, and label them hold with the shared field recorded. 1 to 2 hours (estimate).
4. **Apply two or three knockouts** - Run job-related yes/no requirements such as work authorisation, location or a must-have credential, and record pass and fail counts. Check LL144 scope first if a tool scores or ranks for an NYC role.
5. **Run the adverse-impact check** - Compute each group's selection rate per knockout from voluntary self-ID data, divide by the highest rate, and flag anything under 0.80 with the raw counts. About 30 minutes (estimate).
6. **Skim and rank the survivors** - Read survivors only, at 30 to 60 seconds each, and rank a top 40 to 60 against the must-haves.
7. **Corroborate load-bearing claims** - Check one or two claims per candidate against public evidence and mark each confirmed, unconfirmed or contradicted, with the source named.
8. **Hand off the screen list** - Give the hiring manager 10 to 25 ranked names, each showing the claim, the source and the flags cleared or still on hold.

### Stages 1 to 3: shrink the pile without reading it

Freeze first, because ATS records keep changing as new applications arrive and colleagues act on them. A frozen export with a recorded count lets you show later that nobody was dropped without a trace. Anything that arrives after the freeze goes into a second batch, not into the frozen file.

For exact dedupe, Greenhouse's Duplicate tag matches on email address, phone number and LinkedIn profile URL by default. Site Admins can add name for more sensitive matching and choose which criteria the tag uses. Those three fields are a sensible default in any ATS.

> **Rule:** Review, then merge
>
> In Greenhouse you cannot un-merge candidates, and other systems may be just as unforgiving. Never bulk-merge: check name, work history and location for each pair, and write the reason next to it before the merge runs.

Cluster detection is a different job from dedupe. A **duplicate** is one person applying twice. A **submission cluster** is several different names sharing one phone number, email address or near-identical resume text. FBI guidance tells employers to cross-check HR systems for applicants with the same resume content or contact information. It also notes that phone numbers were reused across resumes for supposedly different applicants, especially voice-over-IP (VoIP) numbers, which run over the internet rather than a carrier line. Label each cluster "hold" and record the shared field. No validated threshold for near-duplicate text has been published, so treat text similarity as a reason to look, not a verdict.

### Stages 6 to 8: read, corroborate, hand off

Read only what survives the knockouts, and rank against the must-haves rather than polish. AI-assisted phrasing is not a ranking factor. The question at this stage is whether the claimed experience matches the role, not whether the sentences sound human. Then corroborate the top 40 to 60 and hand off the best 10 to 25. Both stages have their own sections below.

## Which fraud signals justify a hold, and which justify a reject?

Hold on structural signals. Reject only when a documented knockout fails or public evidence contradicts a claim. Never act on a textual signal by itself. The official advisories list warning signs but, in the excerpts available, do not say how to combine them, so the combination rules here are my own judgement as a practitioner.

Each signal proves something narrow, and each has an innocent version:

| Signal | What it proves | When it lies |
|---|---|---|
| Same phone or email under different names | Shared contact point | Agency, bootcamp or career centre number |
| Same resume text under different names | Shared source text | Common bootcamp template |
| Inconsistent name spelling, location or contact | Records don't line up | Transliteration, recent move |
| Typos and unusual nomenclature | Very little alone | Second-language writer |
| Repository on the applicant's profile | Presence, not authorship | Copied repositories |
| Simple portfolio website | Nothing alone | Plenty of genuine freelancers |

The sources behind these rows are specific. The FBI notes reuse of phone numbers and email addresses across resumes for different applicants, and advises checking resumes for "typos and unusual nomenclature." The 2022 State, Treasury and FBI advisory lists inconsistencies in name spelling, nationality, claimed work location and contact information. It also describes fake personas building simple "portfolio" websites to boost their credibility. A later FBI alert adds multiple accounts created with the same identification document or accessed from the same IP address. Those last two sit inside platforms, not in your export, so you will rarely see them at triage.

#### Hold or screen

Horizontal axis runs from No structural signal to Contact or text shared across names. Vertical axis runs from Claim unconfirmed to Claim confirmed.

| Quadrant | What it means |
| --- | --- |
| Screen | Add to the list and rank on merit |
| Hold, check the intermediary | If the shared field belongs to an agency or bootcamp, clear it and screen |
| Screen, marked unconfirmed | A thin footprint is not a flag; raise the claim on the call |
| Hold for review | Keep off the list until cleared; if a claim is contradicted here, reject and record why |

*A structural signal puts a record on hold, a confirmed claim clears it, and only a contradicted claim inside a cluster justifies a reject.*

> The strongest fraud signal is structural, not textual: check who shares a phone number before anyone judges a sentence.

