AI Agents for Lawyers

Regulation & Ethics

EU AI Act Article 4: The AI Literacy Duty for Law Firms, Explained

Article 4 of the AI Act requires every firm whose staff use AI to support their AI literacy. In force since 2 February 2025, supervised nationally since August 2026.

By Dr. Niklas SchmidtUpdated 13 November 20267 min read
On this page
  1. The text, before and after the Omnibus
  2. What the Commission's FAQ actually says
  3. Who enforces it and what happens if you ignore it
  4. What a proportionate programme looks like
  5. Documentation templates
  6. Mistakes that turn a small duty into a large one
  7. What to do this quarter

Article 4 of the EU AI Act obliges every law firm whose staff use AI systems to take measures that support their AI literacy. It has applied since 2 February 2025, was reworded but not deferred by the Digital Omnibus in July 2026, and has been supervised by national market-surveillance authorities since 2 August 2026. No certificate or test is required; a documented, proportionate programme is.

The text, before and after the Omnibus

Most commentary on Article 4 still quotes the 2024 wording. Regulation (EU) 2026/1744 changed it. The two versions side by side show what moved and what did not.

Original Article 4 (in force 2 Feb 2025)Article 4 as amended (in force 27 Jul 2026)
Core duty"take measures to ensure, to their best extent, a sufficient level of AI literacy""take measures to support the development of AI literacy"
Nature of obligationResult-oriented, softened by "to their best extent"Effort-oriented
Who is coveredStaff and "other persons dealing with the operation and use of AI systems on their behalf"Unchanged
Tailoring"taking into account their technical knowledge, experience, education and training and the context the AI systems are to be used in"Unchanged
New paragraphsNoneArt. 4(2)–(3): Commission and Member States support literacy; the AI Board may issue recommendations
Application date2 Feb 20252 Feb 2025, unchanged

The original wording is reproduced in Latham & Watkins' 2025 client note; the amended text is on the consolidated Article 4 page at artificialintelligenceact.eu. The practical difference is small. A firm that could show a serious, proportionate programme was safe under the old text; it is safer under the new one. A firm that has done nothing was exposed under both.

One point the amendment did not touch: deployers of high-risk systems still have a separate, harder duty under Article 26 to ensure staff assigned to human oversight are trained. The Commission's FAQ says so expressly. For law firms that becomes relevant on 2 December 2027, when Annex III obligations for HR and court-facing tools begin; the hub page on the Act has the full timetable.

What the Commission's FAQ actually says

The Commission's AI Literacy Questions & Answers, updated 27 July 2026 to reflect the Omnibus, is the closest thing to official guidance. Its answers, paraphrased:

  • Scope. Article 4 covers any organisation whose staff use AI systems in its professional activity, including general tools like ChatGPT. There is no exemption for small firms or for occasional use.
  • Form. No specific training format, certificate or knowledge test is required. Measures should fit the technical knowledge of the people, the context and the purpose of the systems in use.
  • Content. Understanding of the AI Act itself and of ethics and governance principles is "encouraged". So is understanding the specific risks of the tools in use, which for lawyers means hallucination, confidentiality and over-reliance.
  • Evidence. The Commission recommends an internal record of the training and measures delivered. The FAQ notes a living repository of more than 40 literacy practices submitted by organisations across sectors.
  • Enforcement. National market-surveillance authorities supervise and enforce from 2 August 2026. Penalties are set nationally and must be proportionate.

Nothing in the FAQ requires an external provider, a fixed number of hours or an exam. A firm can satisfy Article 4 with its own materials if they are competent and the record exists.

Who enforces it and what happens if you ignore it

Since 2 August 2026 the market-surveillance authority designated by each Member State has had the power to supervise Article 4. In Austria the KI-Servicestelle at RTR runs the national information hub with a dedicated KI-Kompetenz section; the surveillance role sits in the national framework built around it. There is no dedicated EU fine tier for Article 4; some compliance vendors still cite the 7.5 million euro or 1 percent tier, which applies to supplying incorrect information to authorities, not to literacy failures. Treat that claim as marketing.

The real enforcement of AI literacy in the legal sector has come from courts and bar regulators, not AI Act authorities. The Divisional Court in Ayinde v Haringey [2025] EWHC 1383 (Admin) addressed heads of chambers and managing partners directly on their responsibility for lawyers' use of AI. The SRA's warning notice of 17 August 2026 requires solicitors to understand a tool "well enough to review its output critically". Austria's ÖRAK rules put final responsibility on the lawyer. An Article 4 file is the evidence you produce when one of those bodies asks what the firm did to prevent a hallucinated citation or a leaked client letter. Missing it is a competence problem before it is an AI Act problem; the hallucination guide explains why the failure mode is so common.

What a proportionate programme looks like

Proportionality is the whole design principle. The table compares two firms.

