Compliance isn't a feature
we added later.
This page is for the people who review vendors: legal, security, privacy and HR, in the US and in Europe. We built TalentShield for some of the strictest rules in hiring, GDPR and the EU AI Act. They cover most of what US teams ask about. The rest is answered below, in plain words, with every location and provider named.
- EU AI Act classification
- Treated as high-risk (Annex III, employment). We build for those obligations and don't claim an exemption.
- Automated decisions
- None. Signals and a score for a human to review. No Article 22 decision-making.
- GDPR roles
- Customer is the controller. TalentShield is the processor under a DPA with a named sub-processor list.
- Data location
- App and database in Frankfurt. Reports in Nuremberg. Backups in Helsinki. AI in Microsoft's EU Data Zone, resource and stored data in Sweden. No candidate data is stored outside the EEA.
- Model training on customer data
- Never used to train models. Microsoft may keep prompts for up to 30 days in Sweden, only to detect misuse of its service.
- Audit trail
- Signal library version, detected signals, score, badge and timestamp for every verification.
- Candidate rights
- Exercised through the customer as controller; requests reaching us are forwarded within five business days.
- Certifications
- ISO 27001 and SOC 2 not held yet; compensating controls described in the Security & Architecture Overview.
Six things that are true
on every report.
No automated decisions
TalentShield produces information for a recruiter to review. It does not reject, rank out or filter anyone on its own. Every recorded verdict is a person's.
Every signal has a reason
Recruiters see which checks passed, which raised a flag and why. If a candidate asks, you can explain the result in plain words.
Stored in the EU, under EU privacy law
Application data is stored on EU infrastructure. AI processing runs on Azure OpenAI Service in Microsoft's EU Data Zone: no training; abuse monitoring only, up to 30 days, in Sweden.
Processor, not controller
You stay the controller of candidate data. TalentShield acts as your processor under a Data Processing Agreement with a named sub-processor list.
We log what we did
Signal version, detected signals, score and timestamp for every verification. Every check is logged, so you can answer a DPIA, a candidate request or an auditor with facts.
Only what the check needs
Deterministic checks run locally. Open questions go to the model with identifying data reduced to what the specific check needs. Configurable per customer.
EU-grade privacy. Built for US hiring teams.
GDPR and the EU AI Act are some of the strictest rules in hiring, and we built TalentShield to them from the start. They ask for what US legal and security teams ask about: a person in charge, clear reasons, a record of what happened and care with personal data. So the hard questions already have answers.
- A person decides. TalentShield never rejects, ranks out or filters a candidate.
- Same input, same score. The score comes from fixed rules, so it's repeatable and explainable.
- Every signal explained. What we found, why it matters, how to check it yourself.
- Every check logged. Signal version, signals found, score and time.
- You're in control. Pick which checks run. Optional checks stay off until you switch them on.
- No biometrics. No faces, no voices, no video analysis.
- Honest about certifications. No SOC 2 or ISO 27001 yet. We show you the controls we run instead.
We don't pretend to be low-risk.
AI systems used to filter or evaluate job applications are listed in Annex III of the AI Act. Some vendors in this space describe themselves as "assistive" to stay out of that box. We do the opposite: we assume the classification applies and build for the obligations that come with it. That is a better place to be when your legal team asks, and it is the honest description of what the product does.
- Human oversight by design. Outputs are information for a person. The system has no path to reject, rank out or filter a candidate on its own.
- Transparency to the people affected. Every signal is written so it can be shown to a candidate: what we found, why it matters, how to check it.
- Logging. For every verification we record the signal library version, detected signals, score, badge and timestamp. The recruiter's review (who, what, when) is recorded separately in the decision history, as the recruiter's own.
- Documentation for the deployer. Instructions for use, the description of checks and their data flows, and the division of responsibilities between us and you are part of the customer documentation.
Obligations for Annex III systems apply from December 2027 after the 2026 amendment of the Act. We are not waiting for the date.
You are the controller. We are the processor.
Candidate data lives in your ATS. TalentShield processes it on your documented instructions, under a Data Processing Agreement that lists every sub-processor by name, location and purpose. Public professional profile information is personal data too, and we treat it that way: it is processed under your legitimate interest, only for the checks you switch on, and never bought or enriched for other purposes.
- No automated decision-making in the meaning of Article 22: the product informs a human decision, it does not make one.
- Data minimisation per check: deterministic checks run locally; open checks send the model only what that check needs; enrichment sources are configurable per customer and off by default in pilots.
- Candidate rights (access, rectification, erasure, objection, human intervention) are exercised through you as the controller; requests that reach us are forwarded to you within five business days.
- DPIA support: we provide the description of processing, data flows and safeguards your assessment needs.
Stored in the EU. Every location named.
The application and its database run in Frankfurt. Verification reports are stored in Nuremberg, with encrypted backups in Helsinki. AI processing runs on Azure OpenAI Service in Microsoft's EU Data Zone, with our resource and stored data in Sweden; never used to train models. Microsoft may keep prompts for up to 30 days in Sweden, only to detect misuse of its service. Transactional email leaves through an EU region. Encryption in transit (TLS 1.2+) and at rest (AES-256), access controls with multi-factor authentication, and separate accounts per system are the baseline, not the premium tier.
We do not hold ISO 27001 or SOC 2 yet. We say so, and the Security & Architecture Overview describes the compensating controls in detail.
Sub-processors for candidate data.
