
LGPD Automated Decision: The Right to Review and When to Require It
When an AI-only decision under LGPD Article 20 requires human review, how a data subject requests a review, and what to log so the process can be proven.
Want to apply this in your company?
Book a call with Juan →A decision made solely on automated processing of personal data, if it affects a person's interests, can be challenged by the data subject, and the company must offer a review. Article 20 of the LGPD (Brazil's data protection law) grants this right. In practice, your support process needs a clear path for the request, a person with authority to reassess the case, and a record showing what was reviewed, by whom, and with what result.
Key takeaways
- Not every automated answer is an automated decision. The trigger is an outcome that affects the data subject and was set without meaningful human involvement.
- The data subject can ask for a review, and the company should offer a channel that works in the same place where the decision was communicated.
- The review must be carried out by someone with access to the data, the criteria, and the power to change the outcome.
- The request should be documented with the date, the criteria applied, the data used, the person who analyzed it, and the response sent.
- An AI agent can triage, explain, and collect context, but it should not close on its own a refusal that has been challenged.
What Article 20 of the LGPD requires
Law No. 13,709/2018 recognizes, in Article 20, the right to request a review of decisions made solely on the basis of automated processing of personal data that affect the data subject's interests. The text cites, as examples, decisions intended to define a personal, professional, consumer, or credit profile, or aspects of personality. The law also requires the controller to provide clear and adequate information about the criteria and procedures used, subject to commercial and industrial secrecy.
Read the official text yourself rather than relying only on summaries, including those from other websites and tools. The Brazilian National Data Protection Authority (ANPD) follows this topic, and your legal team should check the authority's official pages before finalizing any internal policy. For the broader regulatory context on AI systems, read the material on Brazil AI regulation: what companies should watch and do now. For the details of Article 20, see LGPD automated decisions: Article 20 and the right to review.
This right does not prohibit a company from using AI to make decisions. It requires that the company know where a decision was automated, can explain the criteria in plain language, and has a way to review the result when the data subject asks.
When a decision requires human review
Before automating a decision, work through these questions in order:
- Does the decision change something concrete for the person, such as an approval or refusal, an individual price, an access block, a charge, a cancellation, eligibility for a benefit, or a risk classification?
- Does the decision reach the customer without anyone analyzing it first, or does a person only confirm what the AI has already determined?
- Is the data personal, such as payment history, purchases, registration details, browsing behavior, or conversations?
- Can the customer reverse the outcome through support on their own, or does it depend on someone with the authority to change it?
- If the answer is yes to two or more of these questions, treat the decision as a case that requires a human review path.
A review that only confirms the AI's result without a real examination is unlikely to meet the purpose of the article. The reviewer needs to see the data, understand the criteria applied, and have the authority to change the decision. If the team lacks this access, the problem lies in the process design, not in the person answering the customer.
How to document the request in support
A review request should become a registered case, not a conversation lost in the chat history. A simple plan for this:
- Identify the contested decision: which outcome was communicated, on what date, and through which channel.
- Record the data subject's request in their own words, without rewriting or summarizing it in a way that changes its meaning.
- Confirm the identity of the requester through the channel the company already uses, without collecting data beyond what is necessary.
- Attach the data and criteria used, or a reference to the decision record that keeps this information.
- Route the case to a person with authority to change the outcome, with a defined internal deadline.
- Record the conclusion, the justification, and the response sent to the data subject.
- Keep the history for as long as the internal policy and applicable obligations require, with access limited to those who need it.
With an agent on WhatsApp, the weakest point is usually the record. The conversation shows what was said, but not necessarily which rule or model produced the refusal, or which version of the instructions was active at that moment. The flow should store the decision in dedicated fields, including the date, the version of the instructions, the data consulted, and the result. This kind of decision record also helps the team find data errors, which are more common than failures in the model itself.
Illustrative scenario: a refused plan change
Illustrative scenario: an internet provider uses an agent on WhatsApp to handle plan changes. An automatic rule refuses the change for customers with a recent payment delay, and the message only states that the change is not possible at this time. The customer asks for a review. Without a record, the human agent cannot see the criterion, does not know which data was used, and replies that the system decided. With the flow described above, the case opens with the communicated outcome, the delay criterion, the date of the last payment, and the review queue. The reviewer finds that the payment cleared after the analysis, approves the change, and records the correction. The error was in the input data, and the record is what made it possible to find it.
Where the agent should stop and hand over
The agent can answer questions about the decision, explain the review procedure, and collect the information needed. It should not restate the refusal as final, promise an outcome, or try to persuade the customer to drop the request. A practical rule set:
- Explain what happened in a factual way, without jargon.
- Open the review request and state the internal deadline for a reply.
- Transfer to a person when the customer disputes the decision, asks for a review, or mentions the LGPD.
- Never use the refusal message as the final answer to a review request.
When your reply to the customer comes after some time has passed, check the platform's sending rules before responding, as explained in the article on the WhatsApp customer service window with AI. To decide which parts of support to automate and which stay with your team, the guide on AI customer support: what to automate and what to keep human offers complementary criteria.
The NIST AI Risk Management Framework organizes AI risk management into the functions of govern, map, measure, and manage. This separation helps a company assign responsibilities: who sets the policy, who maintains the flow, and who answers the review request.
How to apply the method in this area
The method we use has four steps: map the current process and its baseline, limit data and actions, pilot a real flow with records and a success metric agreed in advance, and then hand over the operation and decide whether to expand, adjust, or stop. Applied to the right of review, this means first listing which decisions are already made without a person, including simple refusal rules. Next, limit the agent to triage and explanation. Then pilot the review request in a single area, with a complete record, and expand only when the legal team and operations agree with the results.
Frequently asked questions
Does every decision made by AI give the right to human review? Not necessarily. Article 20 covers decisions made solely on the basis of automated processing that affect the data subject's interests. Information such as opening hours or the status of an order usually does not fall under it, but a credit refusal, an account block, or a denied plan change may. When in doubt, treat the decision as a case that requires a review path and consult your legal team.
Can a WhatsApp agent deny a review request? It should not close the request on its own. It can log the request, explain the procedure, and state the deadline, but the conclusion needs a person with authority to change the outcome. Denying the request through the agent itself is exactly the kind of failure that a proper record makes visible.
Do I need to reveal the algorithm to the customer? You do not need to expose code, models, or formulas. The LGPD asks for clear and adequate information about criteria and procedures, subject to commercial and industrial secrecy. In practice, the company can explain which types of data and which rules weighed on the decision, without disclosing the internal logic of the system.
Where should I start if my company has no defined process yet? Begin by listing every decision that is already made without a person, including simple rules, because they can also be automated decisions. Then define who performs the review, what record will be kept, and how the customer requests an analysis. Only after that should you configure the agent, starting with one type of decision and reviewing the results with your legal team before expanding.