The 75% Auto-Rejection Myth: What Is Actually Reading Your CV, and Why Almost Nobody Will Tell You

Dear MoreThanCareer.de community,

Welcome back to the newsletter your employer's HR software vendor would rather you did not read in the fortnight before the licence renewal meeting.

Save this one. Not out of politeness, and I am not precious about it. Save it because somebody you know is going to open a laptop on a Thursday and find a rejection timestamped 07:12, written and sent before they were awake, for a role they gave up a Sunday to prepare for. They will spend the rest of that day auditing themselves for a mistake they made at seven in the morning. When they tell you about it, you will want this article rather than a sympathetic noise.

Housekeeping: everything I do lives at www.morethancareer.de. Career coaching, interview preparation, salary negotiation training with the actual scripts rather than the theory, reverse recruitment, CV and LinkedIn rewrites. That is the last mention until the end.

Right. Gloves on.

Part XV took apart what CV-screening software actually rejects you for. Part XVII established that German ghosting is mostly arithmetic in a lab coat. Since then I went looking for one specific thing, could not find it anywhere, and the not-finding is this article.

The short version, for anyone reading on a phone between meetings

The 75% figure is fiction and always was. In Germany, exactly 1% of employers use AI to pre-screen applications at all.¹ So far, so reassuring. Then a British regulator opened up real hiring tools and found some of them guessing candidates' ethnicity from their names, and others letting recruiters filter people out by protected characteristic.² New York became the only place on earth to legally require employers to publish what their hiring algorithms do, and 18 employers out of 391 published anything.³ I checked seventeen pharma and CRO careers portals and found precisely one that tells applicants what its system does. Brussels was due to make all of this a regulated high-risk activity on 2 August 2026 and, five weeks before the date, moved it to December 2027.⁴

The machine is smaller than you think. The hole where the accountability should be is enormous.

First, let us bury the number properly, because it is still being sold to you

Seventy-five per cent of CVs rejected by software before a human sees them. You have read it a hundred times, usually two clicks above a button that costs €89.

Follow it back and it dies in a ditch. The earliest version traces to Preptel, a CV-optimisation vendor, published around 2012 with no study, no survey and no method attached. Preptel went out of business in August 2013. The figure outlived the company by thirteen years and counting. A separate 2013 industry piece said 75 of 100 CVs get screened out "by either the ATS or a recruiter", and somewhere around 2014 the recruiter quietly fell out of that sentence and never came back. CIO.com, cited for years as the source, today publishes something else entirely: 75% of recruiters use an ATS.⁵ Same number, different noun, twelve years of panic.

The best part is who has now recanted. Jobscan built an entire business on beating the machine. Its own current guidance states that an ATS "rarely 'rejects' a candidate automatically".⁶ When the dealer stops using the product, you have learned something about the product.

So the number was a Nebelkerze (a smoke candle, the thing you throw to make people look the wrong way). Fine. Now stop celebrating, because the reason the myth survived is that it was pointing vaguely at something real, and the real thing is uglier.

What the British regulator found when it actually opened the machines

In November 2024 the UK Information Commissioner's Office did the thing nobody else has done: it audited live AI recruitment tools and published the results.²

Some tools were "inferring characteristics, including gender and ethnicity, from a candidate's name instead of asking for this information." Some were "allowing recruiters to filter out candidates with certain protected characteristics." Some "collected far more personal information than necessary and retained it indefinitely to build large databases of potential candidates without their knowledge."

Nearly 300 recommendations. All accepted or partially accepted.

Sit with that for a second. A regulator looked inside working hiring software and found products that would guess whether you are a woman, or where your grandparents were born, from the letters in your surname. Not as an emergent quirk of a neural network. As a feature, shipped, sold, and installed by employers who presumably attended a demo.

And when we do know exactly what a system did, we usually know because somebody sued. iTutorGroup had programmed its application software to automatically reject female applicants aged 55 and over and male applicants aged 60 and over. More than 200 qualified people. Settled with the EEOC for $365,000 in September 2023.⁷ No mysterious algorithm. A human being typed in two numbers, and the software did as it was told, faithfully, for as long as nobody looked.

