TLDR: Artificial intelligence now sits on both sides of the hiring table. The advantage belongs to candidates who pair machine fluency with domain depth, and to employers who treat recruitment automation with the same governance discipline they apply to products.
Applicants and recruiters now run on the same class of tools, which changes what each side must prove
The symmetry is the defining feature of hiring in 2026. A candidate drafts a tailored application with a language model; a recruiter screens that application with software built on comparable technology. Both parties gained speed, and both lost a signal they used to rely on, which is why the criteria for standing out have moved. Effort used to carry information. A cover letter that clearly took an evening to write told a hiring manager something real about motivation, because the cost of producing it was high enough to be meaningful. Generative tooling collapsed that cost to nearly zero, so the same artefact now says very little about the person who sent it. The consequence is a shift in what recruiters weigh: specificity that only a genuine practitioner could supply, verifiable history, and the coherence between what a candidate claims and how they discuss it under questioning. Polish stopped functioning as evidence.
Employers absorbed a mirrored problem. Application volumes rose sharply once applying became cheap, which pushed more organisations toward automated first-pass filtering simply to keep pace with the queue. That filtering then produces its own distortion, because a system optimised to find matches in text rewards candidates who write for the system rather than candidates who did the work. The loop tightens from both ends: applicants tune their documents to the filter, the filter loses discriminating power, and employers respond by filtering harder still. The practical escape route runs through the parts of assessment that resist automation entirely — structured conversation, work samples, and reference evidence from people who watched the candidate operate under real pressure.
In regulated industries the stakes attached to that loop run higher than elsewhere. A mis-hire in a quality, safety or regulatory affairs function carries consequences that reach an inspection, and an inspector asking how a decision was made expects a defensible answer about the person who made it. Tolerance for a screening process nobody can explain is correspondingly low, and the same instinct that makes a pharmaceutical company validate its manufacturing systems makes it uneasy about a hiring filter it has never tested. Both sides of the table therefore face the same underlying task, which is proving something that used to be safely inferred. The rest of this piece takes them in turn, starting with the candidate’s controllable variables.
Machine screening rewards structure, and structure is the variable a candidate fully controls
For an applicant, the first reader is frequently software, and software reads a document in a specific mechanical way. Understanding that mechanism turns CV formatting from a matter of taste into a solvable engineering problem, and it costs a candidate nothing in authenticity. Parsing software extracts text and maps it to fields, which means the layout choices that impress a human designer frequently destroy the machine reading. Multi-column layouts interleave unrelated sentences when the text is extracted linearly. Skills rendered as graphics or rating bars carry no extractable text at all. Dates placed in headers or images vanish from the employment history the system reconstructs. A single-column document with conventional section labels, real text throughout and a standard file format survives extraction intact — and it reads perfectly well to the human who opens it second.
Terminology deserves the same deliberate treatment, particularly in regulated functions where the vocabulary is specific and non-negotiable. A quality professional whose CV says “manufacturing standards” describes real experience less legibly than one who names Good Manufacturing Practice directly. Clinical operations experience registers when a document names ICH-GCP; MedTech regulatory work registers when it names the Medical Device Regulation or, for diagnostics, the In Vitro Diagnostic Regulation. The mechanism is worth understanding rather than merely obeying: matching systems index on those tokens because the tokens are the only unambiguous evidence in a document otherwise full of paraphrase, and interviewers reach for the same terms as a first gauge of depth. Use the exact terms your actual work involved.
Where this discipline stops is the boundary worth respecting. Optimising structure and vocabulary makes true experience findable; padding a document with terms you have never worked with makes the first interview a demonstration of the gap. Regulated hiring conversations move quickly to specifics — which submission, which inspection, which deviation, which market — and the distance between a keyword and a lived experience becomes obvious within minutes, because a practitioner answers with texture while a borrower answers with a definition. The tooling helps you present what you have, and the interview tests what you have. Candidates who hold that line arrive with nothing to defend, which is a materially easier position from which to be persuasive.
Preparation is where generative tools earn their keep, and judgement is where a candidate wins the role
The most productive use of a language model in a job search happens before the conversation rather than during it. Treated as a rehearsal partner, the tooling compresses preparation time considerably; treated as a script, it produces exactly the generic answers that trained interviewers screen out. Rehearsal works because it forces retrieval. Feed a model the job description and ask it to interrogate you on the specific competencies the role names, then answer aloud from memory before reviewing anything. The exercise surfaces the stories you cannot yet tell cleanly, which is diagnostic information you can still act on with a week to go. It also stress-tests your reasoning, because a model prompted to challenge an answer will ask why you chose one approach over an available alternative — the same question a hiring manager asks when probing whether a decision was reasoned or inherited.
