LGPD International Data Transfer and LLM Providers Abroad

LGPD International Data Transfer and LLM Providers Abroad

In most cases, yes. When a company in Brazil sends personal data to an AI provider that processes it in another country, that is an international data transfer under LGPD (art. 5, XV, and arts. 33 to 36). This does not prohibit the use. It does require a legal basis from art. 33, an adequate contract with the provider and control over what gets sent.

Key takeaways

• Sending a name, phone number, email or the text of a conversation to an AI API hosted outside Brazil is usually an international transfer.
• The transfer is only allowed if one of the art. 33 grounds applies, such as contractual clauses recognized by ANPD, a country with an adequate level of protection, or specific and prominent consent.
• The most common path for companies is standard contractual clauses under the ANPD regulation, not consent.
• Before negotiating a contract, reduce what leaves your systems. Masked data or data without personal identifiers simplifies the problem considerably.
• Responsibility stays with the company that decides to send the data, even when the provider is large and well known.

What LGPD calls an international transfer

LGPD defines an international transfer as the transfer of personal data to a foreign country or to an international organization of which Brazil is a member. The concept is deliberately broad. It does not matter whether the data moves as a file, an API call or a system integration: if personal data starts being processed by an agent abroad, the rule applies.

The LGPD chapter on the subject (arts. 33 to 36) says a transfer can only happen under listed grounds. Among them are: countries or organizations with an adequate level of data protection, recognized by ANPD; safeguards offered by the controller through specific contractual clauses, standard contractual clauses, global corporate rules, or seals, certificates and codes of conduct; and the data subject's specific and prominent consent, given with prior information about the international nature of the operation. There are also grounds tied to legal obligation, contract performance and the regular exercise of rights, which fit narrower situations.

ANPD has issued its own regulation on international transfers, which details the standard clauses and the procedure. The official text and updates are published on the ANPD website. Read the version currently in force before you settle on a decision, because administrative practice keeps evolving.

When AI use falls under this rule

What follows is a practical reading of the legal text, not legal advice. For high-risk cases, consult a lawyer who specializes in data protection.

AI use usually falls under the rule when:

• a support agent sends the customer's message, with a name or phone number, to a language model API hosted outside Brazil;
• an employee pastes a spreadsheet, a contract or an email containing customer data into a public AI chat;
• a meeting transcription tool sends audio to servers abroad (see recording meetings with AI and LGPD);
• an automation or agent platform stores conversation logs in a cloud located in another country.

It tends not to fall under the rule when:

• the data sent is not personal, such as a generic technical text with no identification;
• the data has genuinely been anonymized, so the person cannot be identified by reasonable means (art. 12);
• the model runs inside your own infrastructure in Brazil, with nothing sent to third parties abroad.

There are gray areas. A provider may store data in a Brazilian region but have support teams or subprocessors in other countries with possible access. Check this in the contract and the provider's documentation rather than assuming it. The same caution applies to data that passes through an intermediary (the support platform) before reaching the model provider: every link in the chain counts.

The ANPD regulation also distinguishes a transfer from what it calls international collection, where the agent abroad collects the data directly from the data subject. If your customer talks directly to a foreign service, the legal design changes. When your company receives the message and passes it to the model, the scenario is a transfer.

What LGPD and ANPD expect from the company

The controller, generally the company that decides to use the AI, needs to demonstrate several points.

• A legal basis for the processing itself. Before discussing transfer, you need a legal basis under art. 7 (or art. 11 for sensitive data) to use the data for that purpose.
• A transfer ground under art. 33. In business practice this is usually standard contractual clauses or another contractual safeguard provided by the law.
• Transparency. Data subjects must be told, in the privacy policy and during the interaction, that data may be processed by providers abroad and for what purpose.
• The principles in art. 6. Purpose, adequacy and necessity mean sending only what the task requires. Security and prevention call for proportionate technical measures.
• A record of processing operations (art. 37). Keep a map of what is processed, where and by whom.
• Accountability. Art. 42 provides for compensation for damages. Hiring a provider does not shift the company's responsibility toward the data subject and ANPD.

If the AI agent makes decisions that affect the data subject, such as credit analysis or screening, add the reading on automated decisions and art. 20. For the broader regulatory picture, follow the Brazil AI regulation overview for companies.

Which contract clauses and terms to check with the provider

When reviewing the provider's contract or data processing addendum (DPA), confirm the items below. A missing item does not mean the use is illegal, but it flags a question that needs an answer before the pilot.

