The EU AI Act for in-house counsel: what applies to users

A legal department that uses Claude, Legora or another generative tool is a deployer within the meaning of the EU AI Act. That gives three obligations that apply today, and a fourth that applies only if the tool is used in a particular way. Here is what applies, with dates, and what is overstated.

Updated by Nils-Erik Jansson

The department's role

The Act distinguishes between two roles. The provider develops an AI system and places it on the market. The deployer uses the system under its own authority in its business [1]. A legal department that has a contract with Anthropic, Legora or Microsoft and lets its lawyers work in the tool is a deployer. That remains the case when the tool is configured with your own playbooks and connectors. Adding your own rules does not change the role.

This means that most of the Act is not addressed to you. The requirements on risk management, technical documentation, logging and CE marking fall on the provider of high-risk systems. For a deployer of a general purpose tool in legal work, what remains is set out in the table.

Provision Date What applies
Article 4, AI literacy 2 February 2025, new wording 27 July 2026 Measures to support AI literacy among staff
Chapter II, prohibited practices 2 February 2025 Ban on certain AI practices
Article 50, transparency 2 August 2026 Disclosure of artificially generated content in certain cases
Chapter III, high-risk, Annex III 2 December 2027 High-risk rules for systems in Annex III
Chapter III, high-risk, Annex I 2 August 2028 High-risk rules for systems in Annex I

Article 4: AI literacy, new wording since 27 July 2026

Article 4 has applied since 2 February 2025. It was given new wording by Regulation (EU) 2026/1744, which entered into force on 27 July 2026 [2]. Providers and deployers must take measures to support the development of AI literacy of their staff and of other persons dealing with the operation and use of AI systems on their behalf. The measures take into account the persons' technical knowledge, experience, education and training, the context in which the systems are used and the persons on whom they are used. The Article states that the obligation does not require providers or deployers to guarantee any specific level of AI literacy of any individual [1] [2].

According to the recitals of the amending Regulation, the earlier wording meant that staff had to have a sufficient level of AI literacy [2]. The Commission is to publish practical examples of how to comply with the new obligation [1].

In practice, you should be able to show that your lawyers know how the tool works, where it falls short and what your own rules are. In my view, a two hour training session with documentation, guidelines and the same for new joiners is enough. There is no requirement for a particular level or a certificate. Article 4 is not among the provisions to which Article 99(3) to (5) attach fines. Under Article 99(1), the Member States lay down rules on penalties and other enforcement measures for infringements of the Act [1]. National rules are not covered here. The Swedish Bar Association's guidance points to the Article as a reason for law firms to train their staff [3].

My advice is not to buy a generic company wide "AI licence" course for the sake of Article 4. Buy or build the training your lawyers actually need for the tools they have, and document it.

Chapter II: prohibited practices, since 2 February 2025

The Act's ban on certain AI practices has applied since the same date [1]. It covers manipulation, social scoring, certain biometric surveillance and similar practices. Two prohibitions have been added and apply from 2 December 2026. They cover AI systems that generate sexual material of identifiable persons without their consent, and such material involving children as referred to in Directive 2011/93/EU [1] [2]. A legal department does not come close to the prohibitions with contract review. But anyone who is asked by the business about a tool that assesses customer behaviour or employees' emotions needs to know that the list exists.

Article 50: transparency, since 2 August 2026

Article 50 has applied since 2 August 2026 [1]. Deployers of systems that generate or manipulate image, audio or video content constituting a deep fake must disclose that the content has been artificially generated. Deployers who publish AI generated text to inform the public on matters of public interest must disclose this. That does not apply where the text has been reviewed by a person and someone holds editorial responsibility for the publication [1]. Providers of generative systems that were on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2) [2].

For a legal department that uses the tools internally and reads everything before it leaves the department, the Article will seldom apply. If the communications team publishes AI written texts about the company's legal matters, you should know about it.

High-risk: later, and only for certain uses

The high-risk rules in Chapter III apply from 2 December 2027 for systems in Annex III and from 2 August 2028 for systems in Annex I. The dates are in Article 113 as amended by Regulation (EU) 2026/1744 [1] [2].

For you, the high-risk rules apply only if the tool is used for a purpose listed in Annex III. Contract review, research, drafting and summaries are not such purposes. Recruitment, decisions about employees and assessing the creditworthiness of natural persons are [1]. The practical rule is simple. As long as the tool is used for legal work on contracts and documents, you are outside the scope. As soon as someone wants to use it for decisions about people, the question should come to you.

What is overstated

Three common misconceptions. First, that everyone who uses generative AI must carry out an impact assessment under the AI Act. That obligation, in Article 27, applies to certain deployers of high-risk systems and does not apply to you [1]. Second, that the Act requires the use of AI in contract work to be notified to a public authority. It does not. Third, that Article 4 means certification. It does not, see above.

What does apply today, independently of the AI Act, is the GDPR. Personal data in prompts is processing of personal data, a data processing agreement and a third country transfer assessment are required, and that is where the real work lies. See the five questions under AI for legal.

The training and the guidelines in the data protection package are documented as measures under Article 4, with a proposal for annual follow-up. For organisations introducing Claude, they are part of the setup. What the guidelines should contain is set out in the guide on an AI policy for the legal department.

Read also AI literacy training for the legal department.

Frequently asked questions

Is a legal department that uses Claude a deployer under the EU AI Act?
Yes. Anyone who uses an AI system under their own authority in their business is a deployer. The party that develops the system and places it on the market is the provider. The obligations of a deployer of a general purpose tool are limited: AI literacy under Article 4, transparency under Article 50 in certain situations, and the high-risk rules only if the system is used for a purpose that the Act classifies as high-risk [1].
Do we have to certify our staff in AI literacy?
No. Since 27 July 2026, Article 4 requires you to take measures to support the development of AI literacy among your staff, taking into account their technical knowledge, experience, education and training and the context in which the systems are used. The Article states that the obligation does not require you to guarantee any specific level of AI literacy of any individual [1] [2]. It sets no certification and no particular form. In practice you should be able to show what you have done: training, guidelines, and the same for new joiners.
Is contract review with AI high-risk?
Not in itself. High-risk classification follows from the area of use in Annex III to the Act, for example recruitment and decisions about employees, or assessing the creditworthiness of natural persons. The power of the tool does not decide it. A legal department that reviews contracts, searches its own documents or writes drafts is outside the scope. If the same tool is used to sort job applicants, the assessment is different. The high-risk rules for Annex III also apply later than the rest of the Act, from 2 December 2027 [1].

Sources

  1. [1]Regulation (EU) 2024/1689 (the AI Act), consolidated text of 27 July 2026, Articles 3, 4, 5, 27, 50, 99 and 113 and Annex III, EUR-Lex, read 2026-11-17; the consolidated text is a documentation tool, the authentic text is in the Official Journal
  2. [2]Regulation (EU) 2026/1744 (Digital Omnibus on AI), recital 8 and Article 1(5), (7), (39) and (40), EUR-Lex, read 2026-11-17; new wording of Article 4, two new prohibitions, transitional rule in Article 111(4) and dates of application in Article 113
  3. [3]Vägledning om användning av generativ AI i advokatverksamhet, Sveriges advokatsamfund, 12 juni 2025, in Swedish; Swedish Bar Association guidance on generative AI, section 5

Back to AI for legal

More under AI for legal