Hiring, and the filters · 4.2

Applicant tracking, and what it filters

Applicant tracking, and what it filters. What is actually the case, and how it compares with what is repeated.

A practical software reference for this part of the discussion is Monitask's overview of employee monitoring system.

A widely repeated claim holds that three quarters of applications are rejected by software before any person sees them. It has no identifiable source, and it describes the systems inaccurately.

What the software actually does

Applicant tracking systems are databases. They receive applications, parse them into fields, store them, and let a recruiter search, filter and rank. The great majority of rejections are made by a person applying a filter, not by the system acting alone.

The distinction matters because the remedies differ. If a machine is rejecting you, the answer is formatting. If a person is filtering on years of experience, the answer is different and no amount of keyword optimisation touches it.

What does reject automatically

Knockout questions. Do you have the right to work here, do you hold this licence, do you have this many years in this field: answers below threshold are discarded without review, and that is a genuine automatic rejection.

Poorly designed knockouts are the commonest way a qualified applicant disappears, particularly where the question demands more years of experience than the technology has existed.

Parsing, which does fail

Text in tables, multiple columns, headers, footers, text boxes and images parses badly or not at all. A well-designed document can arrive as a scramble of fields, and the applicant never learns this happened.

Plain single-column text with conventional headings survives every parser. It is dull advice and it is the only formatting advice supported by how the systems work.

Ranking, and where bias enters

Where a system scores or ranks candidates using a model trained on past hiring decisions, it learns the past decisions, including the ones nobody would defend.

The clearest documented case is a large retailer's experimental recruiting tool, trained on a decade of applications, which learned to downgrade indications that an applicant was a woman. The company abandoned it. The lesson is not about that company; it is that a model fitted to historical hiring reproduces historical hiring.

The regulatory response

New York City has required bias audits of automated employment decision tools since 2023, with publication of results and notice to candidates. The European Union's artificial intelligence legislation places employment-related systems in its high-risk category, with obligations attached.

Both are recent, both are being tested in practice, and the main open question is the same for each: what counts as an automated decision when a person is nominally in the loop but is following a ranked list.

What a candidate can usefully do

Format plainly. Use the words in the advertisement, because filters match strings rather than meanings. Answer knockout questions carefully and accurately. Apply through the employer's own system rather than an aggregator where possible, since re-syndicated applications parse worse.

And treat the rest of the optimisation advice with suspicion, because much of it is sold by people who also sell the fix.

What an employer can usefully do

Audit the knockouts, which are usually inherited and frequently absurd. Look at how many applications the filters remove and sample the removed ones. Check parse failures, which most systems log and nobody reads.

And if a ranking model is in use, establish what it was trained on before establishing anything else.

Why the myth persists

Because it explains a real experience. Applications do disappear without response, in large numbers, and an explanation naming software is more comfortable than the accurate one, which is that a person spent four seconds and moved on.

The aggregator problem

Applications submitted through job boards are frequently re-syndicated into the employer's system through an integration that parses them a second time, and each parse loses something.

Applying directly on the employer's own site avoids one round of it, which is a small effect and free.

The response nobody sends

Most applications receive no reply of any kind. That is a choice about cost, made by employers who would send one if it were free, and the tracking system that filters the applications is also perfectly capable of sending a rejection.

The absence is not a technical limitation. It is a decision that the applicant's time is worth less than the message costs.

Testing your own pipeline

Submit a fabricated application to your own vacancy and see what arrives at the other end. Most recruiters have never done this and most are surprised by the result.

For broader context, consult New York automated hiring-tool rules.