Shadow AI in HR: Using AI Without Breaking the GDPR (2026)

If you run an HR or talent team in 2026, your biggest AI risk is not the AI vendor your procurement has spent six months vetting. It is the senior recruiter, the L&D manager, and the comp analyst who are pasting candidate CVs, performance reviews, and salary tables into a free ChatGPT account on their personal laptop at 11pm because the official tool is slower. This is shadow AI in HR, and it sits at the intersection of the two regulations that have the highest enforcement velocity in the EU right now: the GDPR and the AI Act.

This is the second article in our redaction guide series. The pillar guide covers data redaction across all use cases; this one zooms into HR specifically — recruiting, performance management, employee relations, comp and L&D — where the shadow AI problem is most acute, the regulatory exposure is highest, and the operational fix has the largest single-team payoff. Terminology note: we use "shadow AI HR" and "shadow AI recruiting" interchangeably, and we use "AI resume screening", "AI talent acquisition", "HR generative AI", and "AI chatbot HR" as the operationally distinct sub-categories that the same architecture has to cover. "AI performance management" and "AI performance review" describe overlapping but distinct workflows, both equally affected.

If your firm has already had the "what do we do about ChatGPT" conversation and quietly lost it, this is the article you can hand to the people who actually use AI every day. It covers where shadow AI is happening inside HR (seven concrete use cases), why those usages break the GDPR and AI Act in specific ways, the real enforcement record from 2023-2026 (Workday, EEOC, NYC AEDT, the Italian Garante on Foodinho/Glovo, the European Court of Human Rights), what a sanctioned alternative needs to look like to actually be adopted, and a ten-step rollout that turns the conversation into a sanctioned tooling decision in 90 days.

1. Why HR is ground zero for shadow AI

Shadow AI is the use of consumer-tier AI tools — ChatGPT, Claude, Gemini, Perplexity, Mistral Le Chat, the dozens of GPT wrappers in the App Store — with confidential data, outside IT visibility, typically through personal accounts. Every function in a firm has some exposure to it; HR has the worst. There are four structural reasons.

First, the data HR handles is mostly free text. CVs, cover letters, performance review comments, employee complaints, exit interviews, manager 1:1 notes, calibration discussions — none of these are tabular records the way customer data tends to be. Free text is exactly where generative AI feels most useful, so the temptation to paste it into a chat window is highest. The pillar covered PII redaction as a technical problem; in HR it is also a behavioral one.

Second, HR work is bursty and judgment-heavy. A recruiter screening 200 CVs the week before a hiring deadline has no time to ask Legal whether the new sourcing tool is GDPR-compliant. A people manager writing a difficult performance review at 9pm wants a thinking partner, not a procurement workflow. The official tools, when they exist, were procured for HRIS or ATS use cases — not for this. So ChatGPT HR use spreads.

Third, HR data is heavily regulated as a category. Article 88 of the GDPR explicitly authorises Member States to set more specific rules for processing employee data, and most have. The EU AI Act classifies HR uses as high-risk in Annex III, paragraph 4. Several non-EU jurisdictions have written specific employment-AI laws: New York City Local Law 144 (Automated Employment Decision Tools, AEDT, in force July 2023), Illinois AI Video Interview Act (2020), Colorado AI Act (effective 2026). The compliance surface in HR is genuinely larger than in other functions, and shadow AI use creates undocumented processing across all of it.

Fourth, HR sits between the legal/compliance bar and the operational pressure to be fast. The same HR director who is told to drive AI productivity gains is also responsible for the GDPR records of processing activities (Art. 30), the Article 35 data protection impact assessment for new technologies, the Article 22 right not to be subject to automated decisions, and — in many EU countries — the works council consultation on new monitoring tools. Without a sanctioned alternative that solves both problems at once, the system bends in the direction of shadow.

The statistics quoted about shadow AI in HR are often repeated with a different scope from the one the source used, and where each shadow AI number comes from lists the population, method and date behind the most cited ones.

