Checkr

Company overview

Checkr, founded in 2014 by Daniel Yanisse and Jonathan Perichon and headquartered in San Francisco, is a consumer reporting agency and background screening platform serving gig, staffing and enterprise employers. Its products span criminal, employment, education, MVR/DOT, drug and international screening, wrapped in an API-first platform with pre-built integrations into most major ATS vendors. The AI layer is unusually well-documented at the component level: a proprietary charge classifier that categorizes criminal charges across fragmented US jurisdictions, a charge explainer that renders severity and statutes in plain language, name and role matchers, a document inspector for forgery signals, and an ML turnaround-time predictor. Its Assess product layers customer-defined rulesets (up to 235 filters at Premium tier) on top of the classified records, reducing manual review by up to 85% while returning advisory eligible/review/escalated values rather than final decisions. Pricing is usage-based and publicly listed at $29.99–$94.99 per report with custom enterprise tiers.

Regulatory exposure

Checkr sits in an unusual regulatory position: it is primarily governed by the FCRA as a consumer reporting agency, not as a classic automated employment decision tool, and NYC LL 144 places the compliance duty on the employer rather than the CRA. But the honest reading is that its ML materially influences employment decisions — the charge classifier determines how a record is characterized, and Assess rulesets convert that characterization into an eligibility signal, which is squarely within the kind of candidate evaluation the EU AI Act treats as high-risk under Annex III. Checkr's strongest compliance asset is ISO/IEC 42001:2023 certification, audited by A-LIGN and announced in October 2025, making it the first background screener certified to the AI management system standard; that maps closely to EU AI Act Article 17 quality-management expectations. Against that, there is no published bias audit of any kind, no demographic impact-ratio analysis, no FRIA or deployer guidance, and no mention anywhere on Checkr's site of the EU AI Act, LL 144, Colorado SB 205 or Illinois HB 3773. The Assess page's claim that 'the risk of bias is drastically reduced by 90%' is a marketing figure with no published methodology behind it. Accuracy risk is also live rather than theoretical: Davis v. Checkr Inc. (S.D. Fla., 0:26-cv-60088, filed January 2026) alleges Checkr failed to maintain reasonable procedures to assure maximum possible accuracy and mismatched criminal records to the wrong consumer — precisely the failure mode an identity-matching model creates.

Path to a higher score

The single highest-value move is publishing an independent demographic bias audit of the charge classifier and Assess eligibility outputs — impact ratios by race/ethnicity and sex, ideally intersectional, from a named auditor with a date. Because Checkr's data flows disproportionately affect populations with criminal records, this would be the most consequential fairness disclosure any screening vendor could make, and the 90% bias-reduction claim needs methodology behind it or should be retired. Second, publish an Article 11-style technical pack or model cards for the named components (charge classifier, name matcher, document inspector) with accuracy metrics, known limitations and demographic performance breakdowns; the ISO 42001 certificate proves a process exists but discloses none of its contents. Third, produce explicit deployer guidance — a FRIA template, an EU AI Act high-risk role allocation statement, and per-jurisdiction notes for LL 144, Colorado and Illinois HB 3773 disclosure duties — since Checkr's customers now carry those obligations and get no help with them today. Fourth, add an AI-specific surface to post-market monitoring: a model-update changelog and published drift or fairness monitoring alongside the existing operational status page. Finally, make the ISO 42001 certificate and its statement of applicability publicly downloadable rather than badge-only.