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

Personalized AI infrastructure

Preparing a Brazilian Company for Superintelligence: Data and Approvals

What Brazilian companies should organize first in data, Portuguese processes and approval rights so AI agents can benefit from stronger models later.

Be Human typographic cover with the title: Preparing a Brazilian Company for Superintelligence: Data and Approvals
Conceptual illustration created with AI · Be Human

Superintelligence for Brazilian companies is mostly a preparation problem, not a purchase. A company gets value from more capable AI models only when its agents can reach trustworthy business data, follow processes written in the language its teams really use, and act inside approval rights that a named person can review. Organizing those three things first lets an agent's scope grow as models improve, without rebuilding its memory, tools and permissions each time.

Key takeaways

  • Start from the decisions an agent would make, then trace the data each decision needs.
  • Identify the source of truth for each business record and handle personal data under a documented basis before connecting anything.
  • Write core processes in Portuguese, using the exact terms your staff use, and include the exception paths.
  • Keep every consequential action behind an approval right with a named owner and a readable log.
  • Treat preparation as a managed, reviewable layer that outlasts any single model choice.

What does preparing for superintelligence actually mean?

No one can say with confidence what the next generation of models will do well in your sector, and a plan that depends on a specific capability is fragile. What a leader can control is the layer around the model: the memory an agent is allowed to use, the tools it can call, the people who approve its actions, and the measures that show whether it is doing the job. When that layer is clear, a stronger model is an upgrade to a component. When it is not, every new model becomes a new project with the same unanswered questions.

This is also why the work starts with a role rather than a platform. A useful first question is which decision a stronger agent should make better, such as answering a billing status question, prioritizing a service queue or preparing a supplier comparison. The answer tells you which data is needed, which systems must be reachable, and where a human must stay in the loop. If you want a broader view of the layers involved, the article on building a company superintelligence layer that improves with models describes how they fit together.

Which data should a Brazilian company organize first?

Data preparation is not a general cleanup. It is a targeted inventory of what the first role needs, followed by the fixes that make that data reliable enough for an agent to quote or act on. Most small and mid-sized companies find that their records are spread across spreadsheets, messaging apps, a commercial system and personal inboxes, and that the same customer or supplier appears under several spellings. An agent given that material will reproduce the confusion faster than a person would.

A practical first pass looks like this:

  • Name the source of truth for each entity the role touches, such as customers, products, price lists, orders, contracts and invoices.
  • Mark which records contain personal data, and confirm with counsel the basis on which your company holds and uses them under the Lei Geral de Proteção de Dados (LGPD).
  • Remove obvious duplicates in customer and supplier records, especially where one company appears under two names or two tax registrations.
  • Record the date each policy document was last valid, so an agent does not present an old refund rule as current.
  • Separate internal know-how from public material, because they need different access rules.
  • Keep a short index showing where each document lives and which person maintains it.

The Autoridade Nacional de Proteção de Dados publishes guidance in Portuguese on these obligations, and it is worth reading it with your legal advisor before deciding what an agent may retrieve. Retrieval is also a design choice: the [AWS guidance comparing retrieval and fine-tuning](https://docs.aws.amazon.com/prescriptive-guidance/latest/retrieval-augmented-generation-options/rag-vs-fine-tuning.html) explains why keeping company material in a searchable store is often easier to update and audit than baking it into a model.

Why do Portuguese-language processes need their own work?

Many Brazilian operations run on procedures that were never written down, or were written in a mix of formal Portuguese, shorthand and regional habits. A human employee can ask a colleague what "cancelamento" means in a particular case. An agent reads the words literally and acts on whatever interpretation the text allows. Terms such as devolução, reembolso, estorno, crédito and cancelamento can carry different meanings across teams, and an agent that confuses them will make confident mistakes.

Write each core process as the team actually performs it, in Portuguese, with four elements: the trigger, the steps, the exceptions and the person who decides the exceptions. Include a short glossary of internal terms and the abbreviations used on invoices, tickets and messages. Keep the language plain and dated where it matters. The guidance in the documentation standards for a superintelligence-ready company gives a format that stays readable for both staff and agents.

How should approval rights be divided?