2. Where shadow AI shows up in HR teams

The seven use cases below cover roughly 95% of the shadow AI we see in HR audits. Each one creates a different category of risk and a different opportunity for a sanctioned alternative.

1. CV screening and shortlisting (AI resume screening). A recruiter pastes 50-200 CVs into a chat session and asks "rank these for cultural fit and technical depth on Kubernetes; give me a top 10". The model sees full names, photos, addresses, dates of birth, previous employers, and any other PII the candidate disclosed. Under Art. 22 GDPR a fully automated decision with legal or similarly significant effect requires an explicit lawful basis; the EU AI Act classes CV screening as high-risk; AI recruiting bias — also referenced as AI bias hiring or AI hiring bias depending on the literature — has its own discrimination exposure under national employment law. The shadow version creates undocumented processing of special-category data (some CVs disclose health information, union membership, religious affiliation) on a US LLM endpoint with no DPA. This is the canonical AI talent acquisition shadow workflow.

2. Performance review writing and AI performance management. A manager pastes 12 months of 1:1 notes and asks the model to draft the year-end review. The notes typically include direct quotes from the employee, observed health information ("she was out sick three weeks in Q3"), comp discussions, and sometimes complaints about peers. AI performance review drafting on a consumer endpoint creates a Recital 71 problem (decisions producing legal or similarly significant effects), an Art. 88 problem (employment-context processing), and frequently an Art. 9 problem (special category disclosure). AI performance review bias is a well-documented failure mode: the model amplifies whatever signal is strongest in the underlying notes, and that signal is often the manager's pre-existing impression of the employee, not the employee's actual behaviour.

3. Exit interview synthesis. An HR business partner conducts exit interviews, then pastes the transcripts into a chatbot to summarise themes for leadership. Exit transcripts contain salary specifics, accusations of harassment, named complaints about managers, and sometimes whistleblowing content protected under the EU Whistleblowing Directive (2019/1937). The shadow version exposes both the leaver and the named third parties on a public LLM. AI exit interview handling is also one of the places where employees most expect confidentiality — a breach is reputationally severe.

4. Job description and ad writing. A talent marketer feeds the previous job ad, the hiring manager's brief, and competitor postings into a chatbot to write a new AI job description. PII risk is low here, but discrimination risk is significant: LLMs trained on historical hiring data carry historical gender and age coding that can produce ads that systematically discourage protected classes (the Amazon résumé tool that downgraded "women's chess club" in 2018 is the textbook case). The EEOC's 2023 technical assistance document confirms ads can violate Title VII even without intent.

5. Compensation benchmarking. A comp analyst pastes the firm's salary band table and asks the model to compare against published market data, propose a new structure, or flag outliers. The bands themselves are confidential, the employee-linked rows are personal data, and the resulting model output is now in a US LLM with potential training use. AI compensation benchmarking via shadow AI is also a competitive-information leak, not just a privacy one.

6. Employee Q&A and policy assistance (AI chatbot HR). Employees ask an HR chatbot questions like "I'm on parental leave; what's my entitlement?" — and the answer requires the employee's contract, country, role, and tenure. Many firms have stood up a generative AI HR assistant on a consumer endpoint to handle this without realising the employee-side input contains health information ("I'm on sick leave because…"), union activity ("I'm on union duty leave"), or family details. ChatGPT employee data exposure through these assistants is one of the harder problems to detect because the data leaves through the employee's input rather than the employer's. The Italian Garante has signalled that employee-facing AI assistants are subject to the same Art. 88 employment rules as employer-side processing.

7. Algorithmic management and workforce planning. A workforce planning analyst feeds an LLM the org chart with names, levels, dates of hire, and performance ratings, and asks for restructuring scenarios or a "list employees most at risk of leaving in the next 6 months". This is algorithmic management in the Foodinho/Glovo sense, and it has produced two of the largest employment-AI enforcement decisions in the EU: the Italian Garante's €2.6M fine on Foodinho (2021) and the subsequent Spanish AEPD work on platform workers.

