Brazilian Bar Association AI Guidelines for Lawyers: What to Do

Brazilian Bar Association AI Guidelines for Lawyers: What to Do

The Brazilian Bar Association (OAB) does not ban generative AI in legal practice, but the recommendations from its Federal Council treat the tool as support, never as a substitute for the lawyer. In practice, that means protecting confidentiality and personal data, reviewing everything the machine produces, being transparent with the client, and keeping professional responsibility with the person who signs. This article turns the Brazilian bar association AI guidelines for lawyers into routines a firm can adopt without improvising.

Key takeaways

• Professional secrecy and data protection law (LGPD) decide what can and cannot go into an AI tool, and that matters more than convenience.
• Every AI output is a draft: citations, facts and legal arguments must be checked by a lawyer who answers for the document.
• Transparency with the client belongs in the engagement contract and in daily routine, not in last-minute explanations.
• Personal chatbot accounts used for firm work are the most common risk and the easiest to fix.
• A small pilot with written rules and a usage log works better than a blanket ban or a blanket green light.

What does the Brazilian bar association say about generative AI?

The OAB Federal Council has published recommendations for the use of generative AI in legal practice. They do not create a new law. They are ethical guidelines that rest on the Statute of the Legal Profession, the Code of Ethics and Discipline, and the LGPD, Brazil's General Data Protection Law. The themes that appear most strongly are the duty of confidentiality, protection of personal data, human supervision, the lawyer's technical competence, transparency with the client, and fees that are consistent with the work actually done.

A note on method: this article separates three layers. The first is official documentation, meaning what the OAB and the law say. The second is implementation experience, meaning how controls of this kind are usually built in organizations. The third is hypothesis, used when a point still depends on interpretation. Whenever the text moves away from the first layer, it says so.

Read the current document on the Federal Council's website before you finalize your firm's policy, because guidance can be revised. Also keep in mind that this article is not legal advice and does not replace a consultation with the Ethics Committee of your state section. For the broader regulatory context, it is worth following what is under discussion in the Brazil AI regulation framework.

Confidentiality and LGPD: what can go into the tool?

The duty of confidentiality is the most sensitive point. Pasting a petition, a client email or a case summary into a chatbot means sending confidential information to a third party. The right question is not whether the tool is good, but who the vendor is, how it handles data, whether there is a contract, whether content is used to train models, and how long it is stored.

Under the LGPD, the firm is a data processing agent for its clients and for the parties involved. That brings requirements of purpose, necessity and security. When an AI vendor processes that data, it becomes part of the processing chain. The ANPD, Brazil's national data protection authority, is the official reference for guidance on the subject.

A simple decision guide the firm can adapt:

• Public content (statute text, already published case law): can be used in an approved tool, with review afterward.
• Internal content without identification (petition structure, generic contract template): can be used in a tool approved by the firm.
• Anonymized case data: only in an approved tool, with a proper contract and with the anonymization checked by a person, because names and unusual details can identify someone even without a tax ID number.
• Identifiable personal data, sensitive data, sealed cases and strategic information: do not enter a tool without a defined legal basis, a proper contract and a decision by the responsible partner. In many cases the answer is not to use AI there, or to use a model running on infrastructure you control.

The classification above is an implementation practice, not OAB text. What the OAB states is the principle: confidentiality is not negotiable for the sake of convenience. To document the risk analysis when use involves personal data at scale, see also the article on the data protection impact assessment for AI projects.

How should you review what the AI produces?

Language models write fluently and confidently, even when they are wrong. They can cite a ruling that does not exist, attribute to an article content it does not have, or mix facts from different cases. Responsibility for the document belongs to the lawyer who signs it, which is why the OAB recommendations insist on human supervision.

A review checklist for filings and legal opinions that had AI support:

• Every citation of a statute, binding precedent or case was opened and checked at the primary source (the court, the official government legislation site, an official database), never in the AI itself.
• The facts of the case were checked against the case file and the documents, not against the generated summary.
• The argument and strategy were assessed by a licensed lawyer who agrees with the reasoning.
• Deadlines and calculations were redone outside the tool.
• The text was read in full to remove generic or repeated passages and anything out of the firm's tone.
• There is a record of who reviewed it and when.

For a more general review routine, the checklist for reviewing AI-generated text can serve as a base, adapted to the legal context.

There is an additional risk when the tool reads documents from third parties, such as the opposing party's filing or attachments received by email. Those files can contain hidden instructions that try to divert the model's behavior, which the OWASP list of risks for LLM applications calls prompt injection. This is technical documentation, not OAB guidance, but it reinforces the rule: the AI suggests, the person decides.

