LGPD and Superintelligence: What Brazilian Companies Should Prepare
Brazilian companies planning AI agents under the LGPD should prepare agent memory, data minimization and access records before go-live.

Brazilian companies that plan to run AI agents under the LGPD should prepare three things before any agent goes live: an inventory of what the agent can remember, a rule that limits personal data to what each task needs, and access records that show who or what touched personal data and for what purpose. These controls matter more as agents receive wider tools and longer memory, because the legal duties attach to personal data wherever it flows, not to how capable the system is. The sections below set out what to build now, in what order, and how to keep the records useful to an owner and to a regulator.
Key takeaways
- Inventory what an agent can remember, where that memory is stored and who can read it before the first live run.
- Apply purpose and necessity limits to every personal data field an agent receives, not only to its prompts.
- Keep access records that name the actor, the data category, the purpose and the outcome, without copying personal data into the logs.
- Set retention and deletion rules for memory, logs and transcripts, and test that deletion actually works.
- Name one person who answers LGPD questions about the agent and reviews any decision made mainly by automated processing.
Why more capable agents raise the LGPD stakes
The LGPD governs the processing of personal data, and it does not depend on which model produced the output. An agent that reads a customer message, looks up an order, saves a preference and drafts a reply is processing personal data at each step, even when every step looks routine. As agents gain access to more systems and become better at connecting scattered facts, the set of personal data they touch tends to widen. Whether or not systems eventually reach the level people call superintelligence, a company that builds data controls now will have less to unwind later.
The NIST AI Risk Management Framework offers a useful working vocabulary: govern, map, measure and manage. Governing means someone owns the decision. Mapping means you know where data enters and leaves the agent. Measuring means you can check whether the controls work, and managing means you act when they do not. Use those four verbs as the agenda for your first review.
What the agent remembers, and why it matters under the LGPD
Agent memory is rarely one thing. Usually there are at least four layers: the context of the current conversation, notes the agent saves for later sessions, documents it retrieves from a knowledge base, and any stored summaries or embeddings that a retrieval system keeps. Each layer is a place where personal data can sit, and each one needs a purpose, a retention period and a list of who can read it.
Start with a memory inventory. For every layer, write what it contains, which data subjects appear in it, why the agent needs it, how long it is kept and which people or systems can read it. Many companies discover at this stage that a memory built for convenience holds data that no task requires. That is the cheapest finding of the whole exercise, because it can be removed before launch.
Retention is where the LGPD asks for discipline. Personal data should be kept only for the period needed to fulfil its purpose, then deleted or anonymized, subject to legal exceptions that counsel must identify. Agent memory is often left out of retention schedules because it does not sit in the database the company already manages. Put it in the schedule explicitly. A practical method for finding and removing one person's data from agent memory is described in a guide to deleting agent memory on request.
Data minimization: share less with the agent and keep less of what it sees
The LGPD's principle of necessity asks that processing be limited to what is needed for the stated purpose. Applied to agents, the question is not whether the model could use a field, but whether the task requires it. A refund assistant does not need a customer's full address history, and a scheduling agent does not need the text of past complaints. Limiting inputs at the tool layer is more reliable than asking the model to ignore data it has already received.
- Give each tool only the fields its task needs, and return only the result, not the whole record.
- Mask identifiers such as national ID numbers, phone numbers and email addresses wherever the task allows it.
- Summarize long histories into the facts the task requires, and drop the rest before it reaches memory.
- Keep test data separate from production data, and never paste real customer records into prompts used for evaluation.
The same discipline applies to logs. Logs that copy full prompts turn the audit trail into a second copy of the personal data you were trying to limit. For the detailed approach, see how to redact personal data from agent logs, and use the data minimization guide for AI agents as a checklist for each tool.
Access records you can actually use
Article 37 of the LGPD asks controllers and operators to keep records of their personal data processing, and an access record is the evidence that makes those records credible. For an agent, an access record should answer five questions: which actor touched the data (a named staff member, a service account or the agent's own identity), which category of data was involved, when it happened, for what purpose, and what the outcome was. A log that only says the agent ran is not an access record.