Where the machine genuinely does end your candidacy unseen

Not by reading your CV and disliking your font. By reading the form.

Greenhouse ships an auto-reject function driven by your answers to yes/no questions, and its own help page confirms that people set up to receive applicant notifications "will not be notified of auto-rejected applicants."⁸ Nobody is told. There is no inbox this lands in. Oracle calls them disqualification questions, whose answers "decide if candidates move forward in the selection process or are automatically disqualified", with work authorisation as the worked example.⁹ Recruitee sells knockout questions that eliminate candidates "without any manual work", and then advises delaying the rejection email "to keep the candidate experience positive".¹⁰ A rejection held back a couple of days so it feels considered. Somewhere there is a product manager who thinks that is empathy.

Then there is the filter that should make every reader over 45 put down their coffee. Harvard Business School and Accenture surveyed 2,275 executives across the US, UK and Germany. Forty-eight per cent confirmed that if a work history has a gap of more than six months, the CV is "automatically screened out by their RMS or ATS, based on that consideration alone."¹¹ Eighty-eight per cent agreed that qualified high-skills candidates get filtered out for failing to match the exact criteria typed into the job description.

Nine months after a restructuring. Two years raising children. A year of caring for a parent. One tick box, configured once, by somebody who has since left the company.

Now the German part, where the giant shrinks

Bitkom asked 852 German companies with three or more employees what they actually run. One per cent use AI to pre-select applications. One per cent let AI conduct interviews. Three per cent run AI competence analysis, four per cent have a chatbot.¹ Meanwhile 100% accept digital applications and 63% do video interviews. The German employer has digitised every part of hiring except the part that requires an opinion. You are statistically likelier to meet a German company that still publishes a fax number than one that lets software build its shortlist.

Michael Ende gave us the right word for this. A Scheinriese is a seeming giant: Herr Tur Tur, who looks monstrous from a distance and shrinks to an ordinary, slightly apologetic man by the time you reach him. Every German child knows the story. Every German job seeker forgets it the moment they open a careers portal.

And German law has already drawn a line. The data protection authorities publish, as their textbook example of an unlawful practice, an AI that evaluates all incoming applications and independently issues the interview invitations. Breach of Article 22(1) GDPR.¹² The same guidance says the human in the loop must retain genuine decision-making latitude, and warns in writing that "insufficient staffing, time pressure and a lack of transparency about how the AI reached its result" must not lead to results being adopted unchecked.¹²

Read that as a candidate. The regulator has already written down, precisely, the thing you suspect is happening to you. A recruiter working down a ranked list at four on a Friday afternoon is not oversight. That is a Feigenblatt (a fig leaf) with a login.

The thing nobody has counted, so I counted it

Here is the experiment. Seventeen employers that matter to this audience, checked against the one piece of evidence nobody can spin: the address bar of the actual application page.

Roche, Novartis, Pfizer, GSK and Sanofi all apply through Workday, and so do IQVIA, Parexel, Fortrea and ICON. Bayer runs SAP SuccessFactors for applications plus a separate Eightfold AI portal that will match your uploaded CV against its database. Boehringer Ingelheim and BioNTech are on SuccessFactors. AstraZeneca's entire external careers site runs on Eightfold. Takeda uses Avature, Lonza uses SmartRecruiters, Merck KGaA and Thermo Fisher front-end with Phenom. All public, all verifiable in an afternoon.

Then I looked for the sentence. The one where the employer tells you, in plain words, what the software does to your application. Whether anything is filtered automatically. Whether a person sees everything, or only the top of a ranked list.

Seventeen employers. One. ICON, a CRO, states on its own careers blog that Workday extracts your details, that its parser prefers plain text, and then this: "a human recruiter or hiring manager always reviews shortlisted candidates."¹³ Eleven words, and the load-bearing one is "shortlisted", which quietly concedes the gate sits upstream of the human. Still eleven more honest words than the other sixteen managed between them.

Disclosure, since I am about to be rude about an industry I have spent seventeen years inside: two of the organisations on that list have employed me. I am not throwing stones from outside the glasshouse. I am standing in it, pointing at a window.

