Talkpush

A publicly downloadable 19-page whitepaper authored by Talkpush's Head of Product Adrian Gomez and Product Manager Pablo de Yzaguirre (PDF metadata dated June 2025) documents TalkScore's evaluation process, impact-based scoring approach, CEFR threshold mapping, error typology, MAE performance metrics and stated future research directions.

https://4993999.fs1.hubspotusercontent-na1.net/hubfs/4993999/Whitepapers/Whitepaper%20Consistency%20in%20Language%20Assessment%20How%20TalkScore%20Outperforms%20Human%20Variability.pdf

A public landing page summarises TalkScore's validation headline figures (0.91 pronunciation correlation against expert human raters, 96.9% agreement identifying sub-A2 candidates) and describes the scores as 'explainable' for hiring-manager review.

https://www.talkpush.com/how-talkscore-delivers-fair-consistent-and-scalable-language-assessments

No AI policy, responsible-AI principles page, model card, explainability statement, ISO 42001 certification or EU AI Act technical documentation pack was found on talkpush.com, blog.talkpush.com or help.talkpush.com (whose 24 help collections include none covering AI governance), and the whitepaper states the detailed methodology report is only 'available upon request'.

Have a URL that disputes this?

Talkpush's own whitepaper reports a validation study across 200 candidates showing a Mean Absolute Error of 0.84 versus 1.18 for human evaluators, a 0.783 correlation with the Versant benchmark and 91.6% agreement in classification, but presents no results disaggregated by sex, race, national origin or first language.

https://4993999.fs1.hubspotusercontent-na1.net/hubfs/4993999/Whitepapers/Whitepaper%20Consistency%20in%20Language%20Assessment%20How%20TalkScore%20Outperforms%20Human%20Variability.pdf

Under 'Future Directions' the whitepaper concedes that broader sampling and cultural diversity work remains outstanding, stating that 'Future studies should analyze scoring variations to mitigate bias' and ensure 'fairness across linguistic backgrounds'- an admission that demographic bias testing has not yet been performed.

https://4993999.fs1.hubspotusercontent-na1.net/hubfs/4993999/Whitepapers/Whitepaper%20Consistency%20in%20Language%20Assessment%20How%20TalkScore%20Outperforms%20Human%20Variability.pdf

Marketing copy claims TalkScore can 'remove accent bias while maintaining CEFR alignment' and provides 'unbiased' assessment, a fairness claim that outruns the published evidence since no accent-, nationality- or first-language-disaggregated results are provided anywhere.

https://www.talkpush.com/how-talkscore-delivers-fair-consistent-and-scalable-language-assessments

No NYC Local Law 144 bias audit summary, independent audit by any recognised auditor (BABL AI, DCI, ORCAA, Warden AI, Holistic AI, Credo AI, ConductorAI), or academic third-party audit of Talkpush or TalkScore was located via web search or on talkpush.com, its whitepaper hub or its help centre.

Have a URL that disputes this?

No EU AI Act Article 27 Fundamental Rights Impact Assessment template, DPIA template, deployer-obligation guidance, high-risk-system classification note or EU AI Act content of any kind was found on talkpush.com, its whitepaper hub, its blog or its help centre, and https://www.talkpush.com/trust and https://www.talkpush.com/dpa both return HTTP 404.

Have a URL that disputes this?

The terms place compliance responsibility generically on the customer -'You shall ensure that you have all necessary approvals, permissions or authorisations for the services operated through Talkpush'- but contain no data processing agreement, no AI or automated-decision clauses and no employment-law or fundamental-rights assessment support.

https://talkpush.com/terms/

The privacy policy publishes a compliance@talkpush.com contact for data requests and privacy concerns, the only compliance-facing channel Talkpush surfaces publicly.

https://talkpush.com/privacy/

Talkpush documents a concrete retention control under which 'leads marked as soft-deleted for two years will be permanently removed from our database' (in effect since 21 October 2024), with immediate permanent deletion for candidates who unsubscribe, explicitly framed as GDPR compliance.

https://help.talkpush.com/en/articles/10056196-automatic-deletion-of-soft-deleted-leads-ensuring-compliance-with-gdpr

A blog post announces SOC 2 compliance and states that 'All data is encrypted both in transit and at rest', alongside access restrictions, routine audits and an incident response plan, but names no auditor, no Type 1/Type 2 designation, no report date, no scope and no process for customers to request the report.

https://www.talkpush.com/post/talkpush-is-now-soc-2-compliant-what-this-means-for-you

The privacy policy enumerates the personal data categories processed - name, email, address, phone, work experience and skills, captured as text, audio and video recordings - and refers generically to third-party providers for maintenance, database management, analytics and IVR hosting without naming them.

https://talkpush.com/privacy/

No disclosure of what data trains or tunes the TalkScore or Voice-AI models, no exclusion list, no statement on whether candidate audio or transcripts are reused for model improvement, no subprocessor list, no data processing agreement, no data-residency options and no ISO 27001 or ISO 42001 certification were found across talkpush.com, its privacy policy, its terms or its help centre.