Three practical rules follow from the matrix:

- **A contradicted claim on its own is a hold, not a reject.** Check the contradiction first, because title conventions and employer renames produce false contradictions (more on this below).
- **Do not use "Reject as spam" on held records.** In Greenhouse it automatically rejects future applications from that email address. If the shared field turns out to be a career centre, you will have blocked real people without knowing it. Keep it for confirmed spam.
- **Report what looks like state-linked fraud.** If a cluster with a contradicted claim matches the DPRK IT worker pattern in the advisories, the FBI's IC3 portal is the reporting route.

Clearing these flags does not verify anyone's identity. The FBI has seen these workers use AI and face-swapping technology in video interviews, and it places identity verification "during interviewing, onboarding, and throughout the employment of any remote worker." Greenhouse has added an identity verification integration for this. It belongs after the screen list, not inside it.

## Knockouts, the four-fifths check, and NYC Local Law 144

Use two or three job-related knockouts. For each one, count passes and fails by demographic group and flag any group whose selection rate falls below four-fifths of the highest group's rate. A **knockout** is a yes/no requirement that ends an application by itself, such as work authorisation, location, or a must-have credential.

Pick knockouts you can check from a field, and that you could defend as job-related if anyone asked. Three is the ceiling. Every extra knockout is another chance to screen out a protected group without meaning to, and another ratio to record.

The federal standard is 29 CFR 1607.4(D): "a selection rate for any race, sex, or ethnic group which is less than four-fifths (4/5) (or eighty percent) of the rate for the group with the highest rate will generally be regarded by the Federal enforcement agencies as evidence of adverse impact." The arithmetic for each knockout:

1. **Selection rate** for each group = applicants who passed ÷ applicants in that group.
2. **Impact ratio** = that group's selection rate ÷ the highest group's selection rate.
3. Flag any impact ratio under 0.80.

The same regulation warns that larger differences "may not constitute adverse impact where the differences are based on small numbers and are not statistically significant." On a single requisition, some groups will be small. So the ratio by itself is not the record. The counts behind it are. Use voluntary self-ID data only. If you did not collect it, write "self-ID not collected" in the log rather than guessing demographics from names.

> **Note:** When a tool runs the knockout
>
> NYC Local Law 144 covers tools that issue a score, classification or recommendation used to substantially assist or replace discretionary decisions: the only factor, the most heavily weighted factor, or one that overrules a human. Covered tools need an independent bias audit within the prior year, a public summary of the results, and written notice to candidates at least 10 business days before use. The business using or developing the tool cannot run the audit itself. Check scope with counsel before switching on any auto-ranking for an NYC role.

**Knockout and adverse-impact log (one per requisition)**

```
Requisition:
Frozen export: file name, timestamp, total count
Knockout:
Job-related reason:
Applied by: person / tool (if tool and NYC role: LL144 scope checked yes / no)
Group | Applied | Passed | Selection rate | Impact ratio vs highest
Group A |  |  |  |
Group B |  |  |  |
Group C |  |  |  |
Highest-rate group:
Any ratio under 0.80: yes / no
Small-sample note:
Decision on this knockout (keep / revise / drop) and who decided:
```

*Repeat the block for each knockout; leave group rows blank and write "self-ID not collected" rather than estimating.*

## How do you verify an applicant's work history before the phone screen?

Pick one or two load-bearing claims per candidate and check each against a public source. Record each as confirmed, unconfirmed or contradicted, and name the source. A **load-bearing claim** is the one the hire actually rests on: the employer and role behind the must-have skill, the system they say they shipped, the licence they say they hold.

Not all evidence is equal. Rank it before you trust it:

#### How much a piece of evidence proves

1. **Authored work with dates** - Commit authorship and timestamps, papers or patents naming the person
2. **Third-party listing** - Employer team page, licensing board entry, conference programme
3. **Self-published presence** - Professional profile, repository on their profile, portfolio site
4. **The resume claim** - The starting point, which proves nothing by itself

*Evidence the applicant controls proves the least; dated, authored work in a third-party record proves the most.*

The middle layers are where fraud hides. The FBI documents North Korean IT workers copying company code repositories, such as GitHub repositories, to their own user profiles. A repository showing on someone's profile does not prove they wrote it. Check who authored the commits and when, and whether the dates fit the claimed job.