Approval rights answer a simple question: which actions can the agent take alone, which require a person to review first, and which are never delegated. The model's own text should never be the thing that decides this. Authorization belongs in the systems the agent calls, enforced by the tool or the backend, as the [OWASP authorization guidance](https://cheatsheetseries.owasp.org/cheatsheets/Authorization_Cheat_Sheet.html) recommends for any application that acts on behalf of a user.

A workable decision guide for a first role:

  • Reading a record the role is already permitted to see: no approval beyond the role's standing permission.
  • Drafting a reply or document for staff review: approval happens when a person sends or signs it.
  • Changing a stored record, such as a customer address or a ticket status: approval by the record owner, with the change logged.
  • Issuing a refund, sending a contract or committing company money: named human approval before execution, with a limit the owner sets in writing.
  • Changing the agent's own tools, memory or permissions: a formal change request reviewed by an accountable leader.

The thresholds in this guide should be reviewed as the agent proves itself, not set once and forgotten. The article on human approval thresholds for stronger AI agents describes how to move them deliberately, and the guide on running a permission review for your AI agents shows the evidence to collect.

A five-step preparation plan

The sequence matters more than the speed. Each step produces something that the next one depends on, and skipping the first two usually means redoing the last three.

  • Step 1: Choose one role and write the decision it makes, along with the people who depend on that decision.
  • Step 2: List the data and tools the role needs, marking the owner, the sensitivity and the current quality of each source.
  • Step 3: Write the Portuguese process for the role, including exceptions, the glossary and the person who decides edge cases.
  • Step 4: Define the approval boundary, the actions that are never delegated and one observable success criterion, such as the share of drafts accepted without edits, measured against a baseline you set yourself.
  • Step 5: Log every action, review the log on a fixed schedule, and widen scope only after the evidence supports it.

Notice that none of these steps requires a particular model. They determine what any model would be allowed to see and do, which is the part that stays valuable when models change.

An illustrative scenario: a billing question

The following scenario is illustrative and does not describe a real company. A distributor with a small finance team wants an agent to answer customers' questions about invoice status. Before preparation, the invoices live in two systems, two customer records exist for some buyers, and the refund policy is in an email from a previous year. The agent answers quickly and sometimes wrongly.

In the prepared version, the team names the invoice system as the source of truth, merges the duplicate customer records and dates the refund policy. The process is written in Portuguese, with the difference between estorno and reembolso defined in the glossary. The agent can read invoice status and draft replies, but any refund or credit goes to a named person in finance. Each week, the team reviews a sample of the agent's replies and the log of blocked requests. When the sample consistently matches the policy, the owner decides whether any category of small, clearly rule-based credit can move under a higher limit. The agent's capability is not what changed the outcome. The structure around it did.

Be Human handles this kind of work as a managed layer. An engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review and controlled improvements, so the company keeps control of what the agent knows and what it is allowed to do.

Where should a leader start?

Start with the smallest role that matters to someone in your company, and resist the urge to organize every system at once. The readiness guide for Brazilian companies lays out a sequence for this, and the Be Human AI diagnostic is free to start with, if you want help identifying the first task worth handing to an agent. Take the free Be Human AI diagnostic to find the first task worth handing to an agent.

Frequently asked questions

Do we need new software before we can prepare? No. Most of the early work is an inventory, a glossary and a written decision guide, and it can be done with the systems you already use. New tools become useful after you know which data and actions the first role depends on. Buying first often means paying for capabilities you cannot yet control.

How do we handle personal data in an AI agent under Brazilian rules? Identify which records contain personal data, document the basis on which your company holds and uses them, and limit the agent to the fields it needs. Confirm the approach with your legal advisor and review the guidance published by the Autoridade Nacional de Proteção de Dados. Keep the access log so you can show who or what touched each record.

Should our processes be written in Portuguese or English? Write the operational processes in the language your team uses every day, which for most Brazilian companies is Portuguese. Keep a glossary for internal terms and any English system labels so that the agent and staff read the same meanings. Translation can come later for partners or documentation shared abroad.

Who should approve an agent's actions? The person accountable for the business outcome should approve consequential actions, not the technical team alone. Technical staff should enforce the permissions and keep the logs, while the business owner sets the limits and reviews exceptions. Written thresholds make this division clear and easy to audit.

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 →