
Superintelligence for Unions in Brazil: An Agent With Human Decisions
How Brazilian unions can use AI agents to answer members, log requests and route legal questions to a named person, within LGPD limits.
Want to apply this in your company?
Book a call with Juan →For a union, superintelligence in the sense of systems that outperform people at everything is not the starting point. What works today is an AI agent that answers members on channels such as WhatsApp, responds to public or registration information, records every request in a history, and routes legal questions to a person on the team with the full context. Legal analysis, deadlines and conduct remain with the legal team. The agent exists to reduce queues, lost information and rework.
Key takeaways
- The agent answers, records and routes. It does not interpret the law, promise outcomes or decide whether a member is entitled to something.
- Union membership is sensitive personal data under the LGPD, Brazil's general data protection law, so collection, storage and access must be stricter than in a typical customer service setup.
- Every request becomes a record with defined fields, a status and a named owner, so nothing depends on a staff member's memory.
- Legal handoffs follow explicit triggers, always go to a person, and include a case summary so the member does not have to repeat their story.
- The project starts with a map of current service and a pilot on one real flow, with agreed metrics and a record of every conversation.
What the agent can do and what stays with the team
Before configuring any tool, the board needs a clear division of responsibilities. Settling this in advance avoids case-by-case arguments later, when the agent is already in production.
The agent can handle the following:
- Share office hours, service addresses and general documents that the union has already published.
- Collect the details of a request, such as request type, date, a short description and the documents sent.
- Confirm receipt and state, in general terms, the next step in the internal process.
- Check the status of a request already on file, without exposing data about other members.
- Book a meeting with the team, when the calendar is available for that purpose.
The following should stay with people on the team:
- Any analysis of legal rights, whether a lawsuit is appropriate, procedural deadlines or amounts owed.
- Negotiations, settlements, legal opinions and guidance on how a member should act in a dispute with an employer.
- Cases with signs of urgency, threats, harassment, workplace accidents or questions about sensitive data of third parties.
- Complaints about the union's own service or about the conduct of officers and staff.
The practical rule is simple: if the correct answer depends on interpreting a specific situation, it does not belong to the agent.
Map the service before choosing the tool
The most common mistake is starting with the technology. Before speaking with a vendor, spend at least a few weeks observing real demand. Note which questions arrive, through which channel, at what times, who answers them today, and where information gets lost. This exercise is what our methodology calls mapping and baseline, and it lets you measure later whether the pilot actually improved anything.
A possible mapping routine:
- List the types of contact received (membership, dues, documents, benefits, legal questions, complaints) and estimate the share of each.
- Identify where each type ends up today: a spreadsheet, a notebook, email, a personal WhatsApp group or an internal system.
- Record how long a request takes between first contact and the first human reply, and where it stalls without an owner.
- Mark which information is truly needed to route each case and which is collected out of habit.
- Define who may see each type of record, including officers, legal, finance and front-desk staff.
This map also shows where automation is not justified. A request that today reaches the right person within a day may not need an agent at all. A request that gets lost between three channels is a good candidate.
Member data requires more care under the LGPD
The LGPD treats union membership as sensitive personal data. This changes the design of the system. In practice, the agent should collect only what the request requires, keep it for a defined period, and restrict access to the records.
The team should decide these points in writing, with support from legal counsel or the data protection officer:
- The purpose of each collection, explained in clear text to the member before or at the start of the conversation.
- The legal basis for each processing activity. Consent is not always the only option, so this must be assessed case by case.
- Which data goes into the record and which should never be requested through the chat, such as passwords, full bank details or medical information that is not necessary.
- The retention period and the routine for deletion or anonymization.
- Whether any model or platform provider processes data outside Brazil, which requires additional analysis, as explained in our article on international data transfer with AI providers.
- How a member can ask for review of an automated decision, a topic covered in our article on automated decisions and the right to review.
The National Data Protection Authority (ANPD) publishes guidance on data processing in Brazil and is the official source to follow. Any legal opinion on a specific case should come from the responsible lawyer, not from a generic text.
How to record requests so nothing is lost
The value of the agent shows up in the record. A well-structured request lets the team pick up a case without asking everything again. A minimal record usually includes:
- Member identification, with basic verification and no excess data.
- Request type, chosen from a closed list, with an "other" option that always triggers human review.
- A summary of a few lines, written in the member's own words and without judgment.