The seven use cases share a pattern. The user knows the data is sensitive but feels protected by the fact that the chat window "is just me typing". The model sees the data as input; the provider's logs see it as a request; the training pipeline may see it as a training example; and your firm's records of processing see none of it.

3. The GDPR + AI Act + national stack for HR AI

Four legal layers apply simultaneously to AI use in HR. They are not redundant; each one catches a different failure mode.

GDPR Article 88 (Processing in the context of employment). Member States may provide for more specific rules on the processing of employees' personal data, including for the purposes of recruitment, performance of the employment contract, planning and organisation of work, equality and diversity, health and safety, and termination. Italy (D.Lgs. 196/2003 as amended), France (Code du travail L1121-1 and L1222-4), Germany (BDSG §26), Spain (Organic Law 3/2018), and Belgium have all exercised this prerogative. The practical effect is that EU-wide compliance does not exist — you have to comply with the GDPR and the local employment-data rules.

GDPR Articles 5, 9, and 22 — the substantive backbone for employment AI. Art. 5 sets the principles (lawfulness, fairness, transparency, purpose limitation, data minimisation, accuracy, storage limitation, integrity, accountability). Art. 9 prohibits processing of special categories without an Art. 9(2) basis, which in employment is most often the Art. 9(2)(b) basis "for the purposes of carrying out obligations under employment law" — but only "in so far as authorised by Union or Member State law". Art. 22 grants the right not to be subject to a decision based solely on automated processing, including profiling, which produces legal effects or similarly significantly affects them. The EDPB's Guidelines 04/2024 on Art. 22 clarify that "based solely" is a meaningful filter — token human review at the end of a process does not avoid Art. 22 — and HR shortlisting is the canonical example. GDPR employee data processing under any of these articles requires lawful basis, transparency, and a documented assessment.

EU AI Act Annex III paragraph 4 — high-risk HR uses. The AI Act explicitly lists HR uses as high-risk: (a) AI systems intended to be used for the recruitment or selection of natural persons, in particular to place targeted job advertisements, to analyse and filter job applications, and to evaluate candidates; (b) AI systems intended to be used to make decisions affecting terms of work-related relationships, the promotion or termination of work-related contractual relationships, to allocate tasks based on individual behaviour or personal traits or characteristics, or to monitor and evaluate the performance and behaviour of persons in such relationships. The high-risk classification triggers the full Title III obligations: risk management system (Art. 9), data governance (Art. 10), technical documentation (Art. 11), record-keeping (Art. 12), transparency (Art. 13), human oversight (Art. 14), accuracy and robustness (Art. 15), conformity assessment, registration in the EU database, post-market monitoring, and a fundamental rights impact assessment (Art. 27 for public bodies and certain private deployers). EU AI Act HR obligations are the most demanding compliance regime any HR function has ever had to operate under. The AI high-risk HR classification under Annex III is what triggers all of this; the AI impact assessment HR teams now have to produce — both the GDPR DPIA and the AI Act FRIA — is the new operational documentation burden.

National employment-AI laws. Beyond Article 88 GDPR implementations, several jurisdictions have written AI-specific employment statutes. NYC Local Law 144 (AEDT) requires bias audits and candidate notice for any "automated employment decision tool" used to "substantially assist" hiring or promotion in NYC since July 2023. Illinois AI Video Interview Act (2020) requires consent, AI-use notice, and limits on data sharing for video interview AI. Colorado AI Act (effective 1 February 2026) mirrors much of the EU AI Act's structure for high-risk uses, including employment. Spain has codified algorithmic transparency for platform work (Real Decreto-Ley 9/2021). The Italian Garante has issued specific guidance on biometric attendance, geolocation, and algorithmic management. France's CNIL has issued recommendations on AI in HR specifically (2024).

The practical takeaway: any HR AI deployment in 2026 has to pass simultaneously through GDPR, AI Act, national employment-data law, and (where applicable) an AI-specific employment statute. Shadow AI use bypasses every one of them.

4. Seven specific ways HR shadow AI breaks compliance

If you are presenting this to a board or a works council and you need a concrete list of what shadow AI actually breaks, here is the operational version.

