You have read it so often that it feels like a fact: three-quarters of résumés are rejected by software before a human ever sees them. It sits on the landing page of every résumé service and in most job-search threads. It is the reason people pay for a keyword score, paste the job description into the footer in white text, and rebuild a clean one-page résumé into something a parser is rumored to prefer.

The figure has no study behind it. What screening software actually does is both less sinister and, in one specific respect, more consequential than the folklore says. Knowing the difference changes where a job seeker’s effort should go.

Where the 75% came from

The earliest traceable appearance is a trade-press article from March 2012, which said applicant tracking systems “kill 75% of candidates’ chances of landing an interview as soon as they submit their resumes, according to job search services provider Preptel” [1]. The article gave no sample, no method and no study. The claim was also narrower than the one that now circulates: it was about the odds of an interview, not about résumés being discarded unread.

Preptel was a company that sold a résumé-optimization service to job seekers for $24.95 a month [2]. The number that has shaped more than a decade of job-search advice appears to have begun as a figure supplied by a business with something to sell on the strength of it. Nobody has since produced evidence for it.

What the software does

Mostly it files and sorts. In a Harvard Business School survey of 2,275 executives in the United States, the United Kingdom and Germany, 63% of employers said they use a recruitment management system, and more than 90% of those use it to make a first cut or to rank applicants [3]. Ranking is the word to notice. An application is rarely deleted. It is placed somewhere in a list, and a recruiter with limited time works from the top.

Automatic rejection does exist, and vendors document it plainly. In Greenhouse, a widely used applicant tracking system, auto-reject is triggered by a candidate’s answers to application questions the employer has written, such as whether you hold a required license or can work in a given location. It does not act on the wording or layout of the résumé [4]. The same vendor states that its AI matching feature sorts candidates into categories for a human to review and does not reject anyone by itself [5].

The part of the fear that is justified

None of this means the software is harmless. The same Harvard research found that the majority of employers acknowledged their hiring systems rejected potentially productive applicants. The mechanism is explicit rules written by people: excluding candidates without a degree, for instance, or ranking down anyone who has not been employed in the last six months [3].

Employers are not confident in what their own tools are doing. In a 2025 Greenhouse survey, only 21% of recruiters said they were very confident their systems are not rejecting qualified candidates, and 8% said they have no idea what their algorithms prioritize [6]. How often a machine makes the final call, nobody knows: no official statistic on automated screening exists.

The clearest documented case of harm involved an explicit rule. In the first settlement the US Equal Employment Opportunity Commission reached over hiring software, a tutoring company paid $365,000 after the agency alleged it had programmed its application system to reject women aged 55 or older and men aged 60 or older automatically. More than 200 qualified applicants were turned away [7]. Oversight remains thin. A state audit of New York City’s law on automated hiring tools found that the enforcing agency received two complaints in two years, and that when it reviewed 32 companies it found one compliance problem where the auditors found at least 17 [8]. The European Union’s AI Act classes recruitment and selection systems as high-risk [9].

So the accurate picture is this. You are unlikely to be rejected outright because a parser disliked your layout or missed a keyword. You can be excluded by an explicit criterion you did not meet, sometimes a poorly chosen one. And you can be ranked far enough down a long list that no one reaches you, which for the applicant amounts to the same thing.

Meanwhile, AI changed what the document is worth

A second shift matters more than the software. Writing help works at the screen. In a field experiment involving nearly half a million job seekers, those randomly given algorithmic writing assistance on their résumés were hired 8% more often, with no evidence that employers were less satisfied [10]. In a controlled experiment, language models used as résumé evaluators preferred résumés they had generated themselves over human-written ones of the same quality between 67% and 82% of the time [11]. A clear, well-edited document is an advantage, and using an assistant to produce one is reasonable.

What has been lost is the signal that tailoring used to carry. After an AI cover-letter tool was introduced on a large online labor platform, the link between how closely a letter matched the job and whether the applicant got a callback fell by 51%, and employers shifted their attention toward other signals, including prior work history [12]. On another platform, employers were willing to pay more for workers whose applications were customized to the posting before large language models arrived, and not afterward [13]. When everyone can produce a tailored letter in seconds, a tailored letter says much less about the person who sent it.

