How AI screening actually works — and what it means for your application
There are two stories about AI in hiring, and both are wrong. The first is that a sophisticated intelligence reads your resume, understands your career, and renders judgement. The second is that it is all a con and nothing has changed. The reality is narrower and more useful to know: employers use several distinct tools that do quite different things, and confusing them leads to advice that actively wastes your time.
The oldest and most widespread is the applicant tracking system, and it is the one most misunderstood. In the great majority of configurations an ATS is a database with a search box. It ingests your file, parses it into structured fields — name, employers, dates, titles, education, skills — stores it, and lets a recruiter search and filter. It is a filing cabinet with a query language, not a judge. The persistent claim that it silently rejects most applicants before any human sees them traces back to vendor marketing rather than research, and treating it as a gatekeeping intelligence produces exactly the wrong behaviour: keyword stuffing, invisible white text, and other tricks that fail the moment a person reads the document.
What genuinely goes wrong at this stage is duller and more fixable. Parsers mangle multi-column layouts, because they read in an order that made sense visually and nonsense linearly. Text inside images is invisible to them. Unconventional section headings mean a parser cannot find your experience. Dates in an unusual format break the timeline. None of this is an algorithm judging you; it is a program failing to read a file. A single-column layout with standard headings and real selectable text removes almost all of it.
The second tool is keyword and requirement matching, and this one does score you. A recruiter searches for the skills in the posting, and applications are ranked by how well they match. This is why tailoring works and why generic applications underperform: if the posting says "stakeholder management" and your resume says "worked with other teams", you have described the same experience in language the filter cannot connect. Aligning your vocabulary to the posting where it genuinely describes what you did is not gaming the system; it is answering the question that was asked.
Alongside this sit knockout questions, which are the most decisive and least discussed part of the process. Work authorisation, years of experience, degree requirements, location, licences, notice period. These are usually binary and usually automatic. A candidate who is perfect on every other axis and answers no to a hard requirement is filtered regardless. This is why checking the knockouts before you invest an hour in tailoring is the highest-return two minutes in any application.
The newest layer is genuine language-model screening, and it is worth being precise about what it does. Where deployed, it summarises applications, drafts a comparison against the job description, and sometimes conducts or scores an initial interview. It is more capable than keyword matching — it can recognise that a candidate's described work is relevant even in different words — but it is also newer, less consistent, and increasingly subject to regulation. Several jurisdictions now require disclosure or auditing when automated tools materially affect hiring decisions, which is one reason adoption is more cautious than the marketing suggests.
The practical consequence of all this is that you are writing for two readers at once, and the advice for them is far less contradictory than it appears. Both want the same document: clear structure, conventional headings, the relevant work near the top, specific claims with numbers attached, and the vocabulary the role actually uses. The machine can parse it and the human can scan it in the few seconds they will give it.
What you should not do is optimise for a model of the software that is wrong. Keyword density targets, hidden text, elaborate ATS-beating templates, and paying for a score out of a hundred all rest on the assumption that a hostile algorithm is the obstacle. Usually the obstacle is more ordinary: a mismatch on a hard requirement, a resume that does not make its relevance obvious in the first third of the first page, or a hundred other applicants who tailored theirs.
One asymmetry is worth exploiting. Because AI has made it trivial to generate applications, volume has risen sharply and the average application has become less distinctive. That devalues volume and increases the return on the things automation cannot fake: a specific, verifiable achievement; evidence you understood the company's actual problem; a short message to a named person rather than a form submission. The tools that let everyone apply to five hundred roles have made applying to five hundred roles a worse strategy than it was.
If you want a single rule: write the document a competent human would want to read, formatted so a parser cannot mangle it, using the words the posting used, and check the knockouts before you start. Everything else in this space is either a refinement of that or a distraction from it.
It is also worth being sceptical of any tool that returns a confident number without telling you what it measured. A score out of a hundred feels authoritative and is usually the least useful output available, because it collapses several different situations into one figure — you match none of the requirements, the posting listed no requirements a parser could recognise, or the tool simply has a narrow vocabulary. Those demand completely different responses from you, and a single percentage hides which one you are looking at. What you actually want from a checker is the breakdown: which required skills it found in your resume, which it did not, and which hard conditions you fail outright.
The same scepticism applies to claims about beating specific systems. Vendors advertising that they have reverse-engineered the major platforms are describing something they cannot verify, because those parsing rules are neither published nor stable. The durable advice has not changed in a decade and does not depend on any of it: conventional structure, real text, specific evidence, and honest alignment to what the role asked for.
Frequently asked questions
Do most resumes really get auto-rejected before a human sees them?
There is no good evidence for the widely repeated figures. Most applicant tracking systems filter and rank rather than reject outright, and the claim is best treated as marketing. Hard knockout criteria like work authorisation genuinely can be automatic.
Does keyword stuffing work?
No, and it backfires. Ranking may improve briefly, but the document then reaches a human who can see the padding, and any claimed skill will be probed at interview. Match vocabulary only where it describes work you actually did.
Should I use a plain template or a designed one?
Plain, single-column, standard headings, real text. Columns, text boxes and graphics are the most common cause of a resume parsing badly, and design almost never compensates for that.
Will AI interviews replace human ones?
Not broadly, and regulation is moving toward requiring disclosure and auditing where automated tools materially affect decisions. Treat an AI screening round as a real round and prepare for it the same way.
Sources
Every figure cited above links to its origin. Where a widely-repeated statistic has no study behind it, we say so rather than repeat it.
Try JobStraight free →