- Documents received, with an indication of which are still pending.
- Channel, date and time of contact.
- Status, such as received, awaiting document, sent to legal, answered or closed.
- The current owner, named as a person and not only as a department.
- A log of every routing decision the agent made, including the reason.
This last item is what we call a decision log. It lets you audit later why a case went to one queue and not another, and it is essential for adjusting the agent's rules based on facts.
Legal handoff: triggers and transfer
Routing is the most sensitive part of the flow. Define fixed triggers that do not depend on the mood of the conversation. Examples of triggers that should move a case to a person immediately:
- The member mentions dismissal, sick leave, an accident, harassment or a deadline that is running.
- The member asks what to do in their own case or whether they should sue someone.
- The member shows confusion or distress, or asks to speak with someone.
- The message includes documents containing data about third parties.
- The agent does not have a vetted answer for the question.
The handoff must be complete. The person who takes over receives the summary, the conversation history and the request type, and the member is told who will respond and through which channel. WhatsApp also has rules about the period during which you can freely reply to a conversation, so review what our article on the 24 hour window describes before designing how the member will be contacted again.
The agent also needs protection against hidden instructions in messages or documents sent by members. The OWASP guidance on prompt injection prevention describes the most common risks and defenses. The basic rule is that the agent must never be allowed to perform actions outside the defined flow simply because it read a text.
Illustrative scenario: a document request that became a legal case
Illustrative scenario: a member writes on WhatsApp outside business hours asking for help submitting a document. In the middle of the message, they say they were let go last week and do not know whether they are entitled to anything.
The agent confirms that it received the document request and records it with the correct type. When it detects the mention of dismissal and the question about rights, it triggers the legal handoff. The message to the member explains that the case has been forwarded to a person on the team, that a reply will come through the agreed channel, and that in the meantime they can keep the documents they already have. The agent does not comment on deadlines or chances of success.
The next day, the staff member opens the record with the summary, the attached document and the history, and makes contact without asking the member to repeat everything. If the case is urgent, the team adjusts the trigger rule for the next pilot cycle. The gain here is not a fast answer about rights, which the agent never gives. The gain is that the question was not lost.
How to run the pilot safely
The methodology we apply follows four steps: map and measure the current situation, limit the agent's data and actions, pilot a real flow with a record and agreed metrics, and then transfer the operation to the team and decide whether to expand, adjust or stop. This is a way to organize the work, not a promise of results, and the choice of metrics depends on the reality of each organization.
A checklist for starting the pilot:
- Choose one request type to begin with, preferably a low-risk one.
- Write down what the agent can and cannot do, approved by legal and by the board.
- Test the agent with anonymized real questions and with cases that should be routed to a person, before any member uses it.
- Define a success metric before starting, such as time to the first human reply or the share of requests with a complete record.
- Review records and routed conversations every week, focusing on errors.
- Name an internal owner who can switch the agent off quickly.
- At the end of the agreed period, decide whether the flow continues, needs adjustment or should be closed.
The NIST AI Risk Management Framework is a useful reference for structuring this kind of assessment without relying on intuition. To understand what an agent with clear limits looks like, as opposed to an open-ended chatbot, our guide on how a service agent should be designed helps separate what to automate from what to keep with people.
If the word superintelligence comes up in board discussions, align the vocabulary first. Systems with large capabilities do not replace the responsibility of whoever signs the guidance given to a member. Our text on ChatGPT and superintelligence explains the difference between what exists today and what remains speculation.
Frequently asked questions
Is a union AI agent the same as superintelligence?
No. For a union, the practical starting point is a narrow AI agent with defined tasks, limits and a record of every action. Superintelligence is a broader and still speculative topic, and it does not change the responsibilities of the legal and administrative team.
Can the agent answer a member's legal questions?
No. The agent can collect the request and route it to a person with a summary. Analysis of rights, deadlines, chances of success and conduct belongs to the legal team.
Is union membership protected by the LGPD?
Yes. The LGPD treats union membership as sensitive personal data, which calls for stricter limits on collection, storage, access and retention. The legal team or data protection officer should define these rules in writing before the pilot begins.
Where should a union start?
Start by mapping how requests arrive today and where they get lost. Then choose one low-risk request type, define the agent's permitted actions with legal approval, and run a limited pilot with agreed metrics and weekly review of records.