Personalized AI infrastructure
Superintelligence Prep for Brazilian SMEs: Data, Language, Compliance
A practical starting order for Brazilian small companies: clean records, Portuguese memory, LGPD limits and owned prompts, before any agent gets access.

Superintelligence preparation for a small business in Brazil is less about a future system and more about the present state of your records, language and rules. A practical starting order is to clean the operational data an agent would read, write the company's memory in Portuguese, map personal data under the LGPD before anything is stored, and keep prompts and instructions as files the company owns. Each step is small on its own, and each one makes the next step cheaper and safer.
Key takeaways
- Clean the records a role will read before connecting any agent to them.
- Write company memory in Portuguese, with a short glossary for the terms your team really uses.
- Map personal data, purpose and retention with your legal adviser before memory stores it.
- Keep prompts, instructions and evaluation cases as company files you can export and version.
- Start with one role, read-only tools and one written approval boundary.
Why the order matters more than the tool
Most small companies meet this subject through product demonstrations, which show what a model can do with a clean example. Your company is probably not a clean example. The customer list has duplicates, the price rules live in three spreadsheets and a messaging app, and the refund practice your team follows differs from the one published on your website. A more capable model reads those contradictions faster and acts on them with more confidence, so the weak point moves from the model to the operation.
The NIST AI Risk Management Framework organizes AI work around governing, mapping, measuring and managing risk. For a business owner, its value is the discipline it asks for: name what the system is for, what data it touches and who answers when it gets something wrong. The sequence below applies that same discipline at the scale of a small company. If you want a wider view of what the first stage involves, Superintelligence Readiness for Brazilian Companies: Where to Start covers the early decisions in more depth.
Which records should an agent read first
Begin with a single role, not with the whole company. Write down the questions that role answers every week, then list the records needed to answer them. For a sales coordinator, that might be the product catalogue, the approved discount rules and the delivery calendar. Everything else stays out of scope until the first role works.
For each record, check four things:
- Who owns it by name, and who is allowed to change it.
- Whether it has one version or several copies that disagree.
- Whether it contains personal data the role does not need.
- Whether a person already checks it before a customer sees the result.
Illustrative scenario: a regional distributor wants an agent to answer order status questions. Its order data is accurate in the billing system but stale in a shared spreadsheet that the warehouse updates by hand. Before any connection, the team names the billing system as the single source, stops using the spreadsheet as a reference, and assigns one person to correct exceptions. The agent then reads one record instead of two. The naming of owners and sources is the kind of work described in Documentation Standards for a Superintelligence-Ready Company.
How to write memory in Portuguese
An agent's memory is the material it can retrieve when it answers: policies, product notes, approved replies and records of past decisions. In retrieval-based designs, the system searches those documents at answer time and writes its response from the passages it finds. If those passages use inconsistent terms, retrieval picks the wrong one, and the answer still sounds confident.
Brazilian operations carry their own vocabulary. A product has one name on the shop floor, another on the invoice and a third in the contract. Customers may use yet another word. Build a short Portuguese glossary that maps each working term to its official name, and keep it inside the memory so the agent reads it too. Write new memory notes in clear Portuguese with an owner and a date, so that a future reader can tell whether a rule is still in force.
Decide explicitly which language governs. Many teams use English for vendor documentation and Portuguese for daily work. A mixed set of documents with no rule about which version wins is a common source of contradictory answers, and it is easy to fix once someone writes the rule down.
What the LGPD requires before memory stores anything
The Lei Geral de Proteção de Dados, known as the LGPD, governs the processing of personal data in Brazil. An agent that stores a customer's name, purchase history or support conversation is processing personal data. The company that decides how that agent is used is expected to answer for that processing, so the decision belongs to the business owner, with counsel. The ANPD, the national data protection authority, publishes guidance that is worth reading alongside that advice.
Before memory stores anything about a person, answer these questions in writing:
- What specific purpose does keeping this information serve?
