What happens after you press send

Your application is read several times before a person sees it. We mapped the route from the vendors' own manuals, a federal court file and two academic audits, and worked out the short list of things you can still control.

Written by someone who interviews for global-company roles, and once sat exactly where you sit.
Published July 2026. Every figure below is linked to its source at the foot of the page.

Derek Mobley submitted a job application at 12:55 in the morning. The rejection arrived at 1:50 a.m., fifty-five minutes later. He says the same thing happened more than a hundred times.

Those allegations sit in a federal complaint against Workday, whose software handles applications for thousands of employers. Mobley, who is Black, over forty and has an anxiety disorder, argues the pattern was not bad luck. In July 2024 Judge Rita Lin of the Northern District of California let the core of the case proceed, on a theory that had never quite been tested: that a software vendor can be liable as the employer's agent when its tools do the rejecting.

Workday's software is not simply implementing in a rote way the criteria that employers set forth, but is instead participating in the decision-making process by recommending some candidates to move forward and rejecting others.

Order on motion to dismiss, Mobley v. Workday, 12 July 2024

She added a sentence that explains why anyone outside the case should care. "Drawing an artificial distinction between software decisionmakers and human decisionmakers would potentially gut anti-discrimination laws in the modern era." In May 2025 the court preliminarily certified a nationwide collective of applicants aged forty and over. Arguing against that, Workday told the court that 1.1 billion applications were rejected using Workday in the period at issue. The case is still in discovery. Nothing has been proved.

But the number is not really about one company. It is a measure of how much of hiring now happens at machine speed, and how little of it a candidate ever sees. What follows is not a theory of that layer. It is the layer as its own builders document it.

VolumeThe pile got much bigger, very fast

Every account of modern hiring starts with volume, usually with a statistic nobody can trace. So here is one that can be traced. Greenhouse, an applicant tracking company, published a benchmark study in March 2026 drawn from its own platform: over 6,000 North American organisations and 640 million applications between 2022 and 2025.

In 2022 the average open role attracted 116 applications. By 2025 it attracted 244. Over the same period the number of recruiters per organisation in that dataset fell by more than half, from 10.4 to 4.6, so the load per recruiter rose from 146 applications to 746.

A second dataset points the same way from the other end. Ashby, a competing applicant tracking company, reported in April 2026 on 109 million applications across 247,000 jobs since January 2021. The share of applications that reach an interview, it found, has fallen from roughly 7 to 8 percent in 2021 to between 3.6 and 4.7 percent now.

Both companies sell software into this market, which is a reason to read their numbers carefully rather than a reason to ignore them. Both are reporting against their own commercial interest in a hiring process that feels manageable. And they agree.

The routeGate by gate

The software between you and a recruiter is not one system. It is a sequence, and the sequence matters, because the authority to end your application is not distributed evenly across it.

Read that sequence again and notice where the emphasis in public debate sits. Almost all of it is on gate three. Almost none of it is on gate two, which is the only gate where a piece of software both makes the call and ends the process.

Gate twoThe gate nobody calls artificial intelligence

Every major vendor now publishes an AI principles page, and they say broadly the same thing. Greenhouse: AI "can inform, summarize and surface insight, but it is never the final decision-maker." Its bias audit statement is blunter still: its matching tool "does not have the ability to automatically advance or reject a candidate." Its data processing FAQ adds that "humans are required to make the decision for each candidate."

Those statements appear to be accurate. They are also about gate three.

Gate two is a rule, not a model, so it falls outside every AI principles page ever written, and it is where an application is most likely to end without a person ever touching it. The detail that makes this concrete is a scheduling option. SmartRecruiters lets an employer set a delayed rejection: the candidate is marked as rejected immediately, but the rejection email is held back for one hour, two days, or five days. The default is one hour.

None of this is hidden. It is in the product manuals, published by the companies that sell the features, written for the administrators who configure them. It is simply not written for you, and nobody has an obligation to translate it.

DisclosureWere you told?

Mostly, no. In May 2026 Greenhouse published a survey of 2,950 active job seekers across the United States, the United Kingdom, Germany, Australia and Ireland. Unless it noted otherwise, the figures it reported reflect its 1,200 American respondents.

Nearly two-thirds had now been interviewed by an AI, up thirteen percentage points in six months. Seven in ten said they had not been clearly told upfront that AI would be evaluating them, and one in five only found out once the interview had started. Of those who completed one, just over half never heard anything back.

The lawThe law increasingly agrees with them

Disclosure is not merely a courtesy in a growing number of places. Illinois has required it since January 2020 for video interviews analysed by AI: an employer must notify the applicant beforehand, explain "how the artificial intelligence works and what general types of characteristics it uses," and obtain consent. Maryland has required a signed waiver since October 2020 before an employer may create a facial template during an interview.

New York City went furthest. Local Law 144 took effect in January 2023 and has been enforced since July that year. An employer using an automated employment decision tool must commission an independent bias audit conducted within the previous year, publish a summary of the results on its website, and notify candidates who live in the city "no less than ten business days before such use," along with the qualifications and characteristics the tool will assess. Candidates may request an alternative process.

