Alva Labs

Company overview

Alva Labs AB is a Stockholm-based assessment vendor founded in 2017 by Malcolm Burenstam Linder, Peter Schierenbeck and Bjorn Strom, and reports 650+ hiring teams as customers. Its product is an adaptive logic test built on Item Response Theory and Bayesian statistics, a Five-Factor Model personality test, coding tests, and structured interviews, now surrounded by candidate management, CV screening and analytics modules. The company describes assessments calibrated on roughly 1.5 million candidates with the full item bank recalibrated every 18 months. Pricing is published as a per-job-position model across 'Grow' and 'Scale' plans (with stated overage rates of EUR 370 and EUR 320 per additional position), though base plan prices are not statically listed, placing it in the mid tier rather than enterprise-only.

Regulatory exposure

As an EU-established provider of AI systems used to evaluate candidates, Alva sits squarely in Annex III high-risk territory under the EU AI Act, and it engages that exposure more directly than almost any peer: it is certified to ISO/IEC 42001 by DNV, covering its psychometric tests, candidate platform AI layer, customer platform and internal AI governance, and it explicitly states it built its AI Management System on ISO/IEC 42001 and the EU AI Act, adopting the seven trustworthy-AI values. It also holds ISO/IEC 27001 and ISO 10667-2, keeps candidate data in the EU, and states candidate data is never used to train third-party models. The clearest gaps are US-facing: no NYC Local Law 144 bias audit summary, no independent algorithmic bias audit from a named auditor (BABL AI, Warden AI, Holistic AI, DCI), and no Illinois or Colorado deployer guidance. Fairness claims such as 'minimal group differences across gender, ethnicity, and age' are asserted on product pages without published impact ratios or demographic sample sizes on the open web.

Path to a higher score

The single highest-leverage move is publishing an independent algorithmic bias audit with actual numbers — impact ratios by sex, race/ethnicity and intersectional categories, with sample sizes — as a downloadable summary rather than leaving fairness evidence inside PDF technical manuals and unquantified marketing claims. Surfacing the adverse-impact and differential-item-functioning results as an HTML page (and repeating it annually) would move the bias score materially. Second, Alva should publish EU AI Act deployer-facing material: an Article 27 FRIA template or walkthrough, an Annex IV-style instructions-for-use / system card per AI feature, and jurisdiction guidance for NYC LL 144, Illinois and Colorado. Third, make the sub-processor list publicly readable (it currently sits behind app.alvalabs.io) and add a fairness-monitoring surface alongside the existing uptime status page, so continuous bias monitoring is externally observable rather than asserted.