Nobody has published how often each public source covers a given person. I found no coverage rates for team pages, commit history, patents, publications or licensing boards, so do not treat a gap as a signal. A thin footprint marks a claim unconfirmed. It does not flag the person.

### Title drift: why a mismatch is not a contradiction

Titles vary by country and employer, and an exact-title search misses real people. Refolk's own index shows how big the effect is:

**21,271 - People currently titled Technical Recruiter in the United States, in Refolk's index**

The UK count for the same title is 709, about 30x lower, and probably reflects UK title conventions such as "Talent Acquisition Partner" more than a real shortage of the role.

| Country | Technical Recruiter count | Exact title in top-25 sample |
|---|---|---|
| United States | 21,271 | 21/25 (84%) |
| United Kingdom | 709 | 20/25 (80%) |
| US vs UK | 30.0x | |

*Source: Refolk's index of professional profiles. The UK figure may be depressed by title conventions; samples of 25 are illustrative.*

Even among top results for one title, 16% to 20% held a different exact title. So when an applicant's resume title does not match the employer's public records, compare the function, the employer and the dates before you write "contradicted."

Corroboration is where the minutes go. You are searching one name and one claimed employer across GitHub, LinkedIn and the open web, then repeating that 40 to 60 times. [Refolk](/) takes a plain-English query and returns the matching public profiles from all three in one pass. It also helps to calibrate first by running the role brief as a search. That shows what a real, checkable footprint for this role looks like before you judge an applicant's.

I ran this search: `Senior backend engineers in Austin with Go and Kubernetes experience and public GitHub commits in the last 12 months` - [see the full result list](https://www.refolk.ai/s/652n6zmmc8).

*Returns real engineers whose public profiles and recent commit history show what genuine evidence for this kind of role looks like, which is the yardstick for judging applicants' claims.*

## How this goes wrong: failure modes and false positives

Most triage failures come from treating a proxy as proof. Each one below either wrongly rejects real people or wrongly clears fake ones, and some do both.

1. **Merging two real people.** Shared household phones and family email addresses make two applicants look like one, and in Greenhouse the merge cannot be undone. *Check:* name, work history and location before every merge.
2. **Flagging an intermediary as a fraud ring.** Career-services centres, bootcamps and staffing agencies often submit many candidates under their own phone number. *Check:* whether the shared field belongs to a known intermediary before the hold becomes anything more.
3. **Using AI-writing detection as a filter.** Detectors flag careful second-language writing, because it reads as predictable to them (low "perplexity"). In the study below, essays every detector flagged had significantly lower perplexity. That makes the detector a stand-in for national origin, which is an adverse-impact risk. *Check:* nothing. Do not run it as a gate.
4. **Treating "typos and unusual nomenclature" as a fraud signal.** This FBI indicator has the same trap for non-native writers. *Check:* use it only alongside a contact-reuse match, never on its own.
5. **Trusting an unstable ratio.** The four-fifths ratio swings wildly when groups are small. *Check:* record the counts behind every ratio, not just the ratio.
6. **Calling a vendor score "decision support."** If recruiters in practice follow the score, the tool likely meets LL144's "substantially assist" test, whatever the contract calls it. *Check:* how the score is actually used, not how it is labelled.
7. **Treating "no public profile" as unverifiable.** Career returners, people under NDA and people in regulated jobs often leave thin public footprints. *Check:* mark them "unconfirmed," never "flagged," and ask on the call.
8. **Taking commit history at face value.** Copied repositories show up on a fraudster's profile too. *Check:* commit authorship and dates, not just the repository's presence.

**61.22% - Average share of human-written TOEFL essays that seven AI detectors flagged as AI-generated**

Liang et al. in Patterns also found that at least one detector flagged 89 of the 91 essays (97.80%).

> **Watch out:** The fair process and the accurate process are the same process here
>
> Every textual shortcut in this list fails on the people a fair process most needs to protect: second-language writers, returners and people with quiet public footprints. Rejecting on those shortcuts is unfair and wrong at the same time, so the record should show you did not.

## Handing off the screen list and keeping the process current

Give the hiring manager 10 to 25 ranked rows in the same fixed format. Each row shows the load-bearing claims, their status and source, the duplicate and cluster outcome, and what is still open for the phone screen. The fixed wording makes the list auditable. A manager can disagree with a ranking but can always see why each person is there.