New York tried to fix exactly this. It is worth knowing how that went.

Since July 2023, New York City has required employers using automated employment decision tools to commission an independent bias audit, publish a summary, and notify candidates ten business days ahead.¹⁴ The first law of its kind anywhere on earth.

Cornell researchers, working with 155 student investigators, checked 391 employers. They found 18 published audit reports and 13 transparency notices.³ Under 5%, and under 4%. The authors are careful and so am I: employers have wide latitude to decide the law does not apply to them, so silence is not proof of a breach. They named the phenomenon "null compliance", which is the most diplomatic description of a regulatory vacuum I have ever read.

Then the New York State Comptroller audited the enforcement agency itself, covering July 2023 to June 2025. In two years, the Department of Consumer and Worker Protection received two complaints. Two. In a city of eight million people, about a law touching every large employer in it. The department reviewed 32 companies and found one instance of non-compliance. The state auditors then reviewed the same 32 companies and found at least seventeen.¹⁵

So the world's first algorithmic hiring transparency regime consists of a law almost nobody publishes under, a public almost nobody has informed, and a regulator that spotted one problem out of at least seventeen sitting directly in front of it.

And Europe? Europe kicked it down the road.

For two years every HR conference in the German-speaking world sold the same slide: 2 August 2026, the day recruitment AI becomes a high-risk system under the EU AI Act, with documentation, oversight and obligations attached.

Five weeks before that date, Regulation (EU) 2026/1744, the Digital Omnibus on AI, moved it to 2 December 2027.⁴ Sixteen extra months, auf die lange Bank geschoben (literally pushed onto the long bench, meaning kicked comfortably down the road), granted at the last possible moment to organisations that had in many cases already bought the compliance software. December 2027 is not quite Sankt-Nimmerleins-Tag (Saint Never's Day, the date that never arrives), but it shares a postcode.

The classification did not change. Annex III still lists systems used "to analyse and filter job applications" as high-risk.⁴ Only the clock moved.

Two things were not postponed, and both are yours to use. Emotion recognition in the workplace has been prohibited since 2 February 2025, so any video tool scoring your "enthusiasm" or "confidence" from your face or voice has been unlawful in the EU for eighteen months. And Article 22 GDPR never went anywhere.

A short tour, because your feed makes it sound like everyone else has solved this

The United States has legislated in four directions at once. Illinois required AI notice from 1 January 2026, then withdrew its implementing rules in June 2026 with no new timeline, so the duty exists and the compliant way to satisfy it does not.¹⁶ Colorado passed America's most ambitious AI employment law, delayed it, had enforcement blocked in federal court, then repealed and replaced it in May 2026 with a lighter notice regime starting January 2027.¹⁷ Texas made it unlawful to discriminate with AI on purpose, a standard that would acquit almost any system ever shipped.¹⁸ The EEOC withdrew its AI hiring guidance entirely in 2025.¹⁹ The statutes still apply. The instruction manual has been taken off the shelf.

Switzerland has no AI act and does not want one, yet Article 21 of the revised data protection act still gives you notice of a purely automated decision with significant effects and, on request, review by an actual human.²⁰ India's data protection act has no equivalent of Article 22 at all: no right against automated decisions, no right to an explanation, and the core obligations do not bite until roughly mid-2027.²¹ If you are applying into a Bangalore shared service centre, your European instincts do not travel with the application.

The part that should embarrass this industry specifically

German pharma is the most named-responsibility sector in Europe. We do not merely have processes. We have people, with certificates, whose own names go on the line.

A company placing finished medicines on the market must appoint a Stufenplanbeauftragter under § 63a of the German Medicines Act, a qualified person personally charged with running pharmacovigilance and coordinating the necessary measures when a safety signal appears.²² There is an information officer under § 74a. A responsible person under § 52a. A Qualified Person for batch release. A medical device consultant under § 83 MPDG. A PRRC under Article 15 of the Medical Device Regulation. A narcotics officer, a radiation protection officer, a dangerous goods officer.

