Cangrade

Company overview

Cangrade was founded in 2014 by Gershon Goren and is headquartered at 9 Munroe Avenue, Watertown, Massachusetts, in the greater Boston area. Its core product is a psychometric pre-hire assessment measuring around 50 traits, scored by machine-learning success models built either from an employer's performance KPIs, from internal subject-matter-expert input, or generatively from a job description, producing a fit score that predicts performance and retention. The platform has since expanded into AI resume screening and candidate matching, structured interview guides, video interviewing, interview scheduling, AI reference checking, and onboarding and retention forecasting. Named customers include Wayfair, FDNY, and Lamar Advertising. The company is privately held with a small team, does not publish pricing, and sells on custom subscription quotes.

Regulatory exposure

Cangrade's product is squarely in scope of the major AI-hiring regimes: it is an AEDT under NYC Local Law 144, a covered AI system under the amended Illinois Human Rights Act and (for its video interviewing module) the Illinois AI Video Interview Act, an ADMT under Colorado's replacement legislation, and would be Annex III high-risk employment AI under the EU AI Act if deployed in the EU. Against that exposure, the company markets unusually categorical claims — "zero statistically significant adverse impact on any legally protected group" and a "0% chance of introducing bias" — that are not matched by published evidence. Cangrade publishes no independent bias audit, no adverse-impact study with sample sizes, dates, or demographic breakdowns, and no Local Law 144 bias-audit summary; it does not appear in the ACLU's crowd-sourced LL 144 audit tracker, and its Ethical AI page names no external auditor. The single public artifact offered in support is US Patent 11,429,859 ("Systems and processes for bias removal in a predictive performance model", granted August 2022), which describes a method for iteratively dropping bias-contributing variables — evidence of technique, not of outcome. Compliance documentation is explicitly gated behind the account team. The EU AI Act is absent from the site entirely, including from Cangrade's own state-by-state AI hiring regulation guide, and no FRIA or Article 26 deployer material exists. Security posture is comparatively better documented: SOC 2 Type 2, US-only data residency on AWS, a commitment that candidate data is never used to train models outside the customer's organization, a 24-hour breach-notification SLA, and named subprocessors in the privacy policy — though no ISO 27001, no ISO 42001, no published DPA, and no trust center.

Path to a higher score

The highest-value single move is to commission and publicly post a named independent bias audit in Local Law 144 format — auditor, date, data source, and intersectional sex-by-race/ethnicity selection-rate and impact ratios — since the current categorical "bias-free" marketing rests entirely on ungated internal testing and a method patent. Alongside it, release the underlying adverse-impact study as a downloadable document with sample sizes and test dates. Second, publish an Article 11-grade technical pack: a model or system card covering intended purpose, training-data provenance and demographic composition, validity coefficients, known limitations, and instructions-for-use, ideally anchored to NIST AI RMF or ISO/IEC 42001. Third, close the EU gap with an Article 27 FRIA template and Article 26 deployer guidance, a downloadable DPA with a subprocessor list, and a public trust center replacing the account-team gate. Finally, stand up post-market surfaces: a working public status page, a dedicated security-disclosure contact and vulnerability policy, a model-update changelog, and periodic re-published adverse-impact monitoring results so the bias claim is continuously evidenced rather than asserted once.