1. Article 28 — processor without DPA. A consumer ChatGPT account is not a Data Processing Agreement. A senior recruiter pasting CVs is creating processing on behalf of the firm without the Art. 28(3) clauses, without sub-processor disclosure, and without a contractual basis to act on the firm's behalf.
2. Article 30 — invisible processing activity. Records of processing activities for HR generally cover the ATS, the HRIS, the payroll vendor, and the L&D platform. They do not cover the recruiter's personal ChatGPT use. The firm's Art. 30 record is therefore materially incomplete, which is itself a violation independent of any data leakage.
3. Article 9 — undocumented special-category processing. CVs and performance discussions routinely contain Art. 9 data (health, union membership, religion, sexual orientation). Processing it on a consumer LLM creates Art. 9 processing without an Art. 9(2) basis and without the additional safeguards.
4. Article 22 — solely automated decision. If shadow AI is being used to rank candidates and the recruiter implements the ranking, the human review is a stamp; the decision is "based solely on automated processing" within the meaning of Art. 22 and Recital 71. The Court of Justice's SCHUFA ruling (C-634/21, December 2023) confirmed a broad reading of "based solely".
5. Articles 5 and 35 — no impact assessment. A high-risk processing operation requires a DPIA under Art. 35, and the AI Act adds a fundamental rights impact assessment for many HR deployments. Shadow AI use never has one.
6. Article 88 and national employment law. Most EU jurisdictions require some form of consultation with the works council (Italy art. 4 Statuto dei Lavoratori; Germany Betriebsverfassungsgesetz §87; France Code du travail L2312-38) before introducing monitoring or evaluation technology. Shadow AI is monitoring technology introduced without consultation.
7. Cross-border transfers and Schrems II. Consumer ChatGPT, Claude, and Gemini accounts route data to US compute by default. There is no transfer impact assessment, no Standard Contractual Clauses, and the firm is unaware the transfer is happening. This is the easiest single GDPR violation to demonstrate in an audit because the network logs alone are sufficient evidence.

The seven breaks compound. A single CV-screening shadow session can trigger six of the seven simultaneously. The Spanish AEPD has fined for less.

5. Real enforcement and the price of doing nothing

HR teams sometimes dismiss the legal exposure as theoretical. The 2023-2026 record disagrees. The following are the cases that should be in every HR leader's awareness in 2026.

Workday (USA, 2024-2025). The Mobley v. Workday federal class action alleges Workday's AI screening algorithm discriminates against applicants based on race, age, and disability. The court denied Workday's motion to dismiss in July 2024 and certified the collective action in May 2025, accepting the theory that an AI screening tool can be treated as the employer's agent for Title VII purposes. This is the first case to establish that the AI vendor itself can be held liable for discriminatory outcomes, expanding the liability surface beyond the employer.

EEOC technical assistance and settlement record (USA, 2023-2026). The EEOC's May 2023 technical assistance document on AI in employment selection set out the framework for applying Title VII (disparate impact) to AI tools. In EEOC v. iTutorGroup (August 2023, settled for USD 365,000) the EEOC alleged the firm's recruitment software programmatically rejected older female applicants. The EEOC has continued to add AI-discrimination cases to its docket through 2026; the May 2023 strategic enforcement plan named AI-driven discrimination as a priority.

NYC AEDT Law 144 enforcement (USA, 2023-2026). Since 5 July 2023, employers using an "automated employment decision tool" for hiring or promotion in NYC must conduct an annual independent bias audit, publish the summary results, and notify candidates. Enforcement has been measured but real, and the law's existence has materially reshaped vendor behaviour in the US recruiting market. Several firms have withdrawn AI features from the NYC market rather than comply.

Italian Garante on Foodinho / Glovo (Italy, 2021-2024). The Garante fined Foodinho (Glovo Italy) EUR 2.6 million in 2021 for the way its rider-rating algorithm processed personal data without proper Art. 88 basis and without adequate transparency, and again EUR 5 million in 2024 on related matters. The decision is foundational for algorithmic management compliance in Italy and has been cited extensively in subsequent decisions across the EU.

