LGPD Compliance for AI on WhatsApp: Legal Basis, Consent, Records

LGPD Compliance for AI on WhatsApp: Legal Basis, Consent, Records

To use AI on WhatsApp in line with LGPD (Brazil's General Data Protection Law), a company needs to settle three things before the first customer conversation: which legal basis supports each purpose, what notice the customer sees, and what record proves what was processed. Consent is not the automatic answer to everything. Many support conversations rest on contract performance or legitimate interest, and consent is kept for optional uses such as promotional messages.

This article organizes those decisions for the person who has to approve the project. It does not replace review by your legal team or your data protection officer (the encarregado under LGPD), who know your contracts and your history.

Key takeaways

• Consent is one legal basis under LGPD, not the only one. Choose the basis per purpose, not per channel.
• The customer must know they are talking to an AI, what their data is used for and how to reach a person.
• Sensitive data and data about children need extra care. Design the agent so it does not ask for them unless it must.
• Keep useful records: purpose, legal basis, notice version, conversation logs with a retention period, and handoff decisions.
• Start with one small flow with clear limits, and expand only after reviewing what the records show.

What changes when AI joins the WhatsApp conversation?

A human agent on WhatsApp already handles personal data: name, phone number, order history, sometimes documents and photos. When an AI takes over part of that conversation, three things change.

The first is the volume and speed of processing. The agent reads everything the customer writes, including what they should not have sent, such as an ID document or a health detail in a conversation about scheduling.

The second is the supplier chain. The message passes through the WhatsApp platform, the AI model provider, automation tools and internal systems. Each link is a potential processor or sub-processor, and some may process data outside Brazil. That calls for contract review and, where it applies, attention to the rules on international data transfers.

The third is predictability. A trained person knows when to stop and escalate. An agent only knows that if someone defined the limits. LGPD principles such as purpose, necessity and transparency become design requirements, not just policy language.

For the technical side of connecting the channel, see How to Integrate AI into Your Company's WhatsApp. This article focuses on the legal and operational decisions that come first.

Is consent always required? How to choose a legal basis

No. LGPD lists several legal bases in Article 7, and the right one depends on what the agent does with the data. Consent must be free, informed and unambiguous, and the customer can withdraw it. If your company could not keep serving the customer after consent was withdrawn, consent is probably the wrong basis.

A simple decision guide, by purpose:

• Answering questions about an order, delivery, appointment or ongoing contract: usually contract performance, or preliminary steps toward a contract.
• Meeting an obligation the company has under law or regulation, such as providing mandatory information: legal or regulatory obligation.
• Support triage, flow improvement and fraud prevention without invasive use: possibly legitimate interest, backed by a documented balancing test.
• Sending offers, campaigns and promotional content: generally consent, with a simple way to opt out.
• Credit analysis or debt collection: assess credit protection or another basis with legal support, because the risk of challenge is higher.
• Any new purpose that appears later, such as using old conversations to train or tune a model: treat it as a separate purpose and reassess the basis, the notice and the provider contract.

The practical point: record one basis per purpose in a single table. If two purposes use the same data but rely on different bases, the agent and the process must respect that difference. For example, a phone number used to confirm a delivery does not, on its own, authorize a marketing campaign.

Legitimate interest is not a shortcut. It requires checking whether the customer would expect the use, whether a less invasive alternative exists, and whether the customer's rights override the company's interest. Document that assessment in one clear paragraph, signed by whoever makes the decision.

What does the customer notice need to say?

Transparency is what most clearly separates a solid project from an improvised one. LGPD gives data subjects easy access to clear information about how their data is processed. On WhatsApp, that means a short message at the start and a link to the full notice.

A checklist for the first message and the privacy notice:

• State that the conversation is handled by an artificial intelligence assistant.
• Say in plain language what the conversation data will be used for.
• Explain how to reach a person, and that this is possible at any time.
• Link to the full privacy notice, with the company's identification and the contact for the data protection officer.
• Explain in the full notice which suppliers take part in the processing and whether any processing happens outside Brazil.
• State how long conversations are kept and how to request access, correction or deletion.
• Ask for separate, optional consent for any use that depends on it, such as promotional messages.

Do not pile everything into a block of legal text in the first message. What the customer needs in order to decide whether to continue comes first, and short. The detail goes behind the link. Keep the version of the notice that was in force in each period, because it will be your evidence.

Sensitive data, automated decisions and human handoff

Data about health, racial or ethnic origin, religious belief, political opinion, sex life, genetic data and biometric data fall under a stricter regime in LGPD. Data about children and teenagers also requires specific treatment. In a support agent, the risk is usually not that the company wants this data. It is that customers send it on their own.

Three measures reduce that risk:

• Design the flow so it does not ask for this data, and instruct the agent to tell customers not to send it when it is not needed.
• Define what happens when it arrives anyway: route to a person, do not repeat the content, and apply the shortest retention that makes sense.
• Prevent the agent from receiving documents and images when the flow does not need them.