Use structured events rather than free text. The OpenTelemetry observability primer describes traces, metrics and logs as the building blocks of telemetry, and the same design works here: each tool call becomes an event with an actor, a data category, a purpose and a result. Keep these events separate from raw conversation content, and restrict who can read that raw content.
Decide who can read the records and for how long. An access record that contains identifiers is itself personal data, so it needs its own retention period, set with counsel and tied to the audit needs you already have.
A preparation plan, with an illustrative scenario
- Step 1: Write the agent's purpose in one sentence and list every personal data category the role needs.
- Step 2: Inventory each memory layer, tool and channel, with an owner, a retention period and a list of readers.
- Step 3: Remove from tool inputs and memory writes every field the task does not need.
- Step 4: Define the access record fields and redaction rules, then test them on sample interactions.
- Step 5: Set retention and deletion rules for memory, logs and transcripts, and run one deletion test end to end.
- Step 6: Name the person who handles data subject requests and reviews automated decisions, with a written path to the company's encarregado (data protection officer) where one is appointed.
- Step 7: Set a go-live condition: no launch until steps 1 to 6 have written evidence behind them.
Illustrative scenario, not a client case: a regional furniture retailer wants an agent to answer delivery questions on WhatsApp. In the first draft, the agent saves full conversation transcripts to a long-term notes store, copies customer addresses into every tool call and logs complete prompts. Reviewed against the plan, the delivery tool receives only the order number and the delivery zone, the notes store keeps only a preferred delivery window, and the logs record the tool, the data category and the outcome with addresses redacted. The deletion test then confirms that a customer's notes disappear from both the store and the logs.
Who keeps the controls working after launch
Controls drift. Tools gain new permissions, memory accumulates and staff change roles. A checklist written once before launch is not enough, so the controls need a schedule: periodic permission reviews, recurring evaluations of what the agent stores and sends, and a written response for incidents. Under the LGPD, a security incident that may cause relevant risk to people must be communicated to the ANPD and to the affected individuals, so the incident path has to exist before it is needed. The quarterly permission audit for AI agents is a practical starting point for the review cycle, and the ANPD publishes guidance on its own portal.
Be Human designs and operates this layer as part of a managed engagement. The first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then cover monitoring, recurring evaluations, permission review and incident response, so the memory inventory and access records are maintained rather than written once and forgotten. This is a working arrangement with a defined scope, not a guarantee of compliance, and legal interpretation remains with the company's counsel.
Choosing where to start
Most companies should start with one agent, one data flow and one set of records, not with every process at once. Choose the task where the personal data is most predictable and where a mistake is easiest to detect and reverse. If you are not sure which task that is, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Does the LGPD apply to AI agents that run on foreign models? The LGPD applies to processing carried out in Brazil, to data collected in Brazil, and to processing that offers goods or services to people in Brazil, regardless of where the model is hosted. The provider's location changes which contracts and transfer safeguards you need, but it does not remove your duties as controller or operator. Ask counsel to review the vendor terms before any personal data reaches the model.
Do I have to delete everything an agent remembers when a customer asks? Not automatically. The right to deletion applies to personal data processed on the basis of consent or where the purpose has ended, and some records may need to be kept for legal reasons. What you must be able to do is find that customer's data across memory, logs and notes, decide what applies, and document the outcome.
Is a full prompt log the same as an access record? No. A full prompt log often contains more personal data than the audit needs, which makes it a liability in its own right. An access record names the actor, the data category, the purpose and the outcome, with sensitive content redacted or kept in a separate, restricted store.
Who should own LGPD questions for an AI agent? A named person, usually from legal, compliance or operations, should own them, with the encarregado involved where one is appointed. That owner approves the purpose, reviews automated decisions when a person requests a review, and signs off on retention. Write the ownership down, so that decisions do not depend on whoever happens to be available.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


