Skip to content
be-human.ai
← Back to the journal

Personalized AI infrastructure

Superintelligence for Brazilian Companies: Language, Data and Local Rules

Brazilian companies should set up bilingual memory, Brazil-based data location and LGPD duties now, so the company layer stays theirs as AI models change.

Be Human typographic cover with the title: Superintelligence for Brazilian Companies: Language, Data and Local Rules
Conceptual illustration created with AI · Be Human

Superintelligence for Brazilian companies means building a company-owned layer of memory, tools, permissions and operating context that stays under local control while the AI models underneath it improve. Brazilian firms should prepare three things now: bilingual memory that keeps Portuguese operating language and English technical terms aligned, a written decision about where personal and business data is stored and processed, and a map of the Brazilian legal duties, chiefly the LGPD, that attach to each agent action. Companies that settle these three before connecting an agent keep their context when they change models.

Key takeaways

  • Keep company memory in Portuguese as the source of truth, with an English glossary that tools and models can rely on.
  • Decide where each kind of data sits (memory, index, logs, backups, model processing) before any agent touches customer records.
  • Map LGPD duties to each agent role, not only to the software vendor, and have counsel confirm transfers outside Brazil.
  • Keep memory, tool definitions, permissions and test cases in formats you can export, so a model change never means starting over.
  • Begin with one role, one success criterion and human approval for anything external, then widen only with evidence.

Why language comes first

Most Brazilian operations run in Portuguese at the counter, on the phone, in WhatsApp threads and in the back office. Vendor documentation, software interfaces and much of the technical literature arrive in English. An agent that reads a customer note saying a nota ficou pendente and a finance rule written as pending invoice needs a shared meaning, or it will act on the wrong one. Bilingual memory is therefore not a full translation of everything the company knows. It is a controlled vocabulary the company owns.

Language also carries legal weight. The words a company uses for consent, cancellation, complaint, retention and refund determine what an agent may do with a case. Define these terms once in Portuguese, since that is the language of your customers and of most of your contracts, and map each English equivalent to the Portuguese entry rather than letting the two drift apart over time.

How to build bilingual memory in four steps

  • Step 1: List the short set of terms that drive your most frequent decisions, such as order status, refund, consent, complaint, renewal and escalation.
  • Step 2: For each term, record the Portuguese term your team uses, the English gloss used in tools and vendor material, and the words that must not be used as synonyms.
  • Step 3: Store source documents in the language they were written in, with metadata that names the language and links each document to its glossary entries, so retrieval can find both versions.
  • Step 4: Test the agent with real Portuguese messages from customers and staff, including informal phrasing and regional variants, and record every case where it picked the wrong meaning.

This structure follows the retrieval pattern described in Google Cloud's architecture guidance for retrieval-augmented generation, where a model answers from documents the company retrieves at the moment of the question. The quality of that retrieval depends on the quality of the vocabulary behind it. A clean glossary is cheaper to maintain than a clever prompt, and it survives a change of model.

Decide where data lives before anything connects

Data location is a set of decisions, not one decision. The memory store, the search index, conversation logs, tool outputs, backups and the processing region of the model provider may each sit in a different place. Write down each location, the party that operates it and whether personal data passes through it. Most companies find that logs are the surprise, because traces of prompts and outputs often contain full customer messages.

  • Memory store: where summaries, documents and approved facts are kept
  • Retrieval index: where document embeddings or search entries are held
  • Logs and traces: where prompts, tool calls and outputs are recorded for review
  • Backups and restore copies: where recoverable copies of the above are kept
  • Model processing: where prompts are sent and whether the provider retains them

The LGPD (Lei Geral de Proteção de Dados) governs the processing of personal data, including transfers outside Brazil, and the ANPD (Autoridade Nacional de Proteção de Dados) publishes guidance and enforces the law. Read the ANPD material directly and ask counsel which transfer mechanism applies to each provider outside Brazil. Once the locations are written down, Draw Your AI Stack: A One-Page Architecture Map for Leaders is a practical way to show them on one page for the people who must approve them.

