Skip to content
Normalex

Search a tool or a guide

← All guides

Writing the firm's AI policy

What the bar, the AI Act and the GDPR ask of a lawyer who uses artificial intelligence, and how to write it into a policy the whole firm can follow.

Updated on 7 October 2026

AI use policy for the firm
On this page
  1. The essentials
  2. What the bar guidelines say
  3. The AI Act for a law firm
  4. Professional secrecy and the closed environment
  5. Data protection: register, contract, impact assessment
  6. Writing the policy in twelve sections
  7. Common mistakes and open points
  8. Sources

The essentials

  • Using AI is neither forbidden nor mandatory. The OVB and AVOCATS.BE guidelines (January 2025) apply existing deontology: pseudonymise, never enter secret information outside a closed environment, check that the sources cited exist, remain responsible for the output.
  • These guidelines do not require a written policy, and there is no general duty to tell the client that AI is used.
  • Since 27 July 2026, art. 4 of the AI Act, replaced by Regulation (EU) 2026/1744, asks the firm to take measures to support the AI literacy of its staff, with no guaranteed level.
  • Professional secrecy has been governed since 1 September 2026 by art. 352 of the new Penal Code (formerly art. 458).
  • Each AI tool that processes personal data is a processing activity for the GDPR register, needs a processing agreement and, most often, an impact assessment before first use.

The AI use policy writes these rules clause by clause, each with its source, plus a one-page summary for staff.

What the bar guidelines say

The guidelines were written jointly by the OVB and AVOCATS.BE and published in January 2025. They create no deontological rule specific to AI: they apply art. 455 of the Judicial Code (dignity, probity, delicacy) and the codes of conduct. They ask the lawyer:

  1. to know the basics of AI and large language models;
  2. to read the tool's terms of use, in particular on training, transfer and storage of the data, its reuse by the provider, where it is processed, whether the system is open or closed, the provider's liability and intellectual property;
  3. to verify the output, and in particular that the legislation, doctrine and case law cited exist;
  4. to pseudonymise personal data before entering it;
  5. where personal data is essential, to be transparent and to have a lawful basis;
  6. never to enter documents or information covered by professional secrecy or a duty of confidentiality, unless absolutely certain of a closed environment with sufficient safeguards;
  7. to tell the users of a chatbot that they are dealing with an automated system.

The lawyer always remains ultimately responsible for the output. The guidelines address neither billing, nor training beyond knowing the basics, nor the supervision of staff: that is the role of a firm policy. The model charters of the OFABB (March 2026) and the Paris bar (July 2026) suggest a structure: authorised tools, data, verification, internal labelling, training, incidents, control, adoption.

The AI Act for a law firm

A firm that uses an AI tool in its work is its deployer. Four rules concern it:

  • AI literacy (art. 4). Since 27 July 2026, the deployer takes measures to support the development of the AI literacy of its staff and of the persons acting on its behalf. The European Commission says no certificate is needed: an internal record of training is enough, and staff who use a chatbot should know its risks, hallucinations among them.
  • Prohibited practices (art. 5). Since 2 February 2025, AI may not be used to infer the emotions of a person at the workplace, except for medical or safety reasons.
  • Transparency (art. 50). Since 2 August 2026, people using a chatbot are told they are dealing with AI; generated images, sound and video are disclosed; generated text published on matters of public interest is disclosed too, unless a person reviewed it under the firm's editorial responsibility.
  • High risk (Annex III and art. 26). Recruiting and evaluating staff, and AI used to decide in arbitration, are high-risk uses. The deployer's duties apply from 2 December 2027. A party lawyer's research and drafting tools do not fall into this category.

Professional secrecy and the closed environment

The guidelines set the strictest rule: no secret information in an AI tool, unless absolutely certain that the tool runs in a closed environment with sufficient safeguards, for example within the firm's own IT perimeter, with nothing shared outside. A breach of professional secrecy is punished under art. 352 of the Penal Code.

A business online service, covered by a processing agreement, with no training and no retention of prompts, no access by the provider's staff and hosted in the Union, offers strong safeguards. Whether that is enough to make it a closed environment is not settled. Prudence suggests accepting by default only tools installed at the firm or in an infrastructure it controls, and making any other choice an explicit, reasoned decision.

In every case, pseudonymisation reduces the risk: remove or replace names, addresses, numbers and identifying details before sending. Localex does it in the browser, on the device.

Data protection: register, contract, impact assessment

The lawyer who chooses the data and the purpose is the controller; the tool's provider is in principle the processor. Four tasks follow:

  • record each tool in the record of processing activities (art. 30 GDPR), with its purposes, data, recipients, transfers and retention;
  • sign a processing agreement (art. 28) limiting the use of the data to the firm's instructions;
  • check transfers outside the European Economic Area (arts 44 to 46) and prefer storage in the Union;
  • check whether an impact assessment is required (art. 35). It is not automatically mandatory for a firm, but the APD's guidance considers that two criteria call for one in most cases. Generative AI used on files holding health, criminal or family data often meets two.

The AI tools register records each tool's safeguards and flags the points to check; the DPIA for an AI tool screens the need for a DPIA and prepares it.

Writing the policy in twelve sections

  1. Scope: who is bound (partners, associates, trainees, staff, contractors).
  2. Principles: freedom of use, professional judgement, intellectual property.
  3. Authorised tools: a list kept by an AI referent, review of the terms, excluded tools, review at least once a year.
  4. Data and professional secrecy: closed environment, pseudonymisation, data allowed per level of tool, separation of matters.
  5. Verification: existence of the sources cited, review by a lawyer, no reuse without checking, internal labelling of AI-assisted work.
  6. Transparency: information to clients, client instructions, chatbot, published content.
  7. The lawyer's responsibility.
  8. Training and AI literacy.
  9. Security: business accounts, multi-factor authentication, devices.
  10. Incidents: reporting to the referent, personal data breach notified to the APD within 72 hours.
  11. Data protection: register, DPIA, processing agreement, transfers, lawfulness of the model.
  12. Control, adoption and review: checks, breaches, acknowledgement, review.

Common mistakes and open points

  • Thinking every client must be told that AI is used. It is not a general duty; the privacy notice describes the processing.
  • Forgetting AI features built into translation, word-processing or PDF software: they are tools like any other.
  • Relying on case law cited by AI without checking it. The Antwerp Court of Appeal awarded an increased procedural indemnity against a party whose AI-drafted submissions made the proceedings chaotic (25 March 2026, 2025AR774).
  • Open points: the exact day the guidelines were adopted is not confirmed; Belgium had not yet designated its AI Act supervisory authority in August 2026; no rule governs billing for AI-assisted work.

The tools assist the lawyer, who remains responsible for the policy adopted and how it is applied.

Sources

A drafting aid based on the sources cited. To be reviewed and adapted by the lawyer; it is neither advice nor a guarantee of compliance.