We appoint a certified, personally liable human being to supervise a pallet of flammable solvent travelling eleven kilometres down the motorway.

For the software that decides which human beings get to work with the solvent, there is no officer, no signal detection, no adverse event reporting, no periodic safety update. Nobody counts a false rejection, because a false rejection generates no case number, no complaint and no patient. It generates a person at a kitchen table concluding that the fault was theirs.

If a production batch failed at the rate hiring processes fail, we would halt the line, notify the authority, and hold a root cause meeting so tense you could hear the legal department swallow. Because it is only careers, we call it candidate experience and hand it to marketing.

The honest verdict

Not a conspiracy. Not a clean process either.

The 75% is dead and deserves to stay dead. In Germany, software is almost never the thing that rejects you. What you are actually up against is a handful of gates configured years ago by someone who has since resigned, running inside a system nobody in the building can fully explain, upstream of a recruiter holding several hundred applications, in a legal regime that will not require anyone to document any of it until December 2027, inside an industry that would never dream of releasing a product this way.

Your rejection is rarely a decision. It is the absence of one, executed quickly. Which is worse than the robot, and also, for the first time in this series, something you can act on with precision.

Five countermoves, ranked by the evidence behind them rather than by how clever they sound

1. The form is the exam. The CV is the appendix.

The commodity advice obsesses over fonts and margins. The vendor documentation quoted above is blunt: this is the only place in the process where software genuinely ends your candidacy with nobody informed.

The specific action: screenshot every question before you submit anything. Work authorisation gets an explicit answer plus the permit type written into any free-text box, never an implication. If you hold an EU Blue Card, a § 18b or § 18g permit, a Chancenkarte (Germany's points-based opportunity card for jobseekers) or Swiss B or C status, write the words. "Yes, no sponsorship required" beats a blank. Years of experience is a numeric filter, not a modesty test, so count adjacent and academic work honestly. Never leave salary empty, because an empty field and an outlier are treated identically in several systems. And if a question would be unlawful in an interview, that is free information about the employer.

2. Route around the queue, because the arithmetic is not close.

The commodity advice says "network". Thank you. Enormously helpful.

Greenhouse's benchmark data across 640 million applications shows applications per role rising from 115 in 2022 to 244 in 2025, while applications handled per recruiter per year went from 146 to 746 and recruiting teams shrank by more than half.²³ In Germany, StepStone found application interest per posting almost doubled between early 2023 and early 2026, with scientific fields up 126%.²⁴ Against that: personal contacts and own staff were the decisive channel in 27% of successful German appointments, ahead of job boards at 21%,²⁵ and referred candidates clear the first screen at 52% against 35% overall.²⁶ On a specialist board such as BioSpace, a pharma posting drew an average of 4.4 applications in Q2 2026.²⁷ Four. Against 244.

The specific action: halve your mass applications this month and put the recovered hours into twelve named human beings. Hiring managers in your function, former colleagues who have since moved, and the two or three recruiters who genuinely own your specialism in DACH. One warm introduction beats forty cold submissions and always has. No algorithm is involved anywhere in that sentence, which is the point.

3. Make the gap boring before a filter makes it interesting.

The standard advice for older candidates is to delete the graduation year and hope. I have given a version of that advice myself and it is half a measure, because the filter that removes you is usually reading dates in the employment history, not the education section.

The specific action: you cannot argue with a tick box, so take the gap out of the field the tick box reads. Put a dated entry across the period. Interim work, freelance mandates, a named certification with its awarding body, a Weiterbildung (structured professional further training, the German version of continuing education), caring responsibilities stated plainly. Two lines, real dates, no apology, no explanation. Explanations belong in interviews, where a human can be persuaded. And a related number that surprises people: German applicants aged 25 to 34 use AI in their applications at 55%, against 28% of those aged 45 to 54.²⁸ Your younger competition is not sharper than you. They are using a tool you are leaving on the shelf. Write it yourself, then make the machine attack your own draft against the advert.