Polish has limits of its own. In two field experiments on cover letters, language models improved the quality of the letters, most of all for weaker applicants, but the gains did not raise interview invitations. And when recruiters learned that a model had been used, they placed more value on strong letters people had written themselves [14].

Their response has been to check in person. In the Greenhouse survey, 39% of US hiring managers said they were holding more in-person interviews to verify that candidates are who they appear to be, and 32% said they had caught applicants reading from AI-generated scripts [6]. The document gets you ranked. The conversation is where the claims in it are tested.

What to do with this

The practical advice that follows from the evidence is duller than the folklore, and more useful.

  • Treat the application questions as the real filter. They are the one place automatic rejection is documented. Read each one carefully and answer accurately. If you do not meet a criterion marked as required, assume the application may never be read and look for another way in.
  • Keep the document plain. Standard headings, a simple layout, ordinary file formats. Nothing in the evidence rewards clever formatting, and hidden text or pasted keywords are a risk with no demonstrated benefit.
  • Use an assistant for clarity, and supply the substance yourself. Polish helps. A model cannot know what you delivered, so the outcomes, numbers and dates have to come from you.
  • Describe your work in the employer’s terms where they are true. If the posting asks for experience you have, say so in the words it uses. This helps a ranking system and a hurried reader alike.
  • Address a gap directly. If you have been out of work for some months, say what you did. Rules about recent employment are among the filters employers admit to using.
  • Be ready to stand behind every line. Work history is what employers now weigh, and the interview is where they check it. For each claim on the page, know the situation, your part in it, and the result. Each interviewer will probe a different part of it.
  • Do not make the application your only route. A ranked list is the weakest position from which to be noticed. Referred and internal candidates are hired at many times the rate of cold applicants.

Persona Map is where the substance behind the document is kept: a private workspace for the outcomes behind each line of your work history, so that what you say in the room matches what you wrote. The software was never the adversary the folklore describes. It is a filing system with a few blunt rules, sitting in front of a person with too many applications and too little time. Meet the rules, be clear, and spend the effort you save on the parts of the process a person decides.

Sources

  1. Meridith Levinson. “5 Insider Secrets for Beating Applicant Tracking Systems.” CIO.com, republished by Computerworld, March 2012. Link
  2. CIO.com. Review of Preptel’s résumé-optimization service, February 29, 2012. Link
  3. Joseph Fuller. “How to Tap the Talent Automated HR Platforms Miss.” Harvard Business School Working Knowledge, 2021. Link
  4. Greenhouse Support. “Application Rules Overview.” Link
  5. Greenhouse Support. Talent Matching: frequently asked questions. Link
  6. Greenhouse. 2025 AI in Hiring survey, press release. Link
  7. US Equal Employment Opportunity Commission. “iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit.” September 11, 2023. Link
  8. Office of the New York State Comptroller. Audit: Enforcement of Local Law 144, Automated Employment Decision Tools. December 2, 2025. Link
  9. European Commission, AI Act Service Desk. Annex III: high-risk AI systems. Link
  10. Emma Wiles, Zanele Munyikwa & John Horton. “Algorithmic Writing Assistance on Jobseekers’ Resumes Increases Hires.” NBER Working Paper 30886, 2023. Link
  11. Xu, Li & Jiang. Experiment on language-model evaluators and model-written résumés. Working paper, 2025. Link
  12. Cui, Dias & Ye. Study of an AI cover-letter tool on an online labor platform. Working paper, 2025. Link
  13. Galdin & Silbert. Study of application customization on Freelancer.com before and after large language models. Working paper, 2025. Link
  14. Kian Abbas Nejad, Giuseppe Musillo, Till Wicker & Niccolò Zaccaria. “Labor Market Signals: The Role of Large Language Models.” Accepted by the Journal of Labor Economics, 2026. Link