AI Agents for Lawyers

Practice Areas

AI for Employment Lawyers: Advice, Investigations and the AI Act's HR Rules

AI for employment lawyers: advising on AI in hiring and monitoring after the AI Act's Annex III deferral to December 2027, plus investigations and policies.

By Dr. Niklas SchmidtUpdated 13 November 202611 min read
On this page
  1. What employers are actually doing
  2. The AI Act's HR rules: what applies now and what waits
  3. Works councils, Article 22 and the rules that never paused
  4. The client-advice checklist
  5. Workplace investigations with AI: privilege, recording and the memo
  6. Policies, tribunal bundles and the extraction problem
  7. A three-lawyer firm's setup
  8. What to do this month

Employment lawyers advise on AI twice: once on the systems their clients are putting into hiring, monitoring and dismissal, and once on the tools inside their own practice. Littler's May 2026 survey of more than 300 US employers found 68% now have a formal AI-use policy, 37% are reassessing roles because of AI and 79% worry about AI-related litigation. The client side is where that litigation will be argued first.

What employers are actually doing

The Littler numbers are the sharpest picture of the demand side as of September 2026: 68% of employers with formal AI-use policies, up from 38% in 2025; 37% reassessing roles; 20% reducing hiring; 15% planning workforce cuts; 79% worried about AI-related litigation (Littler, 6 May 2026). Every one of those percentages is an instruction waiting to happen: a policy to draft, a consultation to run, a selection process to defend.

Law firms are in the same data. Baker McKenzie cut up to 10% of its global business-services staff, some 600 to 1,000 roles including knowledge, research and secretarial staff, citing its use of AI (February 2026) (RollOnFriday). Clifford Chance cut 10% of UK business services in November 2025, Irwin Mitchell cut its litigation paralegals, and Freshfields cited AI in paralegal redundancies in September 2025. Big Law hired 538 fewer graduates in 2025, the first fall since 2011, and Citi's survey found 67% of large firms expect fewer first-year associates by 2028. Wolf Theiss's Chief Digital & Technology Officer Leonhard Wassiq put the contrary view to Kurier in 2026: firms that hire fewer people now because of AI "will be hit hard" (Kurier). Either way, "AI" as a stated reason for redundancy is now a fact pattern, and the junior lawyers page has the rest of the numbers.

The AI Act's HR rules: what applies now and what waits

The single most-asked client question of 2026 has a date-shaped answer. Annex III of the AI Act lists AI used in recruitment, selection, evaluation, promotion and termination as high-risk. The Digital Omnibus on AI, politically agreed on 6–7 May 2026, confirmed by the Council on 13 May and reported as Regulation (EU) 2026/1744 in force from 27 July 2026, deferred the stand-alone Annex III obligations from 2 August 2026 to 2 December 2027 (Gibson Dunn); DLA Piper's global employment team has published an employment-specific note on the same deferral. What was not deferred matters more for the next fifteen months.

ProvisionHR relevanceStatus as of September 2026
Art. 5 prohibitionsEmotion recognition in the workplace (outside medical or safety uses), social scoring, certain biometric categorisationApply since 2 February 2025; fines up to €35m or 7% of turnover
Art. 4 AI literacyEvery employer whose staff use tools like ChatGPT must "take measures to support the development of AI literacy"; keep an internal training recordApplies since 2 February 2025; reworded as an obligation of effort by the Omnibus; national market-surveillance authorities supervise since 2 August 2026
Art. 50 transparencyHR chatbots must identify themselves as AI; synthetic content must carry machine-readable marksApplies since 2 August 2026, with a watermarking grace period to 2 December 2026 for existing systems
Annex III high-risk (employment)Risk management, data governance, logging, human oversight, deployer duties including informing workers' representatives before putting a system into serviceDeferred to 2 December 2027
Annex I product-embeddedAI inside safety components of regulated productsDeferred to 2 August 2028

The Commission's AI-literacy FAQ (updated 27 July 2026) says Article 4 reaches any organisation whose staff use general-purpose tools, requires no certificates or tests, and expects an internal training record (Commission FAQ). There is no dedicated EU fine tier for Article 4; enforcement is national. The Omnibus page tracks what changed and what did not, and the Article 4 page is the one to send an HR director.

Works councils, Article 22 and the rules that never paused

The deferral moved one set of obligations. It did not touch the law under which most AI-in-HR disputes are actually fought.

GDPR Article 22 still gives an individual the right not to be subject to a decision based solely on automated processing that produces legal or similarly significant effects, and a rejection, a performance rating or a dismissal recommendation generated by a model is the textbook example. The practical questions are the old ones: is there meaningful human involvement, what is the lawful basis, was a data protection impact assessment done, and can the employer explain the logic. The GDPR and AI tools page covers lawful basis and DPIAs; the privacy lawyers page has a DPIA prompt skeleton that transfers directly to an HR deployment.