What should you tell the client about AI use?

The OAB recommendations point to transparency with the client as part of ethical practice. The logic is that the client entrusts the case to the lawyer and has a right to know, in general terms, how the work is done, especially when their data is processed by an external tool. The idea also appears that fees should reflect the work and value actually delivered.

How to turn that into practice (this is an implementation suggestion, not official text):

• Include in the engagement contract a short clause saying which types of AI tools may be used, for what purpose, and that every deliverable has human review.
• Explain in plain language that client data is only sent to tools approved by the firm.
• Address consent when use involves the client's sensitive data. Whether that requires specific consent or whether another legal basis is enough is a point to validate with specialist data protection counsel.
• If the firm uses a WhatsApp support agent, make clear that replies are automated and offer a path to reach a person.

A reasonable hypothesis, still without consensus: the more AI takes part in the work, the more explicit the conversation about fees and deliverables needs to be. It is worth deciding this in writing before the client asks.

Illustrative scenario: the draft written in a personal account

This is an illustrative scenario, unrelated to any real firm. An associate at a small labor law firm needs a draft reply quickly. He opens his personal chatbot account, pastes the initial complaint with the client's name, tax ID and health history, and asks for a summary and a draft. The tool returns a good text that cites a binding precedent. He reviews the wording but does not open the precedent.

There are three failures. Sensitive data was sent to an account with no contract and no firm control. A citation was not checked. And no record of the use was kept.

With the practices in this article, the same work would look different. The summary would come from an approved tool, with the data reduced to what is necessary. The draft would be treated as a draft. The precedent would be opened at the official source. The use would be logged in the firm's system, and the contract would already tell the client that this kind of support exists. The time saved remains, and the risk drops.

How to build the plan at your firm: map, limit, pilot, transfer

An AI policy that works usually starts from a small pilot, not a long document. The method we use in our consulting work follows four moves, applied here to the legal context:

• Map: list which AI tools are already in use, including personal accounts, and for which tasks (document summaries, research, drafting, client support). Note how long each task takes today, to have a baseline for comparison.
• Limit: define which data can go in, into which tools, and who approves exceptions. Restrict permissions and access, and require human review before anything is sent externally. The company policy for employee ChatGPT use shows how to structure what to allow and what to ban.
• Pilot: choose one real, low-risk workflow, for example summaries of public documents or first versions of internal templates. Record the use, the reviewer and the errors found, and agree in advance on what result justifies continuing.
• Transfer: at the end of the pilot, hand the rules, the records and the training to the team and decide whether use expands, is adjusted or is stopped.

This method works for any organization and does not replace the firm's own ethical and legal analysis. When there is managed operation, it is delivered through Be Human, with the same limits on data, permissions and human approval.

What stays with the lawyer?

Even with good tools and good routines, some decisions cannot be delegated. The OAB treats AI as an instrument, and professional judgment stays with the person who is registered and answers to the client and to the Bar:

• Case strategy and the choice of legal arguments.
• Signing and filing documents.
• Communicating relevant decisions to the client.
• Assessing conflicts of interest and ethical risks.
• Deciding when not to use AI.

The article on what not to delegate to AI in business helps think through this boundary in other contexts. For a general risk management structure, the NIST AI Risk Management Framework offers useful vocabulary for governing, mapping, measuring and managing the risks of AI systems.

Frequently asked questions

Does the OAB prohibit lawyers from using AI? No. The recommendations treat generative AI as a tool that can be used, provided confidentiality, data protection, human supervision and transparency are respected. Responsibility for the work remains with the lawyer.

Can I use ChatGPT with client data? It depends on the tool, the plan you have, the settings and the type of data. Personal and free accounts, with no contract and no retention controls, are not suitable for confidential information. Identifiable or sensitive data requires an approved tool, a legal basis and a formal decision by the firm.

Do I need to tell the client I used AI? The OAB recommendations value transparency, and the safest way to meet it is to provide for AI use in the engagement contract and explain in plain language how data is handled. If the case involves sensitive data, it is worth validating with specialist counsel whether specific consent is needed.

Who is responsible if the AI gets it wrong? The lawyer who signs and delivers the work. An invented citation or a swapped fact does not stop being a professional error because it came from a tool. That is why human review, with checks at the primary source and a record of who reviewed, is the center of any AI policy for law firms.

Sources

• ANPD: Autoridade Nacional de Proteção de Dados
• NIST AI Risk Management Framework
• OWASP Top 10 for LLM Applications