AI Agents for Lawyers

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.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. The dates that matter after the Digital Omnibus
  2. Is a law firm a deployer?
  3. Which legal-sector uses could be high-risk
  4. What applies to your firm today
  5. The GPAI Code of Practice and what it means for the tools you buy
  6. Penalties, and who is actually exposed
  7. Austria and the national layer
  8. A compliance checklist for the firm
  9. Advising clients: the second checklist
  10. Where this leaves you

The EU AI Act applies to law firms as deployers of AI systems. As of September 2026, three sets of duties are live: the Article 5 prohibitions and the Article 4 AI-literacy duty (since 2 February 2025) and the Article 50 transparency rules (since 2 August 2026). The high-risk regime that could catch legal-sector tools was pushed back by Regulation (EU) 2026/1744 to 2 December 2027.

The dates that matter after the Digital Omnibus

The original AI Act timetable was rewritten by the Digital Omnibus on AI, published in the Official Journal on 24 July 2026 as Regulation (EU) 2026/1744 and in force since 27 July 2026. The table below is the timetable as it stands now; the full account of what the Omnibus changed is on its own page.

ObligationApplies fromWhat it means for a law firm
Art. 5 prohibited practices2 Feb 2025Live. Rarely engaged by legal work, but check any emotion-recognition or social-scoring feature in HR or intake tools
Art. 4 AI literacy2 Feb 2025; supervised by national authorities since 2 Aug 2026Live for any firm whose staff use ChatGPT, Claude, Copilot, Harvey or Legora
GPAI model obligations (Chapter V)2 Aug 2025; AI Office enforcement powers since 2 Aug 2026Not yours. Falls on model providers, but shapes which vendors are safe to rely on
Art. 50 transparency2 Aug 2026; watermarking grace period to 2 Dec 2026 for existing systemsLive if you run a client-facing chatbot or publish AI-generated audio, image or video
Two new Art. 5 prohibitions (non-consensual intimate imagery, CSAM generation)Transition to 2 Dec 2026Irrelevant to practice, relevant when advising platform clients
Annex III stand-alone high-risk (administration of justice, employment, and others)2 Dec 2027 (was 2 Aug 2026)Deferred. The open question about legal research tools is parked until then
Annex I product-embedded high-risk2 Aug 2028Deferred. Relevant to product-liability and medical-device clients

Sources: Gibson Dunn and the Cloud Security Alliance research note, which puts it in three words: "deferred, not cancelled".

Is a law firm a deployer?

Almost always. The Act defines a deployer as any natural or legal person using an AI system under its authority in the course of a professional activity. A partner who pastes a clause into Claude under the firm's Team plan is using an AI system under the firm's authority in a professional activity. So is an associate running a Legora review, a secretary using Copilot in Outlook, or a marketing team generating LinkedIn images. The Commission's own AI-literacy FAQ, updated 27 July 2026, says in terms that Article 4 covers any organisation whose staff use tools like ChatGPT.

Two further roles matter. A firm becomes a provider if it develops an AI system and places it on the market under its own name: a firm that sells a client-facing agent on subscription, as A&O Shearman began doing with four Harvey-built agents in April 2025, may find itself with provider-side obligations for that product. And a firm that materially modifies a high-risk system can be treated as its provider. For most firms, neither applies today; both are worth a line in the tool inventory so the question is asked before a product launches.

The short answer to "does the Act apply to us" is on the direct-answer page. The rest of this page is about what it requires.

Annex III lists the "administration of justice and democratic processes" area, which covers AI systems intended to be used by a judicial authority, or on its behalf, to assist in researching and interpreting facts and law and in applying the law to facts. Court-facing tools are clearly in scope. Whether a lawyer-side research assistant used to prepare submissions is "on behalf of" a judicial authority is unsettled; the mainstream reading is that it is not, but no authority has ruled and no guidance has been issued. The Omnibus parked the question: nothing in Annex III bites before 2 December 2027.