In Germany, § 87(1) no. 6 BetrVG gives the works council co-determination over the introduction and use of technical devices designed to monitor employee conduct or performance, and a model that scores productivity, flags anomalies in time recording or ranks candidates is such a device. In Austria, § 96 ArbVG requires works-council consent for control measures and systems that touch human dignity, and § 96a covers personnel data systems. When the Annex III duties arrive, the deployer must additionally inform workers' representatives and affected workers before putting a high-risk system into service. The sequencing advice for a client rolling out an HR tool in 2026 is therefore: works council and data protection first, AI Act documentation in parallel so that it is ready by December 2027.

The client-advice checklist

This is the list to run against any AI-in-HR proposal a client brings you, whether it is a CV screener, a monitoring dashboard or a chatbot for policy questions.

  1. Classify the use. Prohibited (Art. 5), high-risk (Annex III, deferred), transparency-only (Art. 50), or none of these. Emotion recognition in the workplace is prohibited now; a CV ranker is high-risk in December 2027; a policy chatbot must say it is a bot today.
  2. Identify the deployer. Under the AI Act the employer is the deployer; the vendor is the provider. Get the vendor's technical documentation and its own compliance timetable in writing.
  3. Check for solely automated decisions. If the system makes or effectively makes the decision, Article 22 applies; design in a human reviewer who can and does overrule it, and document that they do.
  4. Run the DPIA. Processing employee data with a new technology to evaluate or monitor is the standard trigger. Cover the training-data question raised in EDPB Opinion 28/2024: if the model was trained unlawfully, the deployer must assess that too.
  5. Involve the works council before purchase, not before go-live. Co-determination rights bite on the introduction of the system; a signed works agreement is the artefact the client will need later.
  6. Meet Article 4 now. A short, recorded training session for the HR team and the managers who will use the tool; keep the attendance record.
  7. Write the internal policy. Permitted tools, prohibited uses (no employee data in consumer chatbots), who can approve exceptions, and how outputs are checked. Littler's 68% figure means the client's competitors already have one.
  8. Plan the December 2027 file. Logging, human-oversight instructions, accuracy monitoring and worker information: start the folder now so the deferral is used, not wasted.
  9. Fix the dismissal narrative. If AI is the reason for redundancies, the selection criteria and consultation must stand on their own; "the system recommended it" is not a fair reason anywhere in Europe.

Workplace investigations with AI: privilege, recording and the memo

Investigations are where employment lawyers' own AI use is most valuable and most exposed. The value is obvious: an inbox export, a chat log and a dozen interview notes become a dated chronology with a source reference per entry in an afternoon. The exposure comes in two places.

The first is privilege. In United States v. Heppner (S.D.N.Y., bench ruling 10 February 2026, opinion 17 February 2026) Judge Rakoff held that memos a defendant produced with consumer Claude were not privileged or work product, because "Claude is not an attorney" and the tool's terms allowed logging and disclosure; privilege might attach, he indicated, where counsel directs the AI use in the manner of a Kovel agent (Harvard Law Review Blog). Warner v. Gilbarco (E.D. Mich., February 2026), reported by a single vendor blog, went the other way for lawyer-directed work product, treating AI programs as "tools, not persons". The lesson for an investigation is that the HR director must not run the evidence through a chatbot before calling you, and that once you are instructed, the AI work should sit inside the legal engagement, on a tier whose terms you have read. The privilege page sets out both cases.

The second is recording. NYC Bar Formal Opinion 2025-6 (22 December 2025) on AI recording of client meetings and 2026-2 (5 August 2026) on non-client conversations require the consent of all parties, prohibit surreptitious recording and say the default is not to record (NYC Bar). Twelve US states, including California, Florida, Illinois, Maryland, Pennsylvania and Washington, require all-party consent by statute, and Duane Morris' February 2026 note recommends switching transcription off by default, documenting consent, checking vendor retention and training terms, and adding attorney impressions to any transcript (Duane Morris). An investigation interview with a complainant is the last place to discover that a meeting bot's terms allow processing for service improvement. The meeting notes page has the setup.

A workable investigation flow: collect the documents into a matter folder on an enterprise tier; have the model build the chronology and a list of every factual assertion in each interview note with the document that supports or contradicts it; interview with a human note-taker unless every participant has consented in writing; draft the findings yourself, using the model only to check the report's internal consistency and to list the assertions that rest on no evidence. Anthropic's open-source employment plugin ships a "privileged investigation memos" skill built on the same division of labour, alongside hiring and termination review, worker classification and a leave tracker (github.com/anthropics/claude-for-legal); the plugins page explains how the cold-start interview and the review gates work.

Policies, tribunal bundles and the extraction problem

Three routine tasks round out the practice side. Drafting client AI policies is now volume work; the North Carolina Bar's "traffic-light" structure (red: client or employee data in consumer tools; yellow: research and drafting with verification; green: administrative and marketing use) translates directly into an employer policy, and the firm AI policy page has an outline you can adapt.

