Pharma Bloodbath Part XV: everyone fears the AI that reads your CV. Almost nobody knows what it actually rejects you for.

Dear MoreThanCareer.de community,

save this article. This is not about flattering me. It contains the specific screening criteria that decide whether a human ever reads your application, and you will want them open in a second tab the next time you apply. If you would rather have someone go through your CV against these criteria with you, that is exactly what I do at MoreThanCareer.de, from career coaching and interview preparation to reverse recruitment and full CV and LinkedIn rebranding.

Now to the topic. Two audiences today: the people who fear AI will take their job, and the people who suspect AI already took their job application. Both groups deserve better information than they are getting.

The anxiety numbers, and one delicious contradiction

Exactly two thirds of Germans, 66%, expect artificial intelligence to reduce jobs in Germany. Only 5% expect AI to create more jobs than it destroys, and 24% expect no effect at all.¹ That is the result of a representative Forschungsgruppe Wahlen survey for ZDF, and it is about as close to a national consensus as this country gets on anything that is not Feierabend.

Here is the part I find far more interesting: asked about their own job, 58% of the same respondents are relaxed. Only 6% fear their personal position will fall to the machine.¹ Two thirds of the country expects a storm, and almost everyone assumes the lightning will hit the neighbour’s house. Das Sankt-Florian-Prinzip (the Saint Florian principle: spare my house, burn the other one) has apparently completed its digital transformation ahead of schedule.

The company side is less relaxed. In the ifo Institute’s survey, 27% of German firms expect AI to eliminate positions within five years, and in industry, which includes our beloved pharmaceutical sector, the figure rises to 37.3%.² One widely circulated projection puts the number of German jobs directly affected by AI over fifteen years at around 1.6 million, with job creation and reduction roughly balancing out long term.³ Treat that 1.6 million with care: it is a model projection, not a measurement, and models about 2040 have the same reliability as weather forecasts about 2040. The ifo data is a survey of what firms actually plan. That one deserves your attention.

What the machine actually rejects you for

Now the screening side, where I can offer something the anxiety debate rarely does: seventeen years of watching these systems from the inside, fourteen of them in pharma and CRO environments.

First, a myth burial. The claim that 75% of CVs are rejected by the ATS before a human sees them traces back to a marketing claim from a long-defunct software vendor, and serious analyses have repeatedly failed to substantiate it. The robot uprising in recruitment has been postponed. What actually happens is duller, and dullness is good news, because dull systems can be beaten.

The best available research here remains the Harvard Business School and Accenture Hidden Workers study: over 90% of employers use software to make a first cut or ranking, and 88% admit that qualified candidates get screened out because they do not match the exact configured criteria.⁴ Read that sentence again slowly. The candidates were qualified. The criteria were exact. The mismatch was vocabulary.

The four gates where applications actually fail

So what are those criteria? In practice, applications fail at four gates.

• Knockout questions. Answer no to a hard requirement, visa status being the classic for international candidates, and the rejection is automatic and instant, which is why some of you receive a no four minutes after applying at 23:47 on a Sunday.
• Hard filters a recruiter configured. Roughly half of employers in the Harvard study screened out CVs with employment gaps of six months or more.⁴ Not because anyone hates you. Because someone under volume pressure set a tick box in 2019 and nobody has reviewed it since.
• Keyword mismatch. You wrote drug safety, the system searched pharmacovigilance, and as far as the software is concerned you have never worked in the field.
• Parsing failure. Your beautifully designed two-column CV with the photo and the skill bars arrives in the database as alphabet soup. Wer schön sein will, muss leiden (who wants to be beautiful must suffer), but preferably not in the applicant tracking system.

Notice what is missing from that list: malice, and mostly, intelligence. These are volume-management rules written by humans, executed by software, at a scale where nobody audits the casualties.

The bias nobody configured on purpose

It gets less comfortable where actual AI enters. University of Washington researchers tested leading language models on over 500 real CVs and found the models favoured white-associated names in 85% of comparisons and female-associated names in 11%, and never preferred a Black-male-associated name over a white-male-associated one.⁵ Identical CVs. Different names. The machine learned our history and mistook it for a job requirement.

The law is coming. Slowly. With paperwork.

Brussels noticed. Under the EU AI Act, software used for recruitment, screening and candidate ranking is classified as high-risk, which triggers duties around bias testing, documentation and human oversight for providers and for the employers deploying the tools.⁶ Those obligations were due to bite on 2 August 2026, in other words, next week. In June, the EU agreed to postpone them to December 2027 under the so-called Digital Omnibus, with final formalities still pending as I write.⁷ The regulation arrived punctually and the enforcement is running eighteen months late, which is the most European sentence I will write this year.

The Americans, lacking a federal law, are solving this the American way: litigation. In Mobley v. Workday, a court has allowed claims to proceed that an AI screening vendor can be directly liable, as an agent of its client employers, for discrimination by its tools, and a collective action for rejected applicants aged 40 and over has been certified, with core claims cleared to proceed in June 2026.^{8, 9} Nothing is proven yet. But every vendor’s legal department read those rulings, and every serious employer’s procurement team should have.

