Skip to content

AI-Powered Career Tools

how ai can help you pass ats resume screening systems

Use AI to review resume formatting and align verified experience with job requirements. Understand the differences between parsing, eligibility filters, and employer-side AI evaluations.

Updated September 28, 202615 min read7 sections

Share on XShare on LinkedIn

How AI Can Help With ATS Resume Screening—and Its Limits

AI resume tools can help review formatting and compare documented experience with a job posting. ATS documentation and hiring research explain the limits of that assistance: neither a readable file nor a high third-party match score guarantees an interview.

Use AI for specific editing and review tasks, and verify the results yourself. Employer requirements, applicant pools, and screening configurations determine the next stage.

An applicant tracking system (ATS) is recruitment software that helps employers collect, organize, track, and review job applications. Some platforms include resume parsing, automated eligibility filters, candidate search, and optional AI-assisted evaluation tools. The screening process depends on the employer’s configuration, so submitting a resume through an ATS does not necessarily mean AI has scored or rejected it.

Parsing, eligibility filtering, and AI-assisted evaluation are different functions. A formatting problem does not by itself establish that an employer rejected the application.

Can AI really beat an ATS?

Sometimes, and the wins are real even though they are narrow. The strongest documented uses are mechanical: extract the job description’s requirements, check whether your evidence actually covers them, and catch the formatting that breaks parsers.

An AI resume tool cannot reliably predict an interview outcome. It generally lacks the employer’s full selection criteria, screening configuration, applicant pool, and human-review context.

Third-party ATS resume-checker scores are estimates produced using each tool’s own methodology. They should not be treated as the score an actual employer will assign. Employer-side platforms may calculate different assessments based on role-specific criteria, eligibility questions, or AI models configured by hiring teams. Use resume-checker scores as diagnostic feedback, not as a prediction of interview selection. Ashby’s application-review documentation describes employer-side evaluations and criteria matching.

Before paying for a checker, review what it actually assesses, which file formats it inspects, and how it explains its results. Its evaluation is not a replication of every employer’s parser or selection process.

Here is what we would use an AI resume tool for, in order of usefulness.

  1. Reverse-engineering the job description. Feed it the posting and ask which requirements have hard evidence in your resume and which are vague. This is a comprehension problem, not a writing problem.
  2. Formatting review. Where the tool can inspect the original document layout, ask it to review your exported file for the failure modes below: contact details in a header, a two-column layout, text boxes, embedded charts.
  3. Rewriting supported content against a fact sheet you supply. Verify every claim, whether you revise one bullet or the whole document.
  4. Gap-checking against a published taxonomy. The U.S. Department of Labor’s O*NET database, currently at release 31.0, gives you the plain-language task and skill statements that recruiters themselves use for job coding. Matching your ATS resume bullets to real task language beats resume keywords stuffing.
  5. Running an adversarial pass. Ask the model to act as a sceptical screener with a reason to reject you.

That is the whole list. It is short because the evidence is short.

What does an ATS actually do to your resume?

ATS-based hiring may involve application collection, resume parsing, eligibility screening, candidate search, automated evaluation, and human review. These steps do not occur in the same order or combination at every organization. Some employers use knockout questions, while others use AI-assisted tools to evaluate documented qualifications. An ATS can also function without automatically ranking every applicant.

Some ATS platforms support automated rejection based on employer-configured eligibility questions. For example, Greenhouse allows organizations to automatically reject candidates based on answers to specified application questions. However, this does not mean the system automatically rejects resumes for missing arbitrary keywords, formatting choices, or failing a universal ATS score. The cited documentation does not establish a universal 75% rejection rate.

That distinction matters for anyone running their own AI job search with automation. Auto-applying at volume to thousands of roles does not defeat an employer-configured filter, and some ATS platforms rate-limit or block scripted submissions outright.