Two of those entries come with an asterisk, and the asterisks are the point. Illinois's newer law, HB 3773, has been in force since 1 January 2026, but the state's Department of Human Rights withdrew its proposed implementing rules and cancelled the June 2026 hearing on them. The duty exists; the shape of a compliant notice does not.

The European Union classifies recruitment systems as high risk. Annex III of the AI Act names "AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates." Article 26 requires an employer to inform affected workers before putting such a system into service. Those duties were due to apply from 2 August 2026. Four days before this piece was published, a simplification regulation entered into force that moved them to 2 December 2027.

Colorado went further backwards. Its 2024 AI Act, which would have imposed a duty of reasonable care against algorithmic discrimination, was repealed and replaced in May 2026, six weeks before it was due to take effect, by a narrower transparency statute.

EvidenceWhat compliance looks like when you go and look

New York City's law is the most demanding of the lot, and it has been running the longest, which makes it the closest thing available to a natural experiment. Two research teams have gone looking for the disclosures.

The first, published at the ACM conference on Fairness, Accountability and Transparency in 2024, sent 155 student investigators to check 391 employers over two weeks in autumn 2023. They found 18 published bias audit reports and 13 published transparency notices. Eleven employers had both.

A second team, publishing at the same conference in 2025, took the opposite approach and collected every Local Law 144 bias audit it could find anywhere, from the law taking effect until early November 2024. It found 116. It also noted the comparison that gives the number its shape: audits existed for roughly 2 percent of the Fortune 500, while an industry report finds more than 98 percent of those companies use applicant tracking software.

Then, in December 2025, the New York State Comptroller audited the city agency charged with enforcing the law. In two years the Department of Consumer and Worker Protection had received two complaints. The agency reviewed 32 companies and identified a single instance of non-compliance; the state auditors reviewed the same 32 and identified at least seventeen. The agency's own explanation, quoted in the report, closes the circle: the law requires posting "if an employer determines it needs to comply."

So the honest summary of the disclosure picture is not that employers are breaking these laws. It is that, four years into the first serious attempt to make this layer visible, an ordinary candidate still cannot find out what read their application, and the machinery built to tell them has produced two complaints and 116 documents.

PracticalWhat you can actually control

Very little of the above is within your reach, and it is worth saying so before the part that is. You cannot audit a tool you were never told about. What is left is smaller than the advice industry implies and larger than nothing.

• Make the file survive the parse. Greenhouse publishes the list of what breaks its own parser, which is the closest thing to a specification anyone has released: graphics, photos, resumes uploaded as images, tables, headers and footers, text boxes, columned layouts, contact details placed in the header rather than the body, company names without an Inc. or Ltd., and abbreviated job titles. Its example is exact: write "Senior Account Executive," not "Sr. Account Exec." Files over 2.5MB are not parsed at all. A failed parse does not reject you; it means a recruiter has to retype your details, and at 746 applications each, that is not a favour you want to ask.
• Treat the application questions as seriously as the resume. This is gate two, the only place where the software's answer is final, and it is usually a dropdown rather than a sentence. Read every required-qualification question before you begin, answer them accurately, and understand that an honest "no" on a genuine requirement is a knockout you cannot write your way around.
• Use the posting's vocabulary only where it is true of you. The parse pulls skills, titles, dates and employers into fields. Getting your real experience described in the words the posting actually uses is a translation problem worth solving. Inventing the words is not, because gate four is a person, and the interview is where invented words go to die.
• Know which problems no document solves. Work authorisation, a minimum number of years, a required location. Those are gate two questions. When one of them is against you, the useful move is to change the target, not the wording, and anyone who tells you otherwise is selling something.
• Ask, where asking is a right. If you live in New York City and an automated tool is being used on you, the law entitles you to notice ten business days in advance and to request an alternative selection process. In the Greenhouse survey, 46 percent of candidates said they wanted the option of a human interview. Most had never been offered it.

EndingThe last gate is still a person

It is tempting to read all of this as an argument for writing to the machine. It is closer to the opposite. Gates one and two are mechanical, and mechanical problems have mechanical fixes: a layout that parses, an answer given accurately, a title spelled out in full. Those are worth an hour of care and then they are done.

Everything after that is a document being read by someone who has 243 others and will form a view in seconds. What survives that is not polish. It is specificity: a decision you made, a constraint you worked inside, a number that came from your actual week rather than from a rounding-up. The machine layer has made the pile bigger, the queue longer and the reader more impatient. It has not changed what a person recognises when they finally look.

Which sets a test you can apply yourself, tonight, without any of the software described above. Take each line of your resume and ask what question it invites. Then answer that question out loud. If a line cannot survive the question it provokes, it is not ready, whatever the parser thinks of it.

BrightNoon publishes a note a week on something we have actually seen decide an application. Our own work starts from the same test this piece ends on: we do not write a line you cannot answer for in an interview, and clients receive the record of where every line came from. More about how that works.