Talkpush

Company overview

Founded in 2014 by Max Armbruster, Talkpush is a conversational recruitment CRM and AI-powered ATS built for high-volume and frontline hiring, with a strong footprint in BPO, retail and QSR across emerging markets. Its published privacy policy is governed by Hong Kong law, and the company maintains offices across the Philippines, Mexico, Costa Rica, India and San Francisco. The platform captures candidates on WhatsApp, Facebook Messenger, SMS and TikTok, routes them through Autoflow automations, scores spoken English via the TalkScore engine (CEFR A1-C2 across pronunciation, vocabulary, grammar, fluency, comprehension and active listening), and runs outbound AI phone screens through Voice-AI in 28 languages. Pricing is quote-only behind a demo request, with enterprise logos including McDonald's, Walmart and IQOR cited on the homepage.

Regulatory exposure

Talkpush sits squarely in EU AI Act Annex III high-risk territory: TalkScore assigns a proficiency grade that functions as an "objective pass/fail gate" for hiring, and Autoflow can route candidates to shortlist or rejection folders based on attributes without a documented human approval step. That creates a direct tension with the company's own privacy policy, which asserts that "no automated decisions are made by the platform" while simultaneously describing optional Smart Filter and AutoFlow functionality that "automatically processes applicants according to matching criteria determined by the Employer." The published TalkScore whitepaper is a genuine, quantitative validation study (n=200; MAE 0.84 vs. 1.18 for human evaluators; 0.783 correlation and 91.6% classification agreement against Versant) but contains no analysis by protected class, and explicitly defers demographic fairness work to "Future studies [that] should analyze scoring variations to mitigate bias" — which sits awkwardly beside marketing copy claiming the product will "remove accent bias." Accent- and language-based scoring carries well-documented national-origin discrimination exposure under US law, and there is no published NYC LL 144 bias audit, no EU AI Act or Article 27 FRIA material, no DPA, no subprocessor list, and no trust or security page (talkpush.com/security, /trust and /dpa all return 404). A SOC 2 announcement exists but names no auditor, type, date or scope.

Path to a higher score

The fastest wins are governance artifacts Talkpush does not yet have rather than new engineering. Publishing a responsible-AI or AI-governance page with a stated intended purpose, known limitations and out-of-scope uses for TalkScore and Voice-AI would immediately lift the Article 11 score, and upgrading the whitepaper's gated "comprehensive report... available upon request" into a public technical annex would go further. The single highest-leverage move is commissioning an independent adverse-impact audit of TalkScore disaggregated by sex, race/ethnicity and — given the accent-bias claim — first language and nationality, then publishing it as a downloadable report with auditor name and date; the company's own whitepaper already identifies this as the open gap. Adding a trust page that names the SOC 2 auditor, type and date, discloses whether candidate audio or transcripts are used to train or refine scoring models, and lists subprocessors and data residency would close most of the data-governance deficit. Finally, deployer-facing guidance — an Article 27 FRIA starting template, a standard DPA, NYC LL 144 notice language, and documentation of how a recruiter reviews, challenges or overrides a TalkScore result before an Autoflow rejection fires — would move FRIA support, human oversight and customer documentation together.