When Software Sorts the Résumés, Who Owns the Rejection?
A three-person HR team posting a warehouse supervisor role might receive two hundred applications in a week. Fifteen years ago that meant evenings with a highlighter. Today it usually means a subscription: a résumé parser, a chat-based pre-screen, a scheduling bot, sometimes a video interview platform that scores responses on word choice and tone.
These tools are not inherently risky. Unmanaged, they concentrate risk in a place small employers rarely look — the point where a qualified applicant quietly disappears from the pipeline and no one can later explain why.
Where the software actually touches the funnel
It helps to map the points of contact before reviewing any vendor. Screening tools typically intervene in four places:
- Sourcing and ranking. A platform scores incoming résumés against a job description and orders the list.
- Pre-screening. A chatbot or asynchronous assessment asks about availability, certifications, or work authorization, and may auto-decline candidates who answer in certain ways.
- Scheduling and communication. Automation sends rejections, reminders, and interview invitations without a person reviewing each message.
- Evaluation. Video or game-based assessments generate a numeric score that a recruiter may treat as a ranking rather than a data point.
Each of these is an employment decision point, and each produces a record. That is the practical lens: not whether the tool is intelligent, but whether the employer can reconstruct what happened to a given applicant.
The record you may not know you own
Federal retention rules do not disappear because a vendor holds the data. Hiring records generally must be preserved for one year under Title VII and the ADA, and for two years under the Uniform Guidelines for larger employers and under OFCCP rules for federal contractors. Several state human rights agencies set their own periods, and some require disclosure of automated decision tools to candidates.
Ask the vendor three questions in writing: Can we export the full applicant record, including scores and dispositions, in a usable format? How long do you retain it, and does deletion happen on our instruction or yours? If we cancel the contract, what do we receive and when?
If the answer to the first question is a dashboard screenshot, the employer does not have a record. It has a view.
Bias audits and the assurance letter
A vendor's marketing page often references an independent bias audit. That phrase deserves scrutiny. Some jurisdictions, notably New York City, require an annual audit of automated employment decision tools with published results; other states have moved in parallel directions through their civil rights statutes, and additional frameworks are taking effect in 2026. Few of these obligations transfer to the vendor automatically.
The practical questions are narrower than the legal ones. Which protected groups were tested? What was the sample, and was it drawn from a population resembling your applicant pool? What disparity threshold triggers remediation? Who receives the results, and will the vendor notify you if the model changes materially after the audit?
A short internal adverse impact review — comparing selection rates across groups at each stage where the tool screens candidates out — remains a reasonable safeguard even where no statute compels it. It is also the fastest way to notice that a pre-screen question is filtering out a category of applicant no one intended to exclude.
What human review should look like
"Human in the loop" is the most overused phrase in this category. Oversight that consists of a recruiter clicking Approve on a queue of automated rejections is not review; it is ratification.
More defensible practice includes four elements. A named person, not a role description, is accountable for each requisition. That person reviews the full application — not just the score — for every candidate the tool declines before any rejection is sent. The reviewer records a brief reason. And a periodic sample of candidates the tool advanced is reviewed as well, to check whether the ranking is surfacing people the job description would not have predicted.
For employers using AI to draft interview questions or summarize transcripts, add one more rule: no summary is treated as the record. The underlying notes or transcript stay attached.
Five items to settle before the tool goes live
- A written purpose statement. One paragraph describing what the tool decides, what it merely informs, and which decisions remain human.
- A data map. What applicant data enters the system, where it is stored, whether it trains the vendor's models, and how it is deleted.
- Contract terms on audit rights, indemnification, and model change notice. A clause allowing the vendor to alter scoring logic without notice undermines every other control.
- A candidate notice. Plain-language disclosure that automated tools are used, what they assess, and how to request an alternative process or a human review. Several jurisdictions require a version of this; it is good practice everywhere.
- A quarterly file check. Pull ten recent requisitions and confirm each applicant record can be reconstructed end to end, including who made each disposition and when.
Keeping the decision in-house
The appeal of these platforms is speed, and speed is a legitimate business need for a five-person recruiting function. But the screening tool produces a recommendation; the employer produces the decision. That distinction only holds if someone can explain, months later, why a particular applicant was set aside.
Adopting a tool is easy. Adopting the recordkeeping, notice, and review habits that make the tool defensible takes slightly longer — roughly the time of two hiring cycles. That is still a shorter timeline than answering an agency inquiry without the file.