Every third party that can receive candidate data, by name. The full list with transfer mechanisms and dated prior versions is Annex 3 of the DPA.
| Provider | Purpose | Location |
|---|---|---|
| Render Services, Inc. | Application hosting, database, backups | Frankfurt, Germany |
| Hetzner Online GmbH | Storage of verification reports; encrypted backups | Nuremberg, Germany; Helsinki, Finland |
| Microsoft Ireland Operations Ltd. (Azure OpenAI Service) | Language model for document parsing and open-ended checks; no training; abuse monitoring only, up to 30 days, in Sweden | EU Data Zone (resource and stored data in Sweden) |
| Sinch AB | Telephone number type lookup; receives the number only | Sweden |
| Mailgun Technologies (Sinch Email), EU region | Transactional email to your users; never candidate data | EU |
| Apify Technologies s.r.o. | Public profile retrieval, only for customers who switch profile checks on; receives the profile URL only | Czechia |
| SerpApi LLC | Profile discovery search, only if you switch discovery on; receives name and current employer only | United States (Standard Contractual Clauses) |
Last updated 20 October 2026.
What you get for your review.
- Compliance Summary (two pages, the short version of everything below)
- Security & Architecture Overview
- Data Processing Agreement with Annexes (processing description, technical and organisational measures, sub-processors)
- AI Governance Statement
- Terms of Service, Design Partner Agreement for pilots
- Answers to standard security questionnaires, on request
- Privacy Policy
The Privacy Policy is public. The two-page Compliance Summary comes with your demo request: tick the box in the form. The DPA, Security & Architecture Overview and AI Governance Statement are sent on request, usually the same day. Write to privacy@talentshield.app.
Questions from legal,
security and DPOs.
Systems used to filter or evaluate job applications fall under Annex III of the AI Act, and we treat TalentShield as such. We do not claim an exemption. The product is built for the obligations that follow: human oversight by design, explainability of every output, logging, and documentation for the deployer.
Our customer. TalentShield processes candidate data on the customer's instructions as a processor, under a Data Processing Agreement with a named sub-processor list. We do not use candidate data for our own purposes and we do not contact candidates.
Application data and reports are stored on EU infrastructure (Frankfurt and Nuremberg, encrypted backups in Helsinki). AI processing runs on Azure OpenAI Service in Microsoft's EU Data Zone (EU and EFTA countries), with stored data in Sweden. Never used to train models. Microsoft may keep prompts for up to 30 days in Sweden, only to detect misuse of its service.
No. It produces information for a recruiter to review: a score, a badge and the reasons. There is no automatic rejection, ranking out or filtering. Every recorded verdict in the system is a person's. No text shown to recruiters or candidates is generated by a language model; every signal explanation is a fixed template written by us.
A meaningful explanation of what was checked and what was found. Every signal in a report is written to be shown to the candidate: what we found, why it matters, how to verify it.
We wanted one high bar for every customer, and EU privacy law is one of the strictest there is. Candidate data is stored in Germany, with encrypted backups in Finland. AI runs in Microsoft's EU Data Zone, with our resource and stored data in Sweden. Nothing is used to train AI models.
Not yet, and we won't pretend otherwise. We run encryption in transit and at rest, multi-factor authentication for production access and a log of every check. The Security & Architecture Overview describes every control, our hosting and AI providers can show their own certifications, and we answer security questionnaires. Write to security@talentshield.app.
No. TalentShield looks at the application only: the CV file, the application data and, if you switch it on, the public professional profile. It doesn't analyze faces, voices or interviews.
No. It shows a recruiter what it found and why. It never rejects, ranks out or filters anyone, and nothing happens in your ATS beyond a note, a score field and a tag.
No. A background check digs into a person's history: criminal records, credit, past employers. TalentShield doesn't query any of that, and it doesn't verify identity documents or collect biometrics. It checks whether the application you received holds together: the file, the text, the contact details, the timeline. Keep your background check at the offer stage; TalentShield runs earlier, on every application.
TalentShield can't advance or reject anyone. It gives each application signals, a score and a badge, built from fixed rules, and it's designed to be one input a recruiter reviews next to everything else. Whether Local Law 144 applies depends on your roles and on how your team uses the output; that's a call for you and your counsel. If you need a bias audit or documentation for one, we give your auditor what each signal checks, how the score is calculated and the logs.
They're new and still changing, and most ask employers for the same things: tell candidates when AI is used, keep a person in charge, watch for unfair outcomes, keep records. TalentShield is built around those ideas. A person makes every decision, every signal is written in plain words you can share with a candidate, and every check is logged. We give you a plain-language description of what TalentShield checks for your candidate notice. Which law applies to you is a question for your counsel.
TalentShield checks whether an application holds together, not who the candidate is. No signal is defined by race, gender, age, religion, disability, national origin or any other protected characteristic, and we don't infer them. No photos, video, voice or biometrics. One signal is never enough on its own, and a person always reviews the reasons. We recommend running the same checks for every candidate in a role.
We process candidate data only on your instructions and never for our own purposes. We don't sell or share it. Our Data Processing Agreement sets this out.
That's a call for you and your counsel. Many employers mention application and identity checks in their candidate privacy notice. We give you a plain-language description of what TalentShield checks, ready to paste.
Security & Architecture Overview, Data Processing Agreement with sub-processor list, AI Governance Statement, Privacy Policy, and our answers to standard security questionnaires. Start with the Compliance Summary: tick the box in the demo form.
Security questions and vulnerability reports: security@talentshield.app (acknowledged within 2 business days). Data protection, DPA and candidate requests: privacy@talentshield.app. Page last updated 20 October 2026.
See it on your own pipeline.
Leave your email and the ATS you use. We come back within one business day to book a 30-minute demo, on your own applicants if you want. Just have questions? Talk to Mateusz, 30 minutes, no slides.