Some platforms offer optional AI evaluations and ranking. Ashby documents criterion-level evaluations and the percentage of configured AI criteria met. Greenhouse also documents integrations that can add match information to profiles. Their presence in a product catalogue does not establish that every employer enables them.

Clear formatting can reduce extraction problems. Successful parsing does not establish eligibility, guarantee a ranking, or ensure that a recruiter reviews the application.

Which resume formats do ATS vendors actually accept?

Verified from vendor documentation in September and October 2026. File limits differ more than most articles admit.

Vendor Accepted formats Documented size limit Source and last updated
Greenhouse Recruiting .doc, .docx, .pdf, .rtf, .txt 100 MB upload; parsing fails above 2.5 MB Supported formats, 19 Aug 2026; Unsuccessful resume parse, 2 Mar 2026
Workable .doc, .docx, .pdf, .rtf, .html, .odt on its application form 5 MB resume upload limit Application-form file limits; Resume-upload extraction
SmartRecruiters Parse API returns UNPARSABLE_RESUME for image-only files and INCOMPLETE_RESUME when required fields are missing Not stated in public API docs Parse a resume, 21 Apr 2026
Ashby candidate.uploadResume and AI Application Review documented in public API Not stated in public docs candidate.uploadResume, 27 May 2026
Lever, iCIMS, Workday, Oracle Taleo, UKG, SuccessFactors Documentation sits behind login walls; we could not verify limits Unverified —

We could not open Lever’s, Workday’s, Oracle’s or UKG’s candidate-facing upload documentation without an account. Rather than repeat numbers we cannot source, we leave them blank. That is the same standard we apply to statistics.

Greenhouse accepts candidate uploads up to 100 MB but cannot parse resumes above 2.5 MB. An unsuccessful parse may leave the original file attached while missing fields require manual entry. Workable documents extraction from image-based CVs, so such files are not universally unreadable; extraction quality and available fields vary by function and platform.

Does a PDF resume get rejected by ATS?

PDF files are supported by many major ATS platforms, but extraction quality varies with the document’s structure, text layer, and the platform’s parsing capabilities. A PDF upload being accepted does not guarantee every field will be extracted correctly. Follow the employer’s format instructions first.

A published study of roughly 200,000 real resumes collected over several years found that about 1% contained hidden prompt injections aimed at AI screeners, and prevalence rose noticeably over the preceding one to two years — with more than 90% of the injected prompts avoiding explicit instruction language. Source: Zhang et al., published at USENIX Security 2026, 27 May 2026.

The approximately 1% figure describes concealed prompt injections in that study’s dataset, not a measured prevalence across all resumes. It does not establish that adding such instructions reliably improves hiring outcomes.

The practical PDF rule: a real text layer, standard fonts, no scanning, and nothing important in a header, footer or text box.

ATS-friendly and ATS-hostile formatting: what breaks parsers

Greenhouse’s own troubleshooting article is more specific than anything written by a career blogger, so we are using its list as the skeleton.

A formatting element is ATS-hostile when the parser cannot recover the reading order or the field value.

Formatting element Verdict Why
Contact details in the page header or footer Hostile Greenhouse names this as a direct cause of failed parses
Contact details in a text box Hostile Same
Two-column layout Hostile Reading order breaks
Tables for skills or dates Hostile Greenhouse names tables explicitly
Charts, graphs, progress bars for skills Hostile No extractable text
Photos, icons, logos, word art Hostile Greenhouse lists graphics and photos as causes
Letter-spaced text Hostile The parser reads separate letters, not a word
Non-standard section headings (“My Journey”, “Where I’ve Been”) Risky Greenhouse flags resumes without clear sections
Abbreviated titles (“Sr. Account Exec”) Risky Greenhouse flags incomplete job titles
Company names without legal suffixes (no Inc., Co., Ltd, LLC) Risky Greenhouse names this
Embedded or web-linked images Hostile Greenhouse does not load externally linked images for PDF preview
Special characters, ligatures, emoji Undocumented Greenhouse notes some imports fail on unrecognised characters, without specifying which
Standard headings, single column, one font Lower risk Generally easier to extract; verify the exported file

These entries summarize documented risks, mainly from Greenhouse. They are not a guarantee of failure or compatibility across every parser; check the actual exported text and populated fields.