Italian Garante on biometric attendance (Italy, 2022-2025). Successive decisions have established that biometric attendance systems and facial-recognition entry systems in workplaces require a specific legal basis beyond consent (which is invalid in the employment context due to the imbalance of power) and have struck down several deployments. The pattern applies directly to any AI tool that processes biometric data of employees.

CNIL guidance on AI in HR (France, 2024-2025). The CNIL published guidance in 2024 confirming that any HR AI deployment must be preceded by a DPIA, must operate on a lawful basis other than consent, must include human oversight, and must respect the works council consultation requirement. The CNIL's regulatory action against algorithmic management tools in the French public sector has set baselines that private employers are expected to meet.

European Court of Human Rights on workplace monitoring. The Barbulescu v. Romania Grand Chamber judgment (2017) and the more recent López Ribalda v. Spain (2019) set the European baseline for workplace monitoring: the employer must give prior, clear, transparent notice and the measures must be proportionate. AI-based behavioural monitoring inherits this baseline, and shadow AI use of employee performance data falls outside it almost by definition.

OpenAI / Italian Garante (2023-2024). The Garante's 30 March 2023 temporary suspension of ChatGPT, the May 2024 EDPB ChatGPT Task Force preliminary findings, and the December 2024 EUR 15 million fine on OpenAI itself collectively confirm that LLM-side processing is now an actively enforced compliance domain. HR shadow AI inherits all of those risks vicariously.

The compound effect is that an HR director cannot any longer assume "no enforcement happened to us yet". The record is dense enough that the question is when, not whether.

6. What a sanctioned, faster-than-shadow HR AI looks like

The pillar guide laid out the general architecture for safe LLM use: mask PII in the browser before it reaches the model, route through an EU-hosted LLM under a signed DPA, restore real values locally. The HR-specific version of that architecture has four additional requirements driven by the regulatory stack we covered above.

(a) Role-aware redaction. A CV-screening workflow needs to redact the candidate's name, photo, date of birth, address, and any Art. 9 disclosure — but it must not redact qualifications, experience, education, or skills. A performance review draft needs to redact the employee's name and any third-party named in the notes — but must preserve the underlying behaviours. A generic redactor that masks everything destroys the workflow's value. The sanctioned tool must understand HR entity categories at the granularity the work actually needs.

(b) Documented sub-processor chain. Every LLM the tool uses must be in the firm's sub-processor list with the corresponding DPA. Redactic publishes its sub-processor list and changes are pre-notified 30 days in advance. For HR use specifically the firm's DPO needs to be able to demonstrate, at any point, exactly which LLM saw what category of data on which date.

(c) Article 22 hygiene by default. The tool must structurally prevent "based solely on automated processing" outcomes. In practice this means the output is always advisory, every individual decision has a recorded human reviewer with a documented review step, and the bias-audit log captures the human override rate. AI performance review drafting under this discipline produces a draft that the manager edits substantively; the redaction layer makes the draft input-safe; the human review layer makes the output Art. 22-safe.

(d) Works council and transparency artefacts. The tool must be able to produce, on request, (i) the Art. 30 record entry for the processing, (ii) the DPIA template, (iii) the fundamental rights impact assessment under AI Act Art. 27, (iv) the bias-audit report aligned to NYC AEDT format where applicable, and (v) the works council notification draft. The redaction certificate that Redactic generates for each job becomes the evidence base for these documents, since it carries hashes and counts but no PII.

Redactic Secure Chat implements all four. The Chat interface masks PII in the browser via the same multilingual GLiNER pipeline as the document workflow, sends the pseudonymised text to OpenAI / Anthropic / Google / Mistral on their EU endpoints (zero retention, no training on input under signed DPA), restores the real values client-side. The audit certificate records the job. For document workflows — JD writing, performance review drafting, comp analysis, exit interview synthesis — the upload-then-redact-then-AI flow follows the same architecture with the redaction layer in front of the model. The Italian, French, German, Spanish, and English landing pages of redactic.eu cover the use cases directly.