The second Annex III area that reaches every firm of any size is employment. AI used for recruitment screening, task allocation, performance monitoring or termination decisions is high-risk. A firm that uses an AI tool to rank associate applications or to score time-recording narratives for productivity is a deployer of a high-risk system from December 2027, with duties of human oversight, logging, worker information and staff training that survive the Omnibus. The employment-law page covers the HR rules from the adviser's side.

What applies to your firm today

Article 4: AI literacy

The Omnibus reworded Article 4 from ensuring "a sufficient level" of AI literacy to taking "measures to support the development of AI literacy", an obligation of effort rather than result, according to the consolidated text at artificialintelligenceact.eu. It did not defer it. National market-surveillance authorities have supervised it since 2 August 2026. There is no dedicated EU fine tier; the Commission FAQ says no certificates or tests are required and recommends an internal training record. The Article 4 page sets out a proportionate programme for a ten-lawyer and a 200-lawyer firm, and the AI literacy guide covers what to teach.

Article 50: transparency

Since 2 August 2026, an AI system that interacts directly with people must tell them they are dealing with AI, unless it is obvious. If your website runs an intake chatbot, it must say so. Synthetic audio, image and video must carry machine-readable marks; providers of existing generative systems have until 2 December 2026 to implement watermarking, per Help Net Security's summary of the enforcement changes. Over 180 organisations had signed the voluntary code of practice on AI-generated-content transparency by August 2026. Anthropic reportedly added a statistical watermark to Claude Fable 5.1 outputs for this reason, though that rests on a secondary source.

Article 5: prohibitions

The prohibitions have applied since 2 February 2025. None targets legal practice, but two can be triggered by tools a firm buys: emotion recognition in the workplace, and social scoring. If an HR or client-intake vendor advertises sentiment analysis of interviews or calls, ask the vendor for its Article 5 analysis before deployment.

The GPAI Code of Practice and what it means for the tools you buy

The obligations for general-purpose AI models have applied since 2 August 2025. The GPAI Code of Practice, published 10 July 2025, was signed in full by OpenAI, Google, Anthropic, Microsoft, Mistral, Amazon, IBM, Cohere and Aleph Alpha; xAI signed only the safety and security chapter; Meta refused, citing "legal uncertainties", as The Register reported. The signatory list is maintained on Wikipedia.

Since 2 August 2026 the Commission's AI Office can demand documentation, evaluate models, order corrective measures, withdraw a model from the EU market and fine providers up to 15 million euro or 3 percent of global turnover. Providers of models released before 2 August 2025 have until 2 August 2027 to comply, per artificialintelligenceact.eu's enforcement note.

For a firm this is a procurement question, not a compliance one. A model provider that has not signed the Code carries more regulatory risk, and a firm whose contract and workflow depend on that model should know it. The vendor due-diligence page has the questions to ask.

Penalties, and who is actually exposed

InfringementMaximum fineWho is exposed
Prohibited practices (Art. 5)35 million euro or 7% of global turnoverProviders and deployers
GPAI and most other obligations15 million euro or 3%Model providers now; deployers of high-risk systems from Dec 2027
Incorrect or misleading information to authorities7.5 million euro or 1%Anyone answering a regulator's request
Article 4 AI literacyNo dedicated EU tier; national and "proportionate"Deployers, via national market-surveillance authorities

Fines for law firms will be rare. The realistic exposure for a firm in 2026 is different: a regulator's information request that reveals there is no training record, a client RFP that asks for the firm's AI Act position and gets no answer, or a professional-conduct complaint in which the bar treats AI Act non-compliance as evidence of a competence failure. Bar rules, not the AI Act, remain the sharper instrument, as the ÖRAK, BRAK and CCBE guidance shows.

Austria and the national layer

Article 4 supervision sits with national market-surveillance authorities, so the practical questions are national. In Austria the information hub is the KI-Servicestelle at the regulator RTR, set up under § 20c KOG, which publishes a dedicated "KI-Kompetenz" section and issued guidance on the prohibited practices on 4 February 2025 (rtr.at). The Austrian bar layered its own four practice rules on top at its first KI-Tag in March 2026: client data stays on EU servers, disclose AI use to clients, the lawyer keeps final responsibility, no client data for training.