How many keywords do I need on a resume?

There is no published number, and any article that gives you one is guessing.

The cited documentation does not establish a universal keyword density or match percentage needed for selection. Some employer-side tools evaluate role-specific criteria. Use accurate job terminology and supported evidence rather than trying to satisfy an assumed quota.

What is defensible:

  • Use the job description’s exact vocabulary for things you genuinely did. “Patient intake coordination” beats “client-facing duties” if that is what you did.
  • Cover the requirements, not the whole posting. Copying the entire job description into your resume reads as padding to a human and does nothing verifiable for a parser.
  • Match the taxonomy recruiters use. O*NET gives you published task statements for occupations; ESCO does the equivalent for Europe. Use them as a vocabulary source.
  • Spell out abbreviations once. “Search Engine Optimization (SEO)” gives a matcher two forms to work with.

If a tool tells you to hit 75% keyword match, that number is the vendor’s own guess about their own model. Fine as a rough target. Not a fact.

What does research show about racial bias in AI resume screening?

Research shows that some AI-assisted hiring systems can produce racially disparate screening outcomes. A 2026 study examining millions of applications processed by a single hiring technology vendor found significant adverse-impact patterns affecting Black and Asian applicants. The findings demonstrate a serious risk of unequal outcomes, although they should not be generalized to every employer, ATS, or AI hiring system. This study alone does not establish discrimination against older applicants or that every rejected applicant was qualified.

The best single fact in this article:

In a study of 3.4 million people submitting 4 million applications to 1,700 postings across 150 employers, 26% of Black applicants and 15% of Asian applicants applied to positions where the vendor’s recommendations showed adverse impact against their group under the four-fifths criterion. These percentages describe applicants exposed to adverse-impact patterns, not the percentage of all applicants rejected. The researchers estimated that approximately 40,000 additional applications would have advanced if Black and Asian applicants had been recommended at the same rate as the most-favoured group. (Stanford HAI summary, original paper)

Two details in that release deserve attention. First, adverse impact disappears when you pool the vendor’s recommendations across all jobs, and reappears when you analyse each position separately, as the law requires. Aggregate results can obscure position-level disparities; adverse-impact evidence is distinct from a final legal determination. Second, the authors found an algorithmic monoculture effect: applicants submitting multiple applications to that vendor’s postings were rejected from all of them more often than independent employer decisions would predict.

The four-fifths rule is a screening heuristic, not a legal safe harbour. It comes from the Uniform Guidelines on Employee Selection Procedures, 29 CFR 1607.4(D), which states that a selection rate below four-fifths of the highest group’s rate “will generally be regarded by the Federal enforcement agencies as evidence of adverse impact,” while smaller differences can still qualify where statistically and practically significant.

The monoculture research is not the only evidence. A separate FAccT 2026 paper on human-AI collaboration found people viewing resumes for longer raised a not-recommended candidate’s selection chance by 3–4%, and up to 55.6% longer viewing time when no AI recommendation was shown. Earlier work from the same group found that when a biased AI recommends a group, people favour it up to 90% of the time. These experiments suggest that human review can be influenced by AI recommendations; they do not establish that every form of human review is ineffective.

What is the law on AI hiring in 2026?

Regulatory status as of October 2026. This area moved fast and most online summaries are stale.