What decides regulated hires sits beyond anything a model can rehearse for you. Interviewers in pharma, MedTech and finance test how a candidate behaved when evidence was incomplete and the clock was running: the deviation discovered late in a batch record, the claim a commercial team wanted that the data would not support, the moment a submission timeline collided with a genuine safety question. Those probes work because the answer exposes a candidate’s default under pressure, and defaults are what an employer is actually buying. The answers draw on episodes you personally lived through, including the parts that went badly and the parts you would handle differently now. That is the material worth preparing hardest, because it is the material nobody else can supply.
Verification closes the loop. Any figure, framework or regulatory reference a model produces belongs in your answer only after you have confirmed it against a primary source, since a fluent hallucination delivered confidently to a regulatory affairs director costs the entire interview. The risk concentrates precisely where the tooling feels most helpful — article numbers, statutory timelines, guideline names — because those are the details a candidate is least able to check from memory and an expert interviewer is most able to check instantly. Candidates who apply this discipline arrive better prepared and more accurate than they would have been alone, which is precisely the balance the employer side of the table is now trying to strike.
GDPR and the EU AI Act put recruitment automation inside a documented compliance perimeter
For HR leaders in Europe and Switzerland, the legal position shapes the tooling decision before any vendor comparison begins. Candidate data and automated assessment both sit inside frameworks that expect an organisation to know what its systems do and to be able to show it. Data protection comes first because it applies to every applicant record regardless of automation. Under the General Data Protection Regulation, candidate information requires a lawful basis, a defined retention period and transparency about how it is processed, and the European Data Protection Board has consistently addressed how automated decision-making interacts with those obligations. Swiss employers face parallel duties under the revised Federal Act on Data Protection. The operational consequence is concrete: an applicant tracking system quietly retaining rejected candidates indefinitely, or a screening vendor processing data outside the agreed scope, creates exposure independent of whether the model performs well.
The EU AI Act treats certain employment uses as higher-risk, which raises expectations around documentation, human oversight and transparency for systems used in recruitment and selection. The reasoning behind that classification is worth internalising rather than merely complying with: hiring decisions materially affect a person’s livelihood, they are difficult for the affected individual to contest because a rejected applicant sees only the outcome, and errors replicate silently across thousands of applications where a single human reviewer’s mistake stops at one. An organisation that can name which tools touch a hiring decision, what each one does, and who reviewed its output is positioned very differently from one discovering its own stack during an audit.
Practical governance therefore starts with an inventory. List the systems that influence candidate progression, record what each contributes, retain the vendor’s documentation, define who holds decision authority at each stage, and test outcomes for adverse impact across groups at a set cadence. Building that list is usually more revealing than expected, because the systems influencing progression rarely stop at the applicant tracking platform — a sourcing tool, a scheduling assistant and an assessment vendor each shape who reaches a human, and each was frequently procured by a different person in a different year. The inventory answers the regulator, and it also answers the more useful internal question of whether the process is actually working.
Regulated employers already know how to govern a system they depend on
The advantage available to pharma, MedTech and financial-services organisations is that the governance discipline exists internally already. Companies operating under Swissmedic, EMA, FDA and FINMA scrutiny validate systems, document decisions and keep humans accountable as routine practice, so extending that habit to recruitment technology requires transfer rather than invention. Validation transfers most directly. A regulated manufacturer would never deploy a system affecting product quality without evidence that it performs as intended in its specific context of use, and the same logic applies to a screening tool. Ask the vendor what population the model was evaluated against, run a parallel test on a historical requisition where the outcome is known, and compare the shortlist the system produces with the one your team produced.
Where the two shortlists diverge sharply, you have learned something valuable about the tool, the role definition, or both — and the role definition is the more common culprit, since a requisition written loosely gives a matching system nothing stable to work against. Human accountability transfers next, and it needs to be located precisely. A named person should own each rejection decision at each stage, in the same way a named individual signs a batch release or a regulatory submission. That placement matters because accountability diffused across a process becomes accountability held by nobody, and automation makes diffusion effortless: the system suggested it, the recruiter accepted it, the manager never saw it.
Naming the owner restores the scrutiny that makes the whole arrangement defensible, and it also clarifies what the automation is for. Handled this way, the tooling redirects scarce recruiting capacity toward the work that decides outcomes — understanding a hiring manager’s real constraint, assessing judgement in conversation, and holding a strong candidate’s interest through a long regulated process in which competing offers arrive faster. Speed becomes the input, and human attention becomes the output. Organisations that make that trade deliberately are the ones that will still be filling their hardest roles when the tooling stops being a differentiator, because by then every competitor will run comparable software and the difference will sit entirely in what people did with the time it returned.
Edward Galle combines AI-powered matching with specialist human assessment across pharma, MedTech, life sciences, finance and brand-tech roles in Switzerland, the EU and the US. Employers can review our services for companies or talk to our team, and professionals can submit a CV for matching.
References
- European Commission — The EU Artificial Intelligence Act. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai
- European Data Protection Board — automated decision-making and GDPR. https://www.edpb.europa.eu/