There is also the question of automated decisions. Article 20 of LGPD gives data subjects the right to request review of decisions made solely through automated processing that affect their interests. In practice, an agent that negotiates, denies refunds, classifies customers or sets support priority deserves attention. A safe rule to start with: the agent informs, organizes and routes, and a person decides anything with a significant effect on the customer.

That is why human handoff is a governance requirement, not only a customer experience feature. The playbook on WhatsApp AI Agent Human Handoff covers triggers and context transfer. To define which actions the agent never takes alone, use the AI Agent Approval Policy.

A related technical risk: customer messages are untrusted input. Instructions hidden in a message can try to make the agent reveal another customer's data or take actions it should not. The OWASP list of risks for applications built on language models is a good reference for your technical team to review this point.

What records should the company keep?

Article 37 of LGPD requires controllers and processors to keep a record of their processing operations. For a WhatsApp agent, that translates into a few concrete artifacts, which can live in a single document at the start:

• Purpose map: each purpose, the data used, the chosen legal basis, the internal owner and the date of the last review.
• Versions of the privacy notice and the opening message, with the dates each one was in force.
• Record of optional consents: who consented, for which purpose, when, and how to withdraw.
• List of suppliers and processors, with the type of data each receives and where it is processed.
• Conversation logs with a defined retention period, access control and a deletion rule.
• Agent decision log: when it answered, when it declined, when it escalated and why.
• Record of incidents and data subject requests, with response deadline and outcome.

Brazil's data protection authority (ANPD) can require a data protection impact assessment in certain cases. Even when nobody requires one, a lean version is worth writing for the pilot: what the agent processes, what risks exist for the customer, and which measures reduce them. The NIST AI Risk Management Framework can help structure that conversation in a neutral way.

One caution about logs: they are useful for audit and improvement, but they are also an accumulation of personal data. Record what is necessary, mask what you can and define who may read them. Observability concepts such as traces and metrics help you build this without storing more text than needed.

Illustrative scenario and step-by-step plan

Illustrative scenario: an aesthetics clinic wants a WhatsApp agent to confirm appointments and answer questions about procedure preparation. During the pilot, a customer writes that she is pregnant and asks whether she can have a procedure. The agent is configured not to give opinions on health conditions, tells her it will pass the conversation to the team, and hands it off with a short summary. The record shows the trigger that fired, the retention applied to that conversation and who took over. The legal basis for the scheduling flow is contract performance. Promotional messages stay outside the agent and depend on separate consent. Decisions like these, made in advance, are what make the service defensible.

A staged plan to get there:

• Map: describe the current flow, which data moves through it, who has access and what customers ask most often.
• Define purposes and legal bases in a table, validated by legal counsel or the data protection officer.
• Limit: choose which data the agent may read, which actions it may take and what must always go to a person.
• Write the opening message and the privacy notice, and set up a separate, optional consent channel.
• Review supplier contracts and settings: retention, use of data for training, processing location and subcontractors.
• Pilot one real flow, with logs on, one agreed success metric and periodic review of conversations.
• Transfer to operations: assign owners, a review routine and criteria to expand, adjust or stop.

This is the method we follow in agent work: map, limit, pilot, transfer. It does not remove risk, but it forces the company to decide the limits before the agent operates. To structure the pilot, the AI Pilot Charter Template provides a charter that fits the privacy decisions above.

When reviewing the project, separate three kinds of statements: what is in the law and in official ANPD guidance, what comes from implementation experience, and what is still a hypothesis to test in the pilot. Mixing the three is the most common way to end up with a policy that looks complete and fails at the first challenge.

Frequently asked questions

Do I need consent for every conversation with the AI on WhatsApp? Not necessarily. Conversations tied to an order, a contract or an appointment usually rest on other legal bases, such as contract performance. Consent is more common for optional purposes such as marketing, and it should be requested separately and be easy to withdraw.

Can I use past conversations to improve or train the agent? It depends on the original purpose, the notice the customer received and what the model provider's contract allows. Treat it as a new purpose: assess the legal basis, inform the customer and check whether the provider uses the data to train its own models. When in doubt, start with minimized data or data anonymized with care.

Does the customer need to know they are talking to an AI? LGPD has no specific sentence on this, but the transparency principle and the right to clear information support telling them. In practice it is the better choice: it reduces complaints, sets expectations and makes it easier for the customer to ask for a person when they prefer one.

Is my company or the AI vendor responsible? In general, the company that decides what the data is used for is the controller and answers to the customer and to the ANPD. The vendor acts as a processor and has its own obligations, but the contract, the instructions and the choice of vendor remain your decisions. That is why it is worth reviewing the contract and recording who does what.

Sources

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