7. The 90-day rollout for HR leaders

This is the ten-step playbook we use with HR leaders adopting sanctioned AI for the first time. It assumes you have shadow AI in your organisation today (you almost certainly do) and you want to move to a sanctioned alternative in a quarter.

1. Day 0-7. Internal census of shadow use. Survey HR, L&D, and recruiting on which AI tools they actually use, for which tasks, with what data. Promise amnesty for honest answers — you need the truth, not the official answer. Expect ChatGPT in 70%+ of responses, Gemini and Claude in 30-50%, and a long tail of specialised tools.
2. Day 7-14. Map the use cases to the seven categories above. Mark each as high, medium, or low data sensitivity. Mark each as high-risk under AI Act Annex III or not. This is your initial inventory of processing operations.
3. Day 14-30. Compliance baseline. Run a gap analysis against the seven breaks in section 4. The DPO and legal sign the gap analysis. This becomes the business case for the sanctioned tool — and the evidence trail if the regulator arrives in the meantime.
4. Day 30-45. Vendor selection. Apply the 8-point checklist from the pillar guide with HR-specific weighting on (b) and (d) above (role-aware redaction + works council artefacts). Procurement signs the DPA, sub-processor list, and SCC where applicable.
5. Day 45-60. DPIA and FRIA. The DPIA covers the GDPR side and the FRIA covers the AI Act side; both are required for high-risk HR uses. The vendor's redaction certificate, bias-audit template, and sub-processor list feed directly into both documents.
6. Day 60-70. Works council and employee transparency. Notify the works council under the applicable national law, file the consultation, draft the employee-facing notice (Art. 13/14 GDPR plus AI Act Art. 26 transparency). Publish the notice.
7. Day 70-80. Pilot with one team. Pick the team with the highest shadow usage — usually senior recruiters or HR business partners — and onboard them to the sanctioned tool. Measure adoption against the baseline shadow usage. The single test of success is whether shadow use drops by week 2.
8. Day 80-85. Adjust the friction profile. If the sanctioned tool is slower than the shadow path, fix the friction (SSO, browser extension, default models, latency). The architecture wins only if it is faster, not just safer.
9. Day 85-90. Wider rollout. Extend to the next 2-3 teams. The shadow ban becomes credible only after the official tool has been seen working in the pilot team.
10. Day 90+. Continuous compliance. Monthly review of the bias-audit report, quarterly review of the sub-processor list, annual update of the DPIA/FRIA. The sanctioned tool's logs feed the regulator-facing artefacts; the shadow census is repeated quarterly to make sure shadow usage stays low.

Two things to expect. First, the works council step is where rollouts most often slow down — start it on day 60 even if the technical pilot is not ready, and run both in parallel. Second, the friction tuning in day 80-85 is usually the difference between a successful rollout and one that flatlines at 20% adoption. The shadow path will not stop being attractive just because it is now officially banned; it stops being attractive when the sanctioned path is genuinely faster.

8. Frequently asked questions

Can I just block ChatGPT on the corporate network?

You can, and many firms have. It does not solve the problem. The shadow usage moves to personal phones, personal laptops on the home network, and personal accounts on corporate machines. Network bans without a sanctioned alternative are a compliance theatre that produces lower visibility, not lower risk. The pillar guide's section on shadow AI covers the dynamic in detail.

Is GDPR Article 22 actually triggered by AI shortlisting?

The Court of Justice's SCHUFA ruling (C-634/21, December 2023) confirmed a broad reading of "based solely on automated processing". If the human decision-maker in practice rubber-stamps the AI output, the decision is treated as solely automated. The EDPB Guidelines 04/2024 reinforce this. The safe assumption is that any AI shortlisting tool needs Art. 22 hygiene built in: documented substantive human review, explanation, contestability.

Does the EU AI Act apply to my use of ChatGPT for HR even though I did not build it?