**Screen list row for the hiring manager**

```
Rank:
Name:
Applied: date, source
Load-bearing claim 1:
  Status: confirmed / unconfirmed / contradicted
  Source: (e.g. commit authorship and dates on the public GitHub graph; employer team page)
Load-bearing claim 2:
  Status:
  Source:
Duplicate check: none found / merged with record (reason)
Cluster check: none / shared field cleared as known intermediary (name it) / on hold
Knockouts passed:
Open questions for the screen:
Identity verification: not yet performed; scheduled for interview stage
```

*One block per candidate in rank order; keep the status words exactly as written so rows can be compared.*

If the list comes up short of 10, do not relax the knockouts to fill it. Go back to the ranked 40 to 60 and corroborate further down. If that still falls short, source outbound for the same brief. The calibration query you ran in Refolk for corroboration doubles as that outbound list.

To keep the method current, re-check the things that change underneath it, not the numbers in this guide:

- **Benchmarks.** Ashby's Talent Trends Report updates its applications-per-hire figures by role type. Compare them against your own logged counts each quarter.
- **ATS duplicate rules.** Greenhouse lets Site Admins change which criteria tag duplicates. Confirm the live configuration before each high-volume requisition.
- **Fraud advisories.** The FBI's DPRK IT worker alerts have added new indicators over time. Read the latest one and add any field-level signal you can check to the cluster stage.
- **Automated decision rules.** Check the DCWP's guidance on Local Law 144 before any new scoring feature goes live for an NYC role.

#### Before the screen list goes to the hiring manager

- [ ] The frozen export file and its count are recorded, and late applications sit in a separate batch.
- [ ] Every merge has a written reason and was reviewed against name, work history and location first.
- [ ] Every cluster hold names the shared field and whether it belongs to a known intermediary.
- [ ] No candidate was rejected on an AI-writing score, on typos, or for having no public profile.
- [ ] Every knockout has pass and fail counts per group, an impact ratio, and the raw numbers behind it.
- [ ] Any automated scoring or ranking used for an NYC role has been checked against Local Law 144 scope.
- [ ] Every screen list row shows each claim's status and a named source.
- [ ] Nobody in an unresolved cluster has been marked "Reject as spam."
- [ ] The hiring manager knows identity verification has not yet been performed.

## Frequently asked questions

### How do you spot fake job applicants without rejecting real ones?

Start with structural signals rather than writing style. FBI guidance points to the same resume content or contact details turning up under different applicant names, including reused VoIP phone numbers. Put those records on hold, check whether the shared field belongs to an agency, bootcamp or career centre, and corroborate one or two load-bearing claims against public evidence. Reject only when a claim is contradicted, never on typos, an AI-writing score or a thin public profile alone.

### Should I use an AI detector to screen AI-generated resumes?

No. In a study published in Patterns, seven detectors flagged an average of 61.22% of human-written TOEFL essays as AI-generated, and at least one detector flagged 89 of the 91. Careful second-language writing looks machine-like to these tools, so a detector score works as a proxy for national origin. Gartner found 39% of candidates use AI while applying, most of them legitimately. Judge claims, not prose.

### How many applications per hire is normal for one role?

Ashby reports recruiters processing 291 applications per hire on average, with the average holding above 300. Role type moves the number most. In the Americas, technical roles draw 610 per hire and business roles 229, and remote startup roles draw more than in-office ones. Published benchmarks disagree with each other, so use your own ATS history for the same role family as your baseline.

### What if an applicant has no public profile to verify against?

Mark the claim unconfirmed, not flagged, and keep them on the list if they are strong on paper. Career returners, people under NDA and people in regulated jobs often leave little public trace, and nobody has published coverage rates for team pages, commit history or licensing boards. Write the unconfirmed claim into the screen list as a question for the phone call. That is the cheapest place to resolve it.

### Does NYC Local Law 144 apply to my ATS auto-ranking feature?

It might. The law covers tools that produce a simplified output such as a score, classification or recommendation that substantially assists or replaces discretionary decisions. That includes cases where the tool is the only factor, the most heavily weighted factor, or overrules a human. If recruiters routinely follow the score, the tool likely meets that test. Covered tools need an independent bias audit within the prior year, a public summary and 10 business days' notice to candidates.

### Should I merge duplicate applications in my ATS?

Only after reviewing each pair. Greenhouse's Duplicate tag matches on email, phone and LinkedIn URL by default, and merges cannot be undone. A shared household phone or a family email address can tag two real people as one record. Before merging, check name, work history and location, and write down why you merged. If the pair might be two different people, leave them separate and note the shared field.

---

*From the Refolk guide library. I revise these guides rather than replacing them, so the current version is always at https://www.refolk.ai/guides/inbound-applicant-triage-playbook*