Element10-lawyer firm200-lawyer firm
Tool inventoryOne page listing each tool, plan, data terms and permitted usesManaged register with owner per tool, reviewed quarterly
PolicyTwo pages: permitted tools, confidentiality rule, verification rule, escalationFull policy with role annexes for lawyers, secretaries, BD, IT, HR
Core trainingOne half-day session for everyone, delivered by a partner or an external trainerRole-based tracks: partners (risk and supervision), associates (prompting and verification), business services (confidentiality and Article 50 marking)
DepthHow LLMs work, hallucination, confidentiality settings, verification routine, the firm's rulesThe same, plus tool-specific modules per platform in use and an agent-and-automation track for builders
OnboardingNew joiners read the policy and attend the next sessionMandatory module within 30 days of joining; access to tools gated on completion
RefreshAnnual, or when tools changeAnnual refresh plus updates on new tools or incidents
RecordDated attendance list with the agenda attachedLearning-system export plus retained materials
OwnerNamed partnerAI governance lead reporting to management
CostA day of partner time or a short external courseInternal programme plus vendor training bundled into platform contracts

The half-day format works because the syllabus is short. Lawyers do not need to understand transformer architecture; they need to know what a large language model does with their text, why it invents citations, which of the firm's tools train on inputs, and what the verification routine is before anything leaves the building. Firms that want a ready-made version of the core session use external programmes: the AI Lab for Lawyers course covers fundamentals, prompting, research and drafting, agents and a regulation module in eight recorded hours, a syllabus that maps closely onto what an Article 4 record needs to show. Junior lawyers, who use the tools most and supervise least, are the group most worth over-investing in; the junior lawyers page explains why.

Documentation templates

Three documents are enough. Copy, adapt, date.

Training record

DateSessionFormatDelivered byAudienceAttendeesMaterials retained
2026-09-30AI fundamentals and firm rulesLive, 3 hPartner XAll fee earners and staff27 of 29 (2 absent, rescheduled 14 Oct)Slides v3, policy v2
2026-10-14Make-up sessionLive, 3 hPartner X22 of 2As above
2026-11-05Legora onboardingVendor webinar, 1 hVendorCorporate team9 of 9Recording link
2027-01-20New joiner moduleRecorded, 2 hSelf-study3 new joiners3 of 3Completion log

Policy clause

AI literacy (Article 4, Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744)

1. The firm supports the development of AI literacy of all partners, employees and
   contractors who use AI systems on its behalf.
2. Every person with access to a firm-approved AI tool completes the firm's core AI
   session before first use and a refresher at least every twelve months.
3. Role-specific training is provided where a person's use of AI carries additional
   risk, including supervision of AI-assisted work, client-facing systems and HR tools.
4. The firm keeps a dated record of sessions delivered, attendance and materials, and
   makes it available to the compliance partner and, on request, to a competent
   authority.
5. [Name] is responsible for this clause and reviews it annually.

Tool inventory line

Tool: Claude (Team plan) | Owner: [partner] | Data terms: not used for training; not ZDR
Permitted: drafting, summarising, research with verification | Prohibited: client
identifiers in consumer accounts | Training module: Core + Claude onboarding
Article 50: n/a (internal only) | Reviewed: 2026-09-18

A law-firm AI policy written in a day can carry all three; the policy is the frame, the record is the evidence.

Mistakes that turn a small duty into a large one

  • Treating the free ChatGPT account as outside scope. It is inside. Staff using personal accounts on firm work are using AI "on behalf of" the firm; the confidentiality checklist shows why those accounts are also the riskiest.
  • Buying a certificate instead of building a record. A certificate proves one person sat a course. A record proves the firm ran a programme fitted to its tools and roles, which is what Article 4 asks.
  • One session in 2025, nothing since. Tools changed completely between February 2025 and September 2026. A programme that does not mention agents, memory features or the firm's current platforms is out of date.
  • Training lawyers and forgetting business services. Marketing generates synthetic images that need Article 50 marks; HR evaluates tools that will be high-risk from December 2027. Both need their own module.
  • Confusing Article 4 with the GDPR file. They overlap but are separate. The deployer definition sits in the AI Act; lawful basis and DPIAs sit in the GDPR.

What to do this quarter

Book the core session, write the two-page policy, create the training record and inventory templates above, and hold the session before the year ends. Put the refresh in next year's calendar. For a small firm that is one partner-day of work; for a large one it is a project, but a bounded one. Either way the output is a file you can hand to a regulator, an insurer or a client's procurement team without embarrassment, which as of September 2026 is the standard that matters.

Frequently asked questions

Is AI training mandatory under Article 4 of the AI Act?

Article 4 requires providers and deployers to take measures to support the development of AI literacy among staff using AI on their behalf. Training is the obvious measure but not the only one; guidance notes, supervised onboarding and internal FAQs also count. The Commission's FAQ says no certificate, test or specific course is required. What matters is that measures exist, fit the people and tools, and are recorded.

Does Article 4 apply to a small law firm using ChatGPT?

Yes. The Commission's AI-literacy FAQ, updated 27 July 2026, states that Article 4 covers any organisation whose staff use tools like ChatGPT in a professional setting. Size is relevant only to proportionality: a three-partner firm can meet the duty with a short session, a written note on the risks of hallucination and confidentiality, and a dated record of who attended.

What is the penalty for breaching Article 4?

There is no dedicated EU fine tier for Article 4. Member States set national penalties, which must be effective, proportionate and dissuasive, and national market-surveillance authorities have supervised the duty since 2 August 2026. For law firms the realistic consequences are an information request, reputational exposure in a client audit, and use of the gap as evidence in a professional-conduct or negligence matter.

Did the Digital Omnibus remove the AI literacy obligation?

No. Regulation (EU) 2026/1744 reworded Article 4 from ensuring a sufficient level of AI literacy to taking measures to support its development, which converts a duty of result into a duty of effort. It also added paragraphs on Commission and Member State support and AI Board recommendations. The application date of 2 February 2025 was left untouched.

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