4. Fix the record the system actually scores, not the document you are proud of.

This is the one most people get wrong without ever knowing. In Workday, your candidate profile is the record. The lovingly formatted PDF is an attachment to it. If the parser mangled your job titles in 2021 and you never corrected the profile, the 2021 mess is what gets scored, not the CV you just uploaded. Eightfold, which runs Bayer's matching portal and the whole of AstraZeneca's careers site, does semantic matching against an uploaded CV, so guessing at keywords is wasted effort and uploading a properly parseable file is not.

The specific action: the next time you apply through a system you already have an account in, open the profile itself and read every parsed field before you submit. Fix the titles, dates and employers by hand. Ten minutes, once per employer, and it outlasts every individual application. While you are there, glance at the domain of the application page: myworkdayjobs.com, successfactors.eu, eightfold.ai, avature.net. Being honest, knowing the vendor is worth ten minutes rather than ten hours, and what it buys you is exactly this paragraph and nothing more mystical. Anyone selling you a system-by-system optimisation package is selling a map of a country they have flown over.

5. Ask for the disclosure, with realistic expectations.

I am going to be straight, because this newsletter does not sell miracles: this will not get you the job. Nobody has ever been hired by exercising a data protection right.

What it does do is produce information, and occasionally a correction. Article 22(3) GDPR gives you the right to human intervention, to state your case and to contest a decision. Article 15 lets you ask what data was processed and how.²⁹ Before applying, if you have any human contact at all, one flat question is worth a week of tailoring: "Is any part of the first screening automated, and does a person see every application?" Almost nobody asks. The answer, including a refusal to answer, tells you a great deal.

The specific action: after a rejection that could not plausibly have involved a human, send one short, unemotional paragraph asking whether the decision involved automated processing and requesting human review. Not a threat, not a complaint, and never a first move, because you want a job rather than a grievance. An employer who cannot answer that within a month has told you more about itself than any employer branding page ever will.

For the recruiters and TA leaders reading this over somebody's shoulder

Not a moral appeal. Arbitrage, as usual.

Appoint a named owner for your hiring stack, in the spirit of every other named role in this industry, and put that name somewhere candidates can see it. You will be the only employer in your peer group who has. Audit your knockout questions this quarter and delete every one that is not a genuine legal or licensing requirement, because that is where your strongest applicants are vanishing with nobody notified. Bin the six-month gap filter, which removes carers, the over-50s and everyone who survived a restructuring, which in pharma right now is most of the people worth hiring. And if you use scoring, name the reviewer and give them the time to actually review, because German guidance already requires genuine decision-making latitude rather than a signature.¹²

Doing this in 2026 costs a quarter of one person's attention. Doing it in December 2027, with documentation obligations attached and a competitor already advertising their audit, will cost considerably more.

One more, free of charge: 61% of hiring managers now run software to detect whether candidates used software. Two machines are arguing about your suitability and both invoices go to a human being. Da hat jemand den Bock zum Gärtner gemacht (someone has put the goat in charge of the garden).

You have reached the end of a long article about the paperwork behind your silence, which most people will not. So do the small thing rather than the big one: open the last system you applied through, log in, and read your own parsed profile. Fix what the parser got wrong. Ten minutes, and you will apply differently for the rest of the year.

Your turn

• Have you ever been rejected fast enough that no human could plausibly have read anything? What time did it land?
• Recruiters, honestly and with no company names: does your organisation know today which knockout questions are switched on across your open requisitions?
• Which employer, anywhere, has told you in plain language what their system does with your application? I found one out of seventeen. I would like to be proved wrong.
• For readers over 50: has removing dates ever actually helped, or did it only relocate the surprise to the interview?
• Would you rather be rejected in four minutes by software, or ignored for four months by a person?

Prost. Und lasst euch von keinem Scheinriesen den Abend verderben. (Cheers. And do not let a seeming giant ruin your evening.)