Tribunal and court bundles are an extraction job: payslips, rotas, WhatsApp exports and scanned contracts into a chronology and a schedule of loss. Run an OCR pass before the model sees scanned documents; Claude will say "illegible" rather than invent a figure, but only when it is looking at an image rather than a bad text layer. The PDF extraction playbook covers the pipeline. And the filing rule from litigation applies without modification: every authority in a skeleton or a brief is opened and read by the lawyer who signs it. Damien Charlotin's database of hallucination decisions stood at 2,041 on 14 September 2026 and employment tribunals are not exempt from it.

A three-lawyer firm's setup

The most instructive employment-practice deployment in the research is also the cheapest. A German three-lawyer employment and tenancy firm runs self-hosted n8n with Nextcloud, local OCR and a local language model on a German server for mail intake, client onboarding and human-in-the-loop deadline capture, at about EUR 20–30 a month with no licence fees. After six months: mail sorting fell from two hours a day to forty minutes, file creation from 15–40 minutes to a five-minute check, client acknowledgements from one or two days to minutes, zero deadline-relevant items missed, and six to seven hours a week reclaimed. The legal paperwork was a § 43e BRAO contract with the provider plus an Article 28 GDPR processing agreement (case study). The n8n page shows how the same design works on a hosted plan for firms that do not want to run a server. Module 4 of the AI Lab for Lawyers course builds an intake-and-deadline workflow of this kind in n8n without code.

What to do this month

Pick the client with the largest planned AI rollout in HR and run the nine-point checklist against it; the works-council and Article 22 items usually surface a gap the vendor's compliance deck did not mention. Inside the practice, move investigations onto an enterprise tier with written terms, set every meeting bot to off by default, and put the OCR-then-extract pipeline in front of the next bundle. Then put 2 December 2027 in the diary for every high-risk HR system a client has told you about; the deferral is the time to build the file, and the clients who use it will be the ones still able to explain their systems when the obligations arrive.

Frequently asked questions

Is AI in recruitment high-risk under the EU AI Act?

Yes, AI used for recruitment, selection, evaluation, promotion and termination decisions is listed in Annex III. Under the Digital Omnibus agreed in May 2026 and reported in force from 27 July 2026, the obligations for stand-alone Annex III systems were deferred to 2 December 2027. Until then the prohibitions in Article 5, the Article 4 literacy duty and the Article 50 transparency duties apply, and GDPR and national employment law govern the deployment.

Do works councils have to be involved before an employer uses AI in HR?

In Germany and Austria, usually yes where the system can monitor conduct or performance: § 87(1) no. 6 BetrVG gives the works council co-determination over technical devices designed to monitor employees, and § 96 ArbVG requires works-council consent for control measures that touch human dignity. Once the AI Act's Annex III duties apply, deployers must also inform workers' representatives before putting a high-risk system into service.

Can I use AI to write a workplace investigation report?

You can use it to organise evidence, build a chronology and draft sections you then rewrite, provided the tool sits on a business or enterprise tier that does not train on inputs and the work is directed by counsel. A federal court held in February 2026 that a party's own consumer-chatbot memos were not privileged; work product directed by lawyers is on stronger ground. Keep the findings and credibility assessments in the lawyer's own words.

Can I record and transcribe investigation interviews with an AI note-taker?

Only with the consent of everyone recorded. NYC Bar Formal Opinions 2025-6 and 2026-2 require consent of all parties, prohibit surreptitious recording and say the default should be not to record; a dozen US states require all-party consent by statute. Document the consent, review the vendor's retention terms, and add the lawyer's own impressions to any transcript so work-product protection is clear.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Regulation & Ethics

The EU AI Act for Lawyers and Law Firms: What Applies in 2026

The EU AI Act applies to law firms as deployers. Article 4 literacy and Article 50 transparency are live; high-risk duties are deferred to December 2027.

9 min read

Regulation & Ethics

The Digital Omnibus on AI: What Changed in the AI Act and What Did Not

The Digital Omnibus AI Act amendment, Regulation (EU) 2026/1744, deferred high-risk duties to December 2027 but left transparency, literacy and the prohibitions in force.

6 min read

Use Cases

AI Meeting Notes and Transcription for Lawyers: Consent, Privilege and Setup

AI meeting notes are ethical for lawyers only with all-party consent, a no-training data path and a retention rule. Bar opinions, tools and a file-note prompt.

8 min read

Regulation & Ethics

Attorney–Client Privilege and AI: Heppner, Warner, the NYT Logs and Public Chatbots

Consumer chatbot memos are not privileged. US v. Heppner, Warner v. Gilbarco and the 20-million-log OpenAI order, plus rules that keep privilege intact.

10 min read

Tools

Claude for Legal Plugins Explained: The 12 Practice-Area Plugins and 20+ Connectors

Claude for Legal gives paid Claude users 12 open-source practice-area plugins, 150+ skills and 20+ connectors to iManage, NetDocuments and CoCounsel.

8 min read

Tools

n8n for Law Firms: Building Legal Automations Without Code

n8n is a self-hostable workflow builder with AI nodes. A German firm runs intake and deadlines on it for €20–30 a month; three law-firm workflows, node by node.

9 min read