  • EU AI Act. Employment AI remains high-risk under Annex III. The Digital Omnibus amendment delayed application of high-risk obligations from 2 August 2026 to 2 December 2027 for standalone systems, with 2 August 2028 for high-risk AI embedded in products. The Council gave final approval on 29 June 2026.
  • Colorado. SB 26-189, scheduled to take effect January 1, 2027, establishes protections for individuals affected by covered automated decision-making technologies. These include specified disclosure obligations, access to and correction of certain personal data, and opportunities to request meaningful human review and reconsideration after covered adverse decisions. The precise content of post-decision disclosures is subject to implementing rules. These are enacted protections with defined scope and conditions, not rights created solely by draft rules.
  • Illinois. The Illinois Human Rights Act amendments from HB 3773 took effect 1 January 2026, barring AI that discriminates on protected characteristics even unintentionally. IDHR published proposed notice rules on 15 May 2026, then withdrew and postponed them on 2 June 2026.
  • New York City. Local Law 144 has required bias audits and candidate notice since July 2023.

Colorado’s enacted law requires specified notices and post-decision disclosures for covered technologies and adverse decisions. Implementing rules clarify the content of those disclosures. This should not be presented as an unconditional right to an explanation for every rejection.

Under Article 86 of the EU AI Act, affected individuals may be entitled to a clear and meaningful explanation when a qualifying decision based on certain high-risk AI systems produces legal effects or similarly significant consequences affecting their fundamental rights. This protection is subject to the article’s scope, conditions, and exceptions. The high-risk AI obligations applicable to employment-related systems are scheduled to apply from December 2, 2027. This is not a general right covering every unsuccessful application.

The 10-step pre-submission ATS check

Run this in order. It takes about twenty minutes.

  1. Check the employer’s accepted file formats first. If the posting permits PDF or DOCX, use a clean, text-based document with standard headings and a logical reading order. Avoid image-only files unless the platform explicitly supports them. After uploading, verify that any automatically populated fields contain the correct information.
  2. Check the contact block. Name, email, phone, location in the body of the first page. Nowhere else.
  3. Confirm single column. No sidebars, no two-column skills lists, no floated text boxes.
  4. Scan for graphics. Delete charts, progress bars, rating stars, icons carrying meaning, and your photo unless the role is a performance one.
  5. Check file weight. Under 2.5 MB if Greenhouse is the application platform, since that is where its parser stops.
  6. Use conventional section headings. Experience, Education, Skills. Not “My Journey” or “Where I’ve Been.”
  7. Write full job titles. Senior Account Executive, not Sr. Account Exec. Add legal suffixes to company names where accurate.
  8. Remove letter-spacing. Some templates space letters for style. Parsers read them as separate characters.
  9. Mirror the job description’s vocabulary for work you genuinely did, and identify which required qualifications you can truthfully evidence. Do not invent missing qualifications.
  10. Have AI run a hostile pass, then read the whole thing yourself and cut anything you cannot defend in an interview.

Where the evidence does not support the advice

Three claims we could not verify, and will not repeat.

  • No published ATS rejection rate. The “75% of resumes are rejected by ATS” figure appears constantly and traces to no primary source we could find.
  • No universal keyword threshold. A third-party match target is not an employer-wide shortlist rule. Employer-side criteria assessments can exist without a universal keyword quota.
  • No universal AI-authorship outcome. A result from a particular model or research setup should not be treated as proof that every employer rewards or penalizes AI-assisted resumes.

The sources support specific conclusions: O*NET provides a published occupational taxonomy; Greenhouse documents parsing risks and eligibility-question auto-rejection; Ashby documents optional AI criteria evaluations; and the Stanford research identifies adverse-impact and correlated-rejection patterns in one vendor’s data. Colorado’s enacted law establishes covered disclosure, data-correction, and human-review protections from January 1, 2027, with detailed disclosures subject to implementing rules.

The gap between what gets sold as ATS advice and what vendors actually publish is wide. We would rather tell you that than fill it with a confident number. Anyone offering genuine ATS optimization as a service owes you the same disclosure: which vendor’s parser did they test against, on which date, and with what error output.

Sources

Filed under

AI-Powered Career ToolsATS Optimization

Share on XShare on LinkedIn

Keep reading