Sources and references

Primary, official and institutional:

[1] Bitkom, Die Bewerbung läuft fast überall schon digital, aber meistens noch ohne KI, 10 March 2025. Representative telephone survey, 852 German companies with 3+ employees: bitkom.org

[2] Information Commissioner's Office, AI tools in recruitment: audit outcomes report, 6 November 2024: ico.org.uk and the accompanying statement: ico.org.uk

[3] Wright, Muenster, Vecchione, Qu, Cai, Smith, Metcalf & Matias, Null Compliance: NYC Local Law 144 and the Challenges of Algorithm Accountability, FAccT '24, 3 June 2024: arxiv.org

[4] Regulation (EU) 2026/1744 of 8 July 2026 (Digital Omnibus on AI), in force 27 July 2026: eur-lex.europa.eu; EU AI Act Annex III, point 4: ai-act-service-desk.ec.europa.eu

[7] EEOC, iTutorGroup to Pay $365,000 to Settle EEOC Discriminatory Hiring Suit, 11 September 2023: eeoc.gov

[11] Fuller, Raman, Sage-Gavin & Hines, Hidden Workers: Untapped Talent, Harvard Business School / Accenture, September 2021. 2,275 executives and 8,720 workers across the US, UK and Germany: hbs.edu

[12] Datenschutzkonferenz, Orientierungshilfe Künstliche Intelligenz und Datenschutz, 6 May 2024: datenschutzkonferenz-online.de

[14] NYC Department of Consumer and Worker Protection, Automated Employment Decision Tools: nyc.gov

[15] Office of the New York State Comptroller, audit 2024-N-6, issued 2 December 2025: osc.ny.gov

[20] Art. 21 revDSG (Switzerland), in force 1 September 2023: datenschutzpartner.ch

[22] § 63a AMG, Stufenplanbeauftragter: dejure.org

[25] IAB-Forum, Online-Medien sind bei der Personalsuche auf dem Vormarsch, 8 June 2026, based on the IAB-Stellenerhebung: iab-forum.de

[29] Articles 22 and 15 GDPR: privacy-regulation.eu

Employer and vendor documentation, quoted because it is each company's own description of its own product:

[6] Jobscan, 8 Things You Need to Know About Applicant Tracking Systems: jobscan.co

[8] Greenhouse Support, Auto-reject: support.greenhouse.io

[9] Oracle, Disqualification Question: docs.oracle.com

[10] Recruitee, Set up knockout questions to automatically disqualify candidates: support.recruitee.com

[13] ICON plc, Applying to ICON through Workday, 5 January 2026: careers.iconplc.com

Tracing the myth:

[5] CIO.com, What is an applicant tracking system?: cio.com. The Preptel trace and the 2013 bankruptcy, via HR consultant Christine Assaf: blog.hiringthing.com and Jan Tegze's independent trace: jantegze.medium.com

Legal analysis and reporting:

[16] Seyfarth Shaw on the Illinois Department of Human Rights withdrawing its proposed AI rules, June 2026: seyfarth.com

[17] Seyfarth Shaw, Colorado Enacts Artificial Intelligence Replacement Law, May 2026: seyfarth.com

[18] K&L Gates on the Texas Responsible Artificial Intelligence Governance Act, signed 22 June 2025: klgates.com

[19] Cooley on the withdrawal of federal AI employment guidance, 2025: cooley.com

[21] India Briefing on the DPDP Rules 2025 and their phased compliance timeline: india-briefing.com

Platform and vendor research. Commercially interested parties with large samples. Directional, not definitive:

[23] Greenhouse, The Hire Standard benchmark report, March 2026. 640 million applications, 6,000+ companies, North America: greenhouse.com

[24] StepStone Arbeitsmarktbriefing, Dr Christina Langer, 21 May 2026: stepstone.de

[26] Ashby, Recruiting Operations Benchmarks: ashbyhq.com

[27] BioSpace, Q2 2026 Job Market Report: biospace.com

[28] StepStone, KI und Jobsuche, 26 November 2025: stepstone.de

Own research: The seventeen careers portals were checked directly against live public URLs on 29 August 2026, including bayer.eightfold.ai, astrazeneca.eightfold.ai, roche.wd3.myworkdayjobs.com, novartis.wd3.myworkdayjobs.com, iqvia.wd1.myworkdayjobs.com, takeda.avature.net and careers.smartrecruiters.com/Lonza.

© 1 September 2026 Andreas Schulz. All rights reserved.