Not a legal opinion
US hiring is covered by anti-discrimination law, and tools that influence who is considered can fall inside that. Rules differ by role, state, and how the tool is used. Have counsel review a real workflow before you deploy one. Nothing on this page is legal advice or a compliance certification.
What a careful first version does
It reads a résumé or an application and lays out fields a recruiter already looks for: recent titles, dates, skills the person wrote, and location if it is present. A recruiter compares those fields to the role. The tool does not mark the person as hire or reject, and it does not email the candidate.
What enters is the file or the form. What is prepared is a structured summary. The approval point is any status change in the applicant tracking system. Until a recruiter makes that change, the candidate’s status stays as it was.
Bias is not removed by a model
A score trained on “who we hired before” will tend to repeat who you hired before. That is not a neutral reading of a résumé. Do not present a historical hiring score as an objective rank.
Proxies matter. A postal code, a gap explained poorly by the parser, or a university name can stand in for a protected trait even when the form never asks. If you do not have a reason to extract a field, leave it out.
Privacy and the candidate
Say what you collect, how long you keep the file and the summary, and who inside the firm can open them. Candidates did not agree to an unbounded experiment with their work history. A vendor’s default retention is not your policy until you choose it.
Parsing fails on nonstandard layouts, career breaks, and work described outside a job title. Those misses are a reason for a person to read the original, not a reason to treat the empty field as a lack of experience.
Related service: Agentic AI and automation