Yes. The AI Act applies to deployers of high-risk AI systems, not just providers. If your HR team uses an AI system for any Annex III paragraph 4 purpose, you are a deployer and you carry the deployer obligations: human oversight, monitoring, transparency to affected persons, record-keeping, fundamental rights impact assessment in many cases. Using a consumer LLM does not exempt you; it makes compliance harder because the documentation the provider should give you under Art. 13 is absent.

Is the Workday case binding outside the US?

Not directly, but the legal theory — that an AI vendor can be treated as the employer's agent for discrimination purposes — is travelling. The European Equality Bodies network has cited the case in its 2025 commentary, and several EU jurisdictions are considering similar interpretations under the Race Equality Directive (2000/43/EC) and the Employment Equality Directive (2000/78/EC). Plan as if the theory applies.

Can the works council block our AI deployment?

In several EU jurisdictions, including Germany under BVerfG §87(1)(6) and Italy under Art. 4 Statuto dei Lavoratori, yes — works council co-determination on monitoring and evaluation technology is mandatory. Failure to consult can lead to the deployment being suspended retroactively and to additional administrative sanctions. The right strategy is to start the consultation on day one of vendor selection, not after rollout.

Does redacting the CV solve the AI Act problem on its own?

No. Redaction solves the GDPR data-leakage problem for the LLM round-trip. The AI Act obligations — high-risk classification, risk management, human oversight, transparency, bias audit, FRIA — apply independently and are not satisfied by redaction. The two stacks work together: the redaction layer keeps the data minimisation principle clean; the human review and audit layer keep the AI Act obligations clean.

What is "algorithmic management" and how is it different from AI in HR?

Algorithmic management is the use of automated systems to allocate tasks, monitor performance, and make decisions about workers — most often associated with platform work (Foodinho/Glovo) but increasingly applied to traditional employment too. The Italian Garante decisions on Foodinho are the foundational EU enforcement; the EU Platform Work Directive (2024) extends specific transparency rights to platform workers. Any "AI scheduler", "AI shift optimiser", or "AI productivity scorer" in traditional employment carries the same legal exposure even if the firm does not see itself as a platform.

Are exit interview transcripts really high-risk?

Yes. They typically contain Art. 9 information (health, union membership), name third parties who have not consented to processing, and may contain whistleblowing content protected under EU Directive 2019/1937 with a separate confidentiality regime. They are also among the most reputationally damaging data categories to leak. Synthesising them through a consumer LLM is one of the highest-risk single shadow-AI behaviours we see.

How long does the 90-day rollout actually take?

In our experience the technical pilot takes 4-6 weeks and the works council consultation 6-8 weeks in EU jurisdictions where it applies. The 90 days assume parallel execution. If you run them sequentially, plan for 4-5 months. The biggest variable is procurement velocity, not technical complexity.

Where is Redactic's data residency for HR data specifically?

The compute and metadata storage are EU (Hetzner Helsinki + Cloudflare R2 EU jurisdiction). The Secure Chat routes to OpenAI / Anthropic / Google / Mistral on their EU endpoints with zero retention and no training on input under signed DPAs. The redaction certificate, sub-processor list, and DPA are public on the redactic.eu trust pages. For HR-specific use, the same architecture applies; nothing about the HR use case bypasses the no-trace guarantee.

If your firm has shadow AI use in HR — and it does — the cost of doing nothing is no longer an abstract regulatory risk. It is a Workday-style discrimination case waiting to be brought, an EEOC enforcement priority, a CNIL or Garante action template, a works council grievance, and a class action in the discrimination plaintiffs' bar's active pipeline. The fix is not to ban harder; the fix is to make the sanctioned path faster than the shadow one and to put a redaction layer between the team and the model. AI HR compliance and AI discrimination exposure are now interlocking risks — an undocumented shadow workflow is simultaneously a GDPR breach and a potential discrimination claim source. The architecture exists, the regulatory clock is running, and the 90-day rollout is realistic. Start with the census this week.

For the foundational technical detail behind the redaction layer — modes, certificates, no-trace architecture, multilingual detection — see the pillar guide. To see the workflow in action, the Redactic landing covers the four use-case modes with a live demo.