- Is the data needed for that purpose, or could the agent work with less?
- How long will it be kept, and what deletes it?
- How will the company answer a request from the person to see, correct or delete their data?
- Which vendor or hosting location will process it, and under what contract?
Treat these answers as living notes, not a one-time form. A purpose that fits a support role may not cover a marketing use someone adds later. Revisit the answers whenever the role gains a new tool or a new data source.
Owned prompts, permissions and approval boundaries
A prompt is an instruction, and an instruction is a business decision. The company should own the prompts that shape how its agent speaks, what it declines to do and how it hands a case to a person. Store them as versioned files next to the approved answers and the evaluation cases, in a format you can read without the vendor's console. When a prompt changes, record what changed and why. This is also what keeps the company free to move to another provider later.
Evaluation cases are the most underrated asset. These are real questions with the answer the company would accept. They show whether a change made the agent better or quietly worse. Keep them small and specific, and rerun them after every change.
Give the agent the narrowest access that lets it do its job. The OWASP authorization guidance describes the general principle: decide access by policy, check it on every request and deny by default. For a business agent, this usually means read access first, write access only where a person approves the result, and no ability to change customer terms or move money without an explicit rule. The quarterly permission review is where those rules get checked against what the agent can actually do.
Illustrative scenario: a two-location clinic network wants an agent that drafts appointment reminders. The agent can read the schedule and draft messages. It cannot mention prices in a message, and it cannot cancel a booking. A staff member approves anything that changes an appointment. The boundary fits on one page, and the first review is scheduled before launch rather than after the first complaint.
A starting order for the next quarter
Work through these steps in sequence. Each one is a decision you can write down and hand to someone who will carry it out.
- Name one business owner for the role and one person who checks its answers.
- List the records the role reads and the owner of each record.
- Remove duplicates and resolve contradictions in those records.
- Write the Portuguese glossary and the approved answers to the most frequent questions.
- Map personal data, purpose and retention with counsel, in writing.
- Connect read-only tools first, and document every write action that needs approval.
- Build a small evaluation set from real questions and agree on one success criterion.
- Schedule the first permission review and keep prompts and files in company storage.
Be Human designs this layer before the agent receives broad access. A first engagement defines one useful role, the memory and tools it needs, its approval boundaries and an observable success criterion. Managed operations then cover monitoring, recurring evaluations, integration maintenance, permission review, incident response and controlled improvements, with the agent's identity, memory and permissions kept under the company's control. The service is remote and available in English and Portuguese.
Choosing the first task
The best first task is usually not the most impressive one. Look for work that happens often, has a clear owner, carries limited risk if an answer is wrong, and produces a result you can check against a known standard. Order status questions, appointment confirmations and internal policy lookups tend to fit that profile better than pricing decisions or contract negotiations. Once a task is chosen, the rest of the order above tells you what to prepare.
If you are not sure which of these steps your company should take first, the free Be Human AI diagnostic can help you find the first task worth handing to an agent.
Frequently asked questions
Is superintelligence preparation something a small business in Brazil needs to start now? It is not a separate project with its own budget line. The useful work is the same as good operations: records that agree, a shared language for the team and clear permissions. Starting there makes any later capability easier to adopt and easier to stop if it does not help.
Do we need to hire a data scientist before the first agent? Not for the first role. What the first role needs is an owner who knows the process, a person who checks the answers and a written boundary for what the agent may do. Outside help can support the setup and the periodic review without becoming a permanent dependency.
Does the LGPD prevent a company from using AI agents? No. The law sets conditions on why personal data is processed, how much is needed, how long it is kept and how people can exercise their rights. Document those conditions for each role, review them with counsel and update them when the agent gains new data sources.
Should prompts and instructions live in our own files or only inside the vendor's tool? Keep a copy in company storage, versioned and exportable, even if the vendor's tool is where the agent runs. That record lets you explain past behavior, compare changes and move to another provider without rebuilding the company's operating knowledge from memory.
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.