Have a URL that disputes this?

Talkpush implements role-based access control across five levels (Owner, Manager, Manager with Analytics, Limited Manager, Onboarding Admin) governing campaign, analytics and settings permissions, and preserves user records by disabling rather than deleting accounts.

https://help.talkpush.com/en/articles/5844218-user-access-levels

Recruiters can attach predefined shortlist and rejection reasons to each candidate, which the documentation frames as 'ensuring transparency for all involved' and, when the single-reason option is enabled, giving each candidate 'a single, trackable rejection reason for compliance and reporting purposes'- though recording a reason is optional.

https://help.talkpush.com/en/articles/6075416-shortlisted-rejection-reasons

Autoflow maintains a version log where 'The version log shows the setup of the Autoflow that was triggered and the one currently in use', providing a partial reconstruction trail of which automation configuration acted on a candidate, though not a per-execution audit log.

https://help.talkpush.com/en/articles/6793338-faster-hiring-with-autoflow

The privacy policy asserts that 'no automated decisions are made by the platform' while in the same document describing optional Smart Filter and AutoFlow functionality that 'automatically processes applicants according to matching criteria determined by the Employer', a tension unresolved in any public document.

https://talkpush.com/privacy/

No documentation of a recruiter's ability to review, challenge or override an AI-generated TalkScore result, no human approval gate before Autoflow moves a candidate to the Rejected folder, no per-jurisdiction compliance toggles and no candidate-facing AI disclosure or consent mechanism were found in the TalkScore, Autoflow or Voice-AI help collections.

Have a URL that disputes this?

A live public status page tracks the Talkpush API, Job Landing Page and Web CRM with 30- and 90-day uptime figures and publishes dated resolved incidents with remediation notes, including a CRM 500-error incident resolved 17 July 2026 and degraded Singapore-region performance resolved 5 June 2026.

https://status.talkpush.com/

A public 'What's new' changelog publishes dated product release notes (for example 'Sort by Job Match Score' and 'Call Recordings & Insights right in Chat', both May 2025), though entries describe feature releases rather than AI model or scoring methodology changes.

https://www.talkpush.com/updates

The whitepaper claims 'TalkScore undergoes regular audits and refinements' with 'continuous model updates and benchmarking efforts', but no published artefact, cadence, dashboard or result from that monitoring process is made available.

https://4993999.fs1.hubspotusercontent-na1.net/hubfs/4993999/Whitepapers/Whitepaper%20Consistency%20in%20Language%20Assessment%20How%20TalkScore%20Outperforms%20Human%20Variability.pdf

No security.txt, vulnerability disclosure or bug bounty programme, no AI incident reporting channel, and no public fairness or model-performance monitoring dashboard were found; https://www.talkpush.com/security returns HTTP 404 and the only compliance-facing address published is compliance@talkpush.com in the privacy policy.

Have a URL that disputes this?

Talkpush operates an extensive public help centre organised into 24 collections covering Autoflow, TalkScore, Voice-AI, Analytics, Company Settings, Proctoring, Candidate Experience, New Releases and a product glossary, with no login required.

https://help.talkpush.com/en/

The public site provides detailed per-product pages for the ATS, conversational AI, assessments, recruitment automation, onboarding and voice agents, alongside a blog, podcast, events archive, integrations directory and an academy.

https://www.talkpush.com/

A public whitepaper hub offers four downloadable papers on ATS integration, social media sourcing, WhatsApp business messaging and recruiting tool effectiveness, none of which addresses AI governance, bias, fairness or regulatory compliance.

https://www.talkpush.com/whitepaper

At least one help article gives deployer-relevant compliance guidance, explaining the automatic data-deletion behaviour customers rely on for GDPR retention obligations.

https://help.talkpush.com/en/articles/10056196-automatic-deletion-of-soft-deleted-leads-ensuring-compliance-with-gdpr

No data processing agreement, subprocessor list, trust or security centre, NYC Local Law 144 notice guidance, EU AI Act deployer guidance or AI-governance FAQ exists; the help centre has no security, privacy or compliance collection among its 24, and https://www.talkpush.com/dpa, /trust and /security all return HTTP 404.

Have a URL that disputes this?