The short version

  • A number of platforms now run a model over your application and put a written account of its fit in front of the recruiter. Lever markets one such layer; the complaint against Eightfold quotes Eightfold's own materials describing a large language model that scores applicants.
  • That is a different mechanism from keyword matching. It reads for meaning, so synonyms, phrasing and context register in a way exact-match scoring never allowed for.
  • Greenhouse told candidates in January 2025 that its AI does not score or rank applications, nor decide whether you advance. Its Talent Matching feature does sort applicants into match categories, but Greenhouse says it never auto-rejects or auto-advances anyone.
  • The practical shift: the summary becomes the thing a recruiter reads first. Your resume is increasingly the input to a document written about you.
  • What survives summarisation is specific, verifiable, self-contained claims. Vague, decorative and context-dependent lines do not.

Advice about applicant tracking systems has spent a decade describing a mechanism that is now only part of the picture. Keyword matching still exists. Sitting on top of it, in a number of products, is a language model doing something categorically different.

This matters less for the reasons usually given, which tend towards the apocalyptic, and more for one specific and quite practical reason: it changes what a recruiter actually looks at.

What the layer does

In the products that describe it, the model takes the parsed application and produces a short natural-language account of the candidate: what they have done, how it lines up with the requisition, what is missing. Lever's AI-powered screening page mocks up the fit-summary part of this: a candidate at a 94% match, tagged Strong Fit, beside an AI justification that opens with "Strong portfolio alignment with role requirements" and ends "Recommend advancing." Recruiters also use AI for the writing around screening: 33% of those in Employ's 2025 Recruiter Nation survey said they use it to craft interview questions.

This is not confined to vendor marketing. The class action complaint filed against Eightfold AI in January 2026 quotes the company's own materials describing a proprietary large language model built on more than 1.5 billion data points, drawing on LinkedIn, GitHub and resume databases, and producing a likelihood-of-success score from 0 to 5 that employers use to rank applicants. Those are allegations in a live case, but the description of the mechanism comes from Eightfold's published product claims rather than from the plaintiffs. Eightfold's own New York City disclosure describes the same 0 to 5 scale: a match score in half-point steps that can produce a ranked list of candidates for a role.

Two things follow immediately. First, it does not need your exact wording. Semantic matching recognises that platform reliability work and site reliability engineering are adjacent, which the older exact-match approach could not. Keyword stuffing gets worse, not better, against a reader that models meaning.

Second, and less comfortably, it introduces a document about you that you did not write and will never see.

What it is not doing

The obvious fear is that a model is now deciding, the same fear behind the belief that an ATS automatically rejects resumes. The vendors that address this directly say otherwise. Greenhouse's candidate-facing page, published in January 2025, says its AI does not score or rank applications and makes no decision about whether you move forward, and that applications are reviewed by real people in the order received.

Greenhouse now also offers Talent Matching, which does sort applicants into categories from Strong to Limited. Its FAQ is explicit that the feature does not auto-reject or auto-advance any candidate, including those in the lowest categories.

That is a narrow claim about one vendor. The same Greenhouse page concedes that some companies do use AI to analyse, score and auto-reject applications without any human input, and most vendors publish considerably less. But it is the clearest statement available from a major system, and it points at where the influence actually sits.

Core Principle
“A summary does not have to make the decision to shape it. It only has to be the first thing anybody reads.”

The distinction is real but thinner than it sounds. A recruiter reading a generated summary before opening the file forms an impression from that summary. It is not a decision made by the model. It is not an unaffected decision either.

Writing for a reader that summarises

If a summary is being generated from your resume, the useful question is what survives that compression. Summarisation keeps claims that are specific, concrete and self-contained. It discards padding, and it cannot rescue a line whose meaning depends on context you left out.

  • Make each bullet stand alone. A line that only makes sense after reading the two above it loses its meaning when extracted.
  • Put the outcome in the sentence, not implied by it. Action, context and result in one line summarises cleanly.
  • Name technologies, systems and domains explicitly. These are the anchors a summary retains.
  • Quantify where the number is real. Figures survive compression better than adjectives do.
  • Cut the qualifiers. Helped, assisted, contributed to and was involved in all compress to nothing.
  • Drop the keyword padding. A skills list disconnected from any evidence reads as unsupported to a model that weighs context.

None of this is new advice, which is the interesting part. Writing that summarises well is writing that reads well, and it was always what a busy human wanted. The language model layer mostly removes the last excuse for the other kind.

Recommended Action

See what gets extracted from your resume

Optimize your application materials in under 2 minutes.

Try it now

Everything above assumes the text came out of your file intact. A summary is generated from parsed output, so extraction failures propagate into a document about you that nobody will ever check against the original.

How much of this is actually deployed

Worth keeping proportionate. Where these features are switched on is uneven. Greenhouse's Talent Matching, for example, is available on its newer subscription tiers with the Real Talent add-on, so two employers on the same applicant tracking system can be running quite different screening.

Survey data points the same way. In the same Employ report, based on a Zogby Analytics survey of more than 1,200 US recruiters and HR managers in September 2025, 65% said they use AI to augment their recruiting technology. But the share using it for candidate matching fell to 40%, from 55% the year before, and writing tasks such as job description recommendations and communicating with candidates were each cited by 41%. Those figures measure what recruiters use, not what products offer. Employ owns Lever, so this is a vendor surveying its own market.

The direction is still consistent across the vendors that document anything. Greenhouse sorts applicants into match categories, Lever markets fit tiers with written justifications, Eightfold describes a 0 to 5 match score, and Workday grades applicants through HiredScore on an A to D scale, without publishing the model behind it. The adaptation this calls for is the same adaptation that helps everywhere else: be specific, be concrete, and make sure the document parses. That advice is stable regardless of which layer happens to be reading.

Frequently asked questions

Many do, for some tasks. In the 2025 Recruiter Nation survey by Employ, which owns Lever, 65% of more than 1,200 US recruiters and HR managers said they use AI in their recruiting technology, while candidate matching was cited by 40%, down from 55% a year earlier. Writing tasks, such as job descriptions and messages to candidates, were cited slightly more often than matching.

Some can. Greenhouse's candidate page concedes that some companies use AI to analyse, score and auto-reject applications without any human input. It says its own auto-rejection runs on rules recruiters set and control, and its Talent Matching FAQ says that feature does not auto-reject or auto-advance anyone. The best-documented automated rejection comes from screening questions a recruiter configures.

Less than before, but phrasing still matters. Semantic matching recognises related roles and skills, and Greenhouse says multiple terms can map to one calibrated skill. Recruiters still run keyword searches over parsed text, though, so use the posting's own wording for standard skills, once, inside a bullet that shows the work. Placement beats repetition.

Sometimes, through the employer. Greenhouse says organisations can export a candidate packet with the Talent Matching result, including the match category and its reasoning, and that these exports help support candidate access requests. Whether you can get one depends on the employer and on the privacy law where you live. Other vendors document considerably less.

The way you would for a busy human. Make each bullet stand alone, put the outcome in the sentence, name technologies explicitly and quantify only where the number is real. Then check that the file parses with an ATS resume checker, because any summary or match is generated from the extracted text, not from your layout.

Yes, through HiredScore, which Workday describes as AI-driven candidate grading that spotlights the most relevant applicants for a recruiter. Workday defines each A to D grade by how many of the job's basic and preferred qualifications a resume meets, but does not publish the model behind it or the threshold for an A, and nothing it publishes describes the grade removing anyone. The grade affects your position in the recruiter's list. See how Workday's grading works.