Data protection runs in parallel and is not displaced by the AI Act; the GDPR page covers lawful basis, DPIAs and vendor terms. Module 6 of the AI Lab for Lawyers course, run by Dr. Niklas Schmidt of Wolf Theiss in Vienna, covers risks and regulation for practitioners in exactly this layered EU-plus-national setting.

A compliance checklist for the firm

  1. Inventory every AI system in use, including consumer accounts staff use on their own initiative. Record vendor, plan, data terms and who uses it.
  2. Classify each entry: general assistant, legal platform, HR or recruitment tool, client-facing chatbot, content generator.
  3. Confirm no Annex III use is live today; diarise December 2027 for HR tools and any court-facing product.
  4. Write a two-page AI-use policy. The policy page has an outline.
  5. Deliver role-based AI literacy training and keep a dated attendance record with the content covered.
  6. Label every client-facing chatbot as AI and check synthetic-content outputs carry marks.
  7. Ask each vendor for its AI Act position: provider or deployer, GPAI Code signatory status, Article 50 marking.
  8. Name an owner for AI governance and one calendar review a year.
  9. Align the file with GDPR: DPIA where needed, processor terms, retention.
  10. Keep the whole file short enough that a regulator's request can be answered in a day.

Advising clients: the second checklist

Clients hear "the AI Act is delayed" and stop. The memo that earns its fee corrects that in the first paragraph and then does the following.

  1. Map the client's systems to roles: provider, deployer, importer, distributor.
  2. Flag anything that could be Annex III high-risk and set a compliance plan against 2 December 2027, not against "sometime".
  3. Check Article 50 duties that are already live: chatbots, deepfake labelling, synthetic-content marking.
  4. Confirm the Article 4 training programme exists and is documented.
  5. For platform and consumer-facing clients, check the two new Article 5 prohibitions before 2 December 2026.
  6. Set up regulatory monitoring for the delegated acts and guidelines the Commission still owes.

Where this leaves you

For a law firm the AI Act in September 2026 is a modest, documentable set of duties: train people and record it, label your chatbot, know what your vendors are, and keep an eye on December 2027. The pieces that generate fear, high-risk conformity assessments and seven-figure fines, are either deferred or aimed at model providers. The pieces that generate professional risk, hallucinated citations and leaked client data, are governed by bar rules and the GDPR, and those have not moved at all.

Frequently asked questions

Does the EU AI Act apply to law firms?

Yes. A firm that uses AI systems in its professional activity is a deployer under Article 3. That triggers the Article 4 AI-literacy duty, the Article 5 prohibitions and, if the firm runs a chatbot or publishes synthetic content, the Article 50 transparency rules. High-risk obligations would apply only if a firm's system falls within Annex III, and those are deferred to 2 December 2027.

Is the AI Act delayed for law firms?

Only the high-risk chapter. The Digital Omnibus on AI, Regulation (EU) 2026/1744, in force since 27 July 2026, moved Annex III stand-alone high-risk obligations to 2 December 2027 and Annex I product-embedded ones to 2 August 2028. Article 4 literacy, Article 5 prohibitions and Article 50 transparency were not deferred and are already in force.

What are the fines under the EU AI Act?

Prohibited practices carry up to 35 million euro or 7 percent of global turnover; most other obligations up to 15 million euro or 3 percent; supplying incorrect information up to 7.5 million euro or 1 percent. Article 4 has no dedicated EU fine tier; enforcement is national and must be proportionate. Fines are the ceiling, not the norm, for a law firm.

Are legal research tools high-risk AI under the AI Act?

Annex III lists AI used by or on behalf of judicial authorities to research and interpret facts and law. Whether tools used by lawyers rather than courts are caught is unsettled. The question is academic until 2 December 2027 because the Omnibus deferred all Annex III obligations. Vendors of court-facing tools should plan for classification; law firms should ask vendors how they see it.

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