Which local legal duties attach to each agent action

The duties follow the action, not the vendor. An agent that answers a customer about their own records processes personal data. An agent that drafts a collection message, stores a complaint or summarizes a matter involving an employee carries different risks. For each role, list the duties it triggers: the legal basis and purpose, data minimization, handling of access and correction requests, deletion and retention, and the response to a security incident. The NIST AI Risk Management Framework offers a useful structure for naming these risks and assigning owners to them.

Illustrative scenario: a regional equipment rental company gives an agent read access to customer history so it can answer billing questions in Portuguese. The agent may read invoices and payment status but cannot change them. Its memory keeps only what open tickets require, and a deletion request removes the person's data from memory, index and logs on a fixed schedule. During setup the team finds that its logs keep full message text, including national tax identifiers, so it changes the logging rule before the agent goes live.

Keeping the company layer yours when models change

Models change faster than company processes. The durable assets are the glossary, the memory documents, the tool definitions, the permission rules and the test set of real cases with expected outcomes. If these live only inside one vendor's console, a model change means rebuilding them. Keep each one in a format you can export and version: plain documents, configuration files, and a dated record of who approved each change.

Write tool descriptions so they describe what a tool does and what it needs, without depending on one model's quirks. Before any new model reaches customers, run it against the same Portuguese cases and compare results. This is the practical meaning of owning the layer: the company decides when a model is good enough, and it can prove that decision with its own evidence.

Permissions and approvals for Brazilian operations

Permissions should mirror the roles your company already has: who may issue a refund, change a customer record or send a message to an external party. Start with human approval for every external message, payment and record change, and raise thresholds only after the agent has shown reliable results on real work. The step-by-step method in How to Run a Permission Review for Your AI Agents, Step by Step fits this stage well.

Customer messages, uploaded documents and web pages can contain text written to instruct the agent, which is known as prompt injection. The OWASP prompt injection guidance recommends treating retrieved content as data rather than as instructions and limiting what tools the agent can call. In Portuguese operations this matters because a customer can paste a long message that contains commands, and the agent should never treat those commands as the company's policy.

How to start this quarter

Begin with one role, one success criterion and one list of data locations and legal duties for that role. Superintelligence Readiness for Brazilian Companies: Where to Start walks through the first decisions in more detail. Be Human's managed approach follows the same order: it defines the role, builds the bilingual glossary and memory, records data locations and the duties each action triggers, and then operates the layer with monitoring, recurring evaluations, permission reviews and controlled changes. The company keeps ownership of its memory, permissions and records, and it can see what the agent reads, what it may do and what changed.

If you are unsure which first task is worth handing to an agent, the free Be Human AI diagnostic is designed to help you find it.

Frequently asked questions

What does superintelligence for Brazilian companies mean in practice? It describes a company-owned layer of memory, tools, permissions and operating context that stays under local control as frontier AI models improve. It is not a single product you buy. Preparing for it means organizing your language, your data locations and your legal duties now.

Does the LGPD apply to an internal AI agent? Yes, when the agent processes personal data of customers, suppliers, employees or other people, whether the tool is internal or external. Confirm the legal basis and purpose for each role with counsel, and keep records of what data the agent reads and stores. The obligations follow the data processing, not the label on the tool.

Must customer data stay in Brazil? Not in every case, because the LGPD allows international transfers under specific conditions. Decide location by data type and by provider, document the decision, and confirm the transfer mechanism with counsel before go-live. Write the decision down so that a future model or vendor change can be checked against it.

Should an agent's memory be kept in Portuguese or English? Keep the source record in Portuguese, since that is the language of your customers and most of your contracts, and maintain an English glossary for tools and vendor interfaces. The glossary is what lets the agent, your staff and any future model agree on what a term means.

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.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →