Skip to main content
DoneThat

AI Adoption GuideHRHire

Background-check risk extraction

Parses background-check returns, summarizes adverse data, and flags items needing recruiter review.

HR processPlanSourceSelectHireOnboardDevelopRewardExit

By Don, DoneThat’s AI coach · updated

A flag that cites the vendor report line

The job is finished when the ATS holds either a flag that quotes a specific line from the vendor return, or an empty field because the return has no line to quote. A paragraph of summary that cannot be walked back to the packet is not a quality result.

A usable flag names the search or product the vendor ran, the jurisdiction or section header if the report shows one, and the adverse wording as printed. A recruiter who opens the same Checkr or Sterling packet should land on that line without inferring which county, which year, or which possible match the extractor meant.

Empty is a correct output. When every section of the packet is clear or silent, the risk field stays blank. Writing "clear," "no issues," or "pass" is not empty. Those strings show up in later searches as if they were findings, and they cite nothing.

Extraction does not hire and does not reject. It does not auto-disqualify. It does not invent an adverse item that is not on the report. The recruiter still reads the packet.

Load the vendor return before you extract

Open the file the vendor produced, not the status chip next to the candidate name. In Greenhouse or Workday the Checkr or Sterling return is usually an attachment or an embedded report view. A badge that says complete, an email subject, or a column in an ops export is not the report.

Confirm identity first: name, date of birth, and the identifier the vendor printed, against the candidate record. A clean report on the wrong person is not a silent report. Stop and fix the packet before any flag is written.

Then walk the products in the return in the order the vendor listed them. One packet often mixes county criminal, statewide or national criminal, motor vehicle, education, employment, and watchlist sections. A flag belongs to one section. Do not roll four products into a single sentence unless every clause still cites its own line.

If the packet is missing, still pending, truncated, or locked so you cannot read the lines, do not extract. Pending is not silent. A blank page is not a clear driving record. Queue the vendor or the ATS sync and wait.

Copy the line you will cite while the report is on screen, then write the flag. Writing the flag first and hunting for a cite afterward is the usual path to an uncited flag sitting in the ATS.

Cite the report line or leave the field blank

When the vendor printed an adverse line, the flag is a cite, not a verdict.

Illustrative example: A background-check coordinator opens a Checkr packet on a Greenhouse candidate. The county criminal section for Cook County, IL shows a 2018 misdemeanor theft, disposition "supervision, terminated satisfactory 2019." The motor vehicle section lists no violations. Education verification is complete with no discrepancy. The coordinator writes one flag: "County criminal (Cook County, IL): 2018 misdemeanor theft, disposition supervision, terminated satisfactory 2019 (Checkr county criminal section)." Motor vehicle and education stay empty. The coordinator does not add "pattern of dishonesty," does not change supervision into a conviction, and does not mark the candidate ineligible. The recruiter opens the same packet, confirms the line, and follows the company's adjudication and notice steps if the role requires them.

When the vendor printed no adverse line, leave the field empty. Do not encode silence as "N/A," "none," or "green." If a section is unreadable (redaction, cut-off PDF, mismatched candidate), leave the flag blank and send the packet problem to the ops queue as a document issue, not as risk.

If two lines in the same section are adverse, cite both, still as flags for review, still without a hire decision in the text.

Recruiter review is the decision

The flag is a pointer for the person who owns the requisition. Workday and Greenhouse can route a cited flag to that recruiter. They open the vendor line, apply the role's adjudication rules and legal process, and record the outcome. The extractor must not write "do not hire," "fail," "knockout," or "ineligible."

Treating the flag as a disqualify, auto-rejecting, or blocking the offer workflow because an extractor filled a box, skips the human review the process is built on. Even a perfect cite can be a record mismatch, a matter your policy treats as non-disqualifying, or a product the role does not use as a knockout. The coordinator's output is the cite. The recruiter's output is the decision.

Where policy and applicable law require pre-adverse notice, a copy of the report, or an individualized assessment, those steps belong to the recruiter and counsel, not to the extractor. A model flag must not be the event that starts rejection mail.

When the field is empty, review is still required. Empty means the extractor found no report line to quote. It does not mean skip the packet. The recruiter still confirms the right person, the right products for the role, and that the vendor actually finished.

Do not let the flag text become a second personnel file. Keep interpretation, exceptions, and business justification in the recruiter's review notes, tied to the same cite.

Extraction mistakes that look like a complete file

A flag with no report cite. Phrases such as "criminal history present," "driving issues," or "something on the background check" cannot be audited and cannot be replayed. A second coordinator cannot open Checkr or Sterling and land on the line. Uncited flags are defects: rewrite with a quote from the packet, or blank the field. If you cannot paste the vendor's words, you do not have a flag yet.

Treating the flag as a disqualify. Moving the candidate to rejected, stopping localized offer letter generation, or freezing comp recommendation per offer solely because a flag exists, even a well-cited one, is a process error. The flag starts recruiter review. It does not close the requisition. Status changes and offer blocks wait on the recruiter's recorded decision, not on the presence of text in the risk field.

Inventing an adverse item. Inferring an impaired-driving offense from a speeding ticket, reading a dismissal as a conviction, merging two similar names, or filling a quiet education section with "degree probably false" because a school name wrapped oddly, all put findings on the record the vendor did not print. If the report is silent, empty stays empty. Do not complete a sparse education or employment section by guessing what the vendor would have said.

These errors usually come from optimizing for a filled box. Quality is the opposite: a cite you can point to, or a blank field.

After the packet: offers, clauses, and later audits

Risk extraction sits after the vendor return lands and before anyone treats the candidate as cleared to start. It does not replace the vendor's own adjudication screens, counsel's notice letters, or the recruiter's read of the packet.

Keep background-check contingencies in offer and contract language, not inside the flag. If the letter or agreement still needs a check-contingent start, send that wording through contract clause review. Run offer and compensation work after the recruiter has reviewed the packet, not on an uncited summary and not while the return is pending.

When you later inspect who was flagged, who was cleared, and who dropped out, use hiring decision bias audit on recruiter decisions and adjudication outcomes. An invented item or a flag with no cite will distort that review. A quoted vendor line plus empty-when-silent fields keeps the human decision inspectable.

Is this worth automating for you?

Whether this pays back depends on how much time it takes your team today. Most teams estimate that from memory, and the estimate is usually wrong in one direction or the other.

DoneThat reconstructs where the time actually went, with no timers to forget, so you can measure the baseline before committing to a project and check the gain afterward.

Measure the baseline first