• Roles of the parties: does the provider act as a processor, handling data on behalf of your company, or as its own controller for some purposes?
• Use for training: is data sent through the API or the business plan used to train models? Is that off by default or does it depend on configuration?
• Retention: how long are prompts and responses stored, and can retention be reduced or disabled?
• Location and access: in which countries is data processed and stored, and from where can support teams access it?
• Subprocessors: is there a published list, with notice of changes and a right to object?
• Standard clauses: does the contract incorporate the standard clauses required by ANPD or provide another art. 33 safeguard? If the text is the standard clauses, its content should not be altered.
• Security: encryption, access control, access logging and the certifications the provider claims, with documentation available.
• Incidents: deadline and method for notifying your company in case of a breach, so you can meet your own duties to ANPD and data subjects.
• Data subject rights: how the provider helps you handle requests for access, correction and deletion.
• Return and deletion: what happens to the data when the contract ends.
• Audit: what evidence the provider supplies that it does what it promised.

Treat the contract and the provider's official documentation as the primary source. Marketing pages and answers from an AI assistant about its own provider do not replace the contract text.

Illustrative scenario: WhatsApp support with a model abroad

This scenario is illustrative. A services company receives requests for duplicate invoices and appointments on WhatsApp. It wants an agent that understands the message, checks the internal system and replies.

The initial design sends the whole conversation to the model, including full name, phone number and address, and stores logs on the vendor's platform. While mapping the flow, the team realizes the model only needs the intent (duplicate invoice, appointment, question) and the text of the question. Name, phone number and address can stay in the internal system, which assembles the final reply.

After the adjustment, the message is masked before sending, the model returns the intent and a suggested text, and the local system fills in the personal data. What still leaves Brazil is smaller and less identifiable. Even so, because some messages may contain personal data typed by the customer, the company keeps the contract with art. 33 safeguards, an updated privacy policy and minimal log retention. For the channel side, see how to integrate AI into your company's WhatsApp.

A step-by-step plan to decide safely

The method we apply in AI projects follows a sequence: map, limit, pilot, deliver. For international transfer, it translates like this.

• Map: list every flow that uses AI, which personal data goes in, which provider receives it, in which country it processes the data and who has access. Include informal use of AI chats by employees.
• Limit: define the minimum data needed for each task. Mask or replace identifiers before sending, block sensitive categories (health, biometrics, children's data) in flows that do not need them, and restrict what the agent can do.
• Choose the safeguard: with your legal team, define the applicable art. 33 ground and check the provider's contract against the clause list above.
• Pilot: run one small, real flow with logs of what was sent, an agreed success metric and human review of borderline cases.
• Deliver and review: document decisions, update the privacy policy and the record of operations, and define when to reassess (provider change, new feature, contract change). At the end of the pilot, decide whether to expand, adjust or stop.

For structuring risk more broadly, the NIST AI Risk Management Framework offers useful governance vocabulary, and the OWASP Top 10 for LLM applications helps review risks such as sensitive information disclosure. Neither replaces the LGPD analysis, but both complement the technical side.

How to reduce what leaves Brazil

The best way to simplify compliance is to send less. Some measures, in order of effort:

• A clear internal policy on what can and cannot be pasted into AI tools, as described in the company policy for employee ChatGPT use.
• Masking or pseudonymization before the API call, with re-identification done inside your own system. Remember that pseudonymized data can still be personal data.
• Business accounts with reduced retention and no use of your data for training, when the provider offers that setting and the contract guarantees it.
• Models run on your own infrastructure or in a cloud in Brazil for the most sensitive flows, which calls for careful design, as in running Ollama local models in a company.
• Separating tasks: use a model abroad only for content with no personal data and keep the rest in a controlled environment.

None of these measures removes the need to assess the specific case, but all of them shrink the risk surface and make the contract less critical.

Frequently asked questions

Is using ChatGPT or another AI with customer data always an international transfer?

Not necessarily. If what you send contains no personal data, or if the data has been effectively anonymized, the transfer rule does not apply. If it includes a name, phone number, email or content that identifies someone, and the provider processes it abroad, the rule tends to apply.

Do I need the customer's consent to send data to an AI abroad?

Generally that is not the main path. Specific and prominent consent is one of the art. 33 grounds, but it tends to be fragile because it can be withdrawn and must be free and informed. Companies usually prefer contractual safeguards, such as standard clauses, as long as there is also a legal basis for the processing itself.

If the provider stores data on a server in Brazil, does the problem go away?

It shrinks, but it does not disappear automatically. You need to confirm whether processing, support or subprocessor access happens in other countries, since that can amount to a transfer. The answer is in the contract, the subprocessor list and the provider's documentation.

Is the provider responsible if something goes wrong?

No. The company that decides to send the data remains responsible to data subjects and ANPD, and LGPD provides for liability of both controller and processor depending on the case. That is why it is worth documenting the choice of provider, the clauses you verified and the data limitation measures.

Sources

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