AI resume screening that explains itself and leaves the decision to people

A model reads each application against a short list of job-related requirements the hiring manager wrote down, quotes the evidence it found for each one, and hands the recruiter a sorted queue. The recruiter still decides who moves forward.

01the problem and who owns it

A popular posting draws hundreds of applications, and the recruiter has a few minutes for each. Strong candidates with unusual titles get skipped, keyword filters reward people who copy the posting back, and the hiring manager never sees why anyone was cut.

The recruiting team owns throughput, but the risk sits with HR and legal. Any tool that influences who gets an interview is a selection procedure in the eyes of the EEOC, so the design has to be defensible before it is fast.

02what the AI does, step by step

1. Turn the job into written criteriaThe hiring manager and recruiter agree on four to eight requirements that are actually job-related, split into must-have and nice-to-have. Proxies like school names, graduation years, or employment gaps are excluded on purpose.
2. Pull the application from the ATSWhen an application arrives, the workflow fetches the resume text and screening answers. Name, photo, address, and other fields the criteria do not need are stripped before the model sees anything.
3. Map evidence to each requirementFor every criterion the model returns met, partly met, or no evidence, with the exact resume line it relied on. A requirement with no quoted evidence cannot be marked met.
4. Flag what needs a human readCareer changers, non-standard formats, and resumes the parser mangled are flagged for manual review instead of being scored low. Parsing failure is never treated as a weak candidate.
5. Write the summary back to the ATSThe criterion table and a three-line summary land as a note on the candidate record. The candidate's stage does not change; only a recruiter can advance or reject.
6. Log every run for auditInputs, criteria version, model version, and output are stored so the team can reconstruct any screening later and run adverse impact analysis across groups.

03systems it connects to

• Applicant tracking system. Greenhouse, Lever, Ashby, or Workday Recruiting, read and written through their APIs.
• Resume parsing. The ATS parser or a document extraction step for PDFs and Word files.
• Audit store. A separate log or database the recruiting team cannot edit, retained per your records policy.
• Demographic data, kept apart. Voluntary self-identification data stays outside the screening path and is used only for the bias analysis.

04human checkpoints

• Criteria approval. HR signs off on each requisition's criteria before screening starts, checking that every item is job-related and consistent with business necessity.
• Every rejection is human. A recruiter makes each advance or decline decision. The model output is one input, visible to the person who decides.
• Periodic fairness review. HR or outside counsel compares pass-through rates across groups on a schedule and pauses the tool if a gap appears.

05what to measure

• Recruiter agreement. How often recruiters advance candidates the model rated low, and the reverse.
• Time from application to first decision. Measured per requisition, before and after.
• Selection rates by group. Pass rates at the screen stage, compared using the four-fifths rule as a first signal, not a safe harbor.
• Downstream quality. Interview-to-offer rate for screened candidates, which shows whether the criteria predict anything.

06risks and guardrails

• Disparate impact. A neutral-looking rule can still screen out a protected group, and the employer, not the vendor, carries that exposure. Work with employment counsel before launch.
• Local AI hiring laws. If you hire for roles in New York City, Local Law 144 requires an independent bias audit and candidate notice before using an automated employment decision tool. Other states and cities are adding rules, so check each hiring location.
• Accommodation requests. Candidates must have a clear path to request an ADA accommodation or an alternative review, and that request must reach a person.

07build vs buy

Most modern ATS products now ship AI matching or summaries, and for a small team hiring a few roles a quarter that is usually the right call, provided the vendor will share audit results and explain how scores are produced.

A custom build makes sense when you need evidence-quoting summaries rather than opaque scores, when criteria differ sharply by role family, or when you want the audit log and bias analysis under your own control rather than a vendor's dashboard.

08related playbooks

Browse every hr and recruiting playbook or the full library.