Companies Screen Resumes Differently, and You Cannot Know Which
Yes, some companies use software in the first pass over applications. Others have a person read every one. Most sit somewhere between, with software sorting a queue that a human then skims. Which one you are facing is decided per employer, per role, and per hiring manager, is not published, and changes without notice. So the practical question is not “do they use AI” — it is what you can write that works whichever it turns out to be. That answer is stable, and it is short: state plainly what you did, in the words the reader is already holding.
What varies, and why guessing is pointless
The setup differs on at least four axes that you have no visibility into. Whether software ranks or only stores the applications. Whether anyone looks at that ranking. Whether the stated requirements are configured as hard filters or as scoring hints. Whether a recruiter, a coordinator, or the hiring manager does the human pass. Two roles at the same company, posted the same week, can differ on all four.
This is why articles asserting how resume screening works are describing one configuration at one company at one time, and why anything phrased as a universal figure — a share of employers who do it, a share of resumes rejected, a score you have to clear — is not a fact about the world. There is no single system for such a number to describe. How document parsing works as a technical subject is a legitimate topic in its own right; it is just not the thing that decides your application.
The one property that helps in every case
Everything that survives across configurations comes down to a single idea: the document should be unambiguous about what you did.
An automated first pass is looking for evidence that you match stated requirements. A human first pass, done fast, is looking for the same thing. Both fail on the same input — a page where your last role’s actual scope is buried, or implied, or phrased so generically that it fits any job. Neither is fooled by cleverness, and neither needs it.
Concretely, that means:
Say the role plainly, in the ordinary term for it. If your title was internal jargon, put the recognisable equivalent alongside it. This is a real and common problem, and using a different title than the one on your payslip has an honest form that solves it without misrepresenting anything.
Use the posting’s vocabulary where it is genuinely yours. If the advert says “incident response” and you called it “on-call,” write the phrase they used, because that is the phrase in the reader’s head and in whatever they configured. This is not keyword gaming; it is translating your history into their language. The limit is honesty: only for things you actually did.
State scope with a number that exists. Team size, number of sites, size of a budget you actually held. These carry more information per character than any adjective, and unlike performance figures you can usually source them.
Keep the structure boring. Standard section names, reverse-chronological, dates in one consistent format, one column, no text inside images. Every one of these helps a fast human read as much as it helps anything automated — the two audiences want the same document, which is the point most guides get wrong.
Put the qualifying evidence high. If the role requires a specific licence, language, or system, it should be visible without scrolling or searching. Anything that needs hunting for might as well be absent in either kind of pass.
What only works against a guessed-at machine
There is a second category of advice: things that could only help if you were right about the software, and that hurt you the moment a person reads the page. Dense keyword lists with no substance behind them. Repeating one term until it dominates. Pasting the advert’s text into the document. These are the tactics circulated as ways to get past screening, and each of them fails on its own terms — but note the simpler objection: they cost you the human read, which every application eventually gets if it is going anywhere.
The asymmetry is what matters. Writing plainly costs you nothing in the automated case and helps in the human case. Gaming costs you in the human case and is a bet in the automated one.
You can sometimes tell which pipeline you are in
The uncertainty is not total. The route you applied through carries real information about how your application is being handled — a quick-apply on a job board, a company portal with a long form, a named person’s inbox, and a referral are four genuinely different situations, and reading your own application channel is more useful than speculating about software you will never see.
The most useful observation there: the questions on the application form itself are the screen you can actually see. If it asks whether you hold a specific certification and offers only yes or no, that is a filter, it is visible, and it operates whatever else is running. It is also worth noticing what a form asks that your resume does not, since those fields are frequently the ones that decide the first cut.
What no configuration can fix
Two things are worth naming, because people spend effort on formatting when this is the actual problem.
First, you may not meet a hard requirement. No arrangement of words gets a document past a genuine legal or licensing gate. That is not a screening problem; it is an eligibility one, and the right response is to apply where you qualify or to acquire the thing.
Second, the claims still have to hold. A resume optimised into the first pass and then unable to survive a conversation has cost you more than it gained, because the failure happens later, in front of a person, about something you wrote. Whatever you do to get read, every line still has to be one you can walk through from memory, and every figure still needs a source you could point at.
The version to write
One document, boringly formatted, that states your last role’s scope in the first few lines, uses the ordinary words for what you did, carries one specific instance per role, and contains no claim you could not defend. That page performs identically well against a careful human, a rushed human, and a ranking algorithm — not because it is optimised for all three, but because all three are trying to find out the same thing, and it tells them.