Elsewhere: New York City has required bias audits since 2023, the UK has no AI law and is governing through regulator letters and existing equality legislation, China runs an algorithm filing regime, and India currently regulates AI hiring about as much as it regulates the weather. If you apply internationally, the protection around your application varies enormously by postcode.

Five things that actually work

Deliberately practical, and each one traceable to the mechanics above.

1. Mirror the exact vocabulary of the posting, including both acronym and full term. Pharmacovigilance (PV), Regulatory Affairs (RA), the exact job title where truthful. This is not gaming the system. It is translating your experience into the search terms a human configured.⁴ Juniors: this matters most for you, because you have the least material for the ranking to work with.
2. Treat knockout questions as the actual interview. Read them twice. If a question asks about visa sponsorship, answer accurately, and simultaneously target employers who state that they sponsor, because a knockout question at a non-sponsoring employer was never a door, it was a wall with a doorknob painted on. My honest advice for international candidates: research sponsorship policy before applying, not after rejection number forty.
3. Close visible gaps with one honest line. Given that roughly half of employers filter six-month gaps,⁴ a labelled entry such as 2024: family care or 2023: professional reorientation and certification converts a filter trigger into a biography. You owe nobody your medical history. You owe the parser a continuous timeline.
4. Make your CV boring to machines and interesting to humans. Single column, standard headings, no photo dependence, contact details in the body: that is the core of every CV and LinkedIn rewrite I write. Save the design ambition for your portfolio. In der Kürze liegt die Würze (brevity is the spice), and in the parsability lies the interview.
5. For experienced professionals and the 50+ group: remove data that only feeds proxies. Graduation years and roles older than fifteen years add filtering risk and nearly zero information. The Workday litigation is precisely about age-linked proxies in screening.⁸ Until the law catches up, do not volunteer the inputs. Your last fifteen years, framed as risk management for the employer, are the product. Executives: the same logic applies one level up, because executive search firms increasingly run AI-assisted longlisting too, and your public expertise trail is what those systems find.

For recruiters and hiring managers: the delay is not a holiday

The postponement to December 2027 moved the deadline, and moved none of the liability, because discrimination was already illegal under existing law on both sides of the Atlantic.⁸ Four things worth doing now, while your competitors wait for the fine print. Audit your knockout criteria and gap filters, and ask when anyone last checked what they exclude; the Harvard study’s own recommendation is to define six to eight minimum skills that filter candidates in rather than out.⁴ Demand bias-testing documentation from your vendors today, because a vendor who cannot produce it in 2026 will not magically produce it in 2027.⁷ Give your human oversight real authority; a reviewer who rubber-stamps the ranking is decoration, and the AI Act explicitly says so.⁶ And run adverse-impact analysis on your own candidate data, not the vendor’s demo set. The recruiting teams that do this will hire the qualified people their competitors’ filters silently discard. That is not compliance. That is arbitrage.

The bottom line

The 66% are right that AI is changing the labour market, and mostly wrong about how. The near-term risk to your career is not a superintelligence taking your job. It is a mediocre filter, configured by a stressed human in a hurry, quietly sorting your application into the void while the regulation that would govern it waits for a publication slot in the Official Journal. You cannot vote on the timeline. You can make your application legible to the machine and compelling to the human behind it, this week. Hilf dir selbst, dann hilft dir Gott (help yourself, then God helps you). And where God is unavailable, a well-parsed CV performs surprisingly similar miracles.

Your turn

Four questions worth answering honestly, even if only for yourself.

• Have you ever received a rejection within minutes of applying? What do you now suspect triggered it?
• Recruiters and hiring managers: when did your organisation last audit its knockout criteria, honestly?
• International candidates: how has the sponsorship knockout question shaped your search in DACH?
• Readers over 50: have you removed graduation years from your CV, and did it change anything measurable?

Prost. Und passt auf euch auf.

Sources & References

Credible and primary sources:

[1] ZDFheute / Forschungsgruppe Wahlen: Sind Jobs durch KI in Gefahr? (May 2026)

[2] ifo Institut via Statista: AI and employment, sector comparison (37.3% in industry)

[4] Harvard Business School & Accenture: Hidden Workers, Untapped Talent (2021)

[5] University of Washington: racial and gender bias in AI resume ranking (AIES 2024)

[6] EU AI Act (official implementation resource): what the Act means for staffing and recruitment

[7] Ogletree Deakins: EU agreement to delay high-risk employment AI rules to December 2027

[8] Maynard Nexsen: Emerging liability for AI-driven hiring tools, Mobley v. Workday

[9] Wiggins Childs (plaintiffs’ counsel): court authorises notice in Mobley v. Workday

Vendor and directional sources (cited as directional only):

[3] Stellenmarkt.de: Jobabbau durch KI, wo steht Deutschland (1.6 million projection)

© 28 July 2026 